Updated August 16, 2026: Initial publication. Model sizes and memory figures verified against official model releases as of this date; this table is refreshed quarterly because model sizes churn.

Every hardware guide for local AI eventually comes down to one question: will the model fit? This page answers it with numbers, then goes one step further than the spec-sheet math. Where a model class appears in the tables below, we have run it in the StorageReview lab, and the hardware recommendations link to systems we have actually tested on our desktop and laptop Local AI leaderboards.

The short version: parameter count alone no longer predicts hardware requirements. Quantization shrinks weights by roughly 4x, mixture-of-experts models run far lighter than their total parameter counts suggest, and agentic workloads add a memory tax that most sizing guides skip entirely. Here is how to budget all three.

Model Memory Requirements

File sizes below are the common Q4_K_M quantized releases (or the native format where noted) as published on Ollama and Hugging Face in August 2026. The “Memory to Run” column adds working overhead and a moderate context window; figures marked with an asterisk are computed estimates, while the GPT-OSS and DeepSeek 671B figures are the vendors’ own statements.

Model	Parameters	Q4 / Native Download	Memory to Run*	Runs On
Llama 3.1 8B	8B	4.9 GB	~6 to 8 GB	8GB VRAM laptops and up
Phi-4 14B	14.7B	9.1 GB	~11 to 12 GB	16GB VRAM or 64GB shared-memory laptops
GPT-OSS 20B (MXFP4)	21B (3.6B active)	12 to 14 GB	~13 to 16 GB	16GB VRAM, shared-memory laptops
Mistral Small 3.2 24B	24B	15 GB	~18 to 20 GB	24GB RTX PRO laptops
Gemma 3 27B	27B	17 GB	~20 to 22 GB	24GB RTX PRO laptops, unified-memory desktops
Qwen3 30B-A3B	30.5B (3.3B active)	19 GB	~22 to 24 GB	24GB VRAM (tight) or unified memory
QwQ 32B / DeepSeek-R1 32B	32B	20 GB	~22 to 24 GB	24GB VRAM (tight) or unified memory
Llama 3.3 70B / R1 Distill 70B	70B	43 GB	~48 to 50 GB	96GB+ unified memory, RTX PRO 6000 towers
GPT-OSS 120B (MXFP4)	117B (5.1B active)	65 GB	~65 to 80 GB	96GB to 128GB unified memory, 96GB GPUs
DeepSeek-R1 671B (full)	671B (37B active)	404 GB	450+ GB	Out of reach for desktops; rack-scale territory

*Estimates assume roughly 8K of active context and about 15 percent runtime overhead on top of the weight file. Larger contexts cost substantially more; see the agentic tax below.

Two patterns worth noticing. First, mixture-of-experts changes the math: GPT-OSS 120B carries 117 billion parameters but activates only 5.1 billion per token, so it runs on a single 96GB to 128GB memory pool, and OpenAI states the target as a single 80GB GPU. We have run it on the HP Z2 Mini G1a, a 1-liter machine with no discrete GPU at all. Second, the full DeepSeek-R1 at 671B parameters is in a different universe: 404GB just to download at Q4, roughly 450GB of combined memory to run it acceptably. Nothing on any of our leaderboards runs it, and no desktop should try.

What Runs Where

8GB VRAM laptops (RTX PRO 500 to 2000 class). Comfortable with 7B to 8B models at Q4 and the smaller Phi and Gemma variants. The ceiling is real: in our testing, a 13B model that wanted about 12GB of graphics memory simply did not finish on 8GB cards. Our Best Laptops for Local AI page covers the field.

24GB VRAM systems (RTX PRO 5000 laptops, desktop RTX cards). The sweet spot for 14B to 27B models with context to spare, and the fastest way to run them; our Dell Pro Max 18 Plus leads that field at 185 tokens per second on Phi. The 32B class fits at Q4 but leaves little room for context, which matters more than you think once agents are involved.

Unified and shared memory systems (96GB to 128GB). AMD Strix Halo machines can assign up to 96GB of system RAM to the GPU, and GB10 systems like the NVIDIA DGX Spark carry 128GB of coherent memory. This class runs 70B models at Q4 and GPT-OSS 120B, trading speed for capacity. The Best Desktops for Local AI leaderboard ranks them, and the same architecture reaches laptops in the HP ZBook Ultra G1a 14, which loaded DeepSeek-R1 70B in our testing.

RTX PRO 6000 workstation towers (96GB per card, up to 384GB tested). Speed and capacity at once: 70B models at higher quantization with full context, GPT-OSS 120B with headroom, or several models resident simultaneously for multi-agent pipelines. Our Best Desktop Workstations leaderboard covers the class, topped by the four-GPU-capable HP Z8 Fury G6i.

The Agentic Tax: Context Is a Second Memory Budget

Sizing from the weight file alone works for short chat sessions. Agents break that assumption. An agent loops: it reads tool output, files, and prior steps back into its context window, and every token held in context costs memory in the KV cache on top of the weights. The costs are not small.

Model Class	KV Cache at 32K Context	KV Cache at 128K Context	Rough Cost per 1,000 Tokens
8B (Llama 3.1 8B, FP16 KV)	4.19 GB	16.78 GB	~0.13 GB
70B (Llama 3.3 70B, FP16 KV)	10.49 GB	41.94 GB	~0.33 GB

Read that table against the sizing one above and the problem is obvious. A 70B model at Q4 is 43GB of weights, but at a full 128K context it needs roughly 85GB in total, which is double the weights. An 8B model at 128K spends more memory on context (16.78GB) than on its own Q4 weights (4.9GB). KV cache quantization helps, with FP8 halving those figures and INT4 quartering them at some quality cost, but the planning rule stands: for agentic work, budget context like a second model.

Serving multiplies the tax again. If a system hosts several agents or users concurrently, each request holds its own KV cache. vLLM, the serving stack we benchmark with on our Local AI leaderboards, pre-allocates 90 percent of GPU memory by default precisely because KV space is what determines how many concurrent requests survive without throughput-killing preemption. A machine that runs one chat comfortably can be undersized for three agents.

System RAM: The Forgotten Third Leg

For GPU-only inference, system RAM just needs to stay out of the way, and 32GB to 64GB is a comfortable floor on the machines we test. It becomes decisive in two cases. Offloading, where layers that do not fit in VRAM spill to CPU, works through memory-mapped model files, so system RAM should at least match the model file size plus OS headroom. And on unified memory machines, system RAM is the GPU memory, which is exactly why 64GB and 128GB configurations dominate our Local AI leaderboards. Ollama’s own floor guidance runs 8GB of RAM for 7B models, 16GB for 13B, and 32GB for the 33B class; treat those as minimums, not targets.

Storage: The Leg Everyone Skips

Model libraries get large quickly. The ten models in our sizing table total roughly 240GB in their common quantizations, and a working library with a few variants, embedding models for retrieval, and vector stores for agent memory can pass 500GB before any project data arrives. On top of that, agentic workflows generate real scratch traffic: logs, checkpoints, and retrieval indexes that live alongside the models.

Load time is where drive class shows up. The arithmetic is simple: a 65GB GPT-OSS 120B file reads in about 109 seconds at SATA speeds, roughly 17 seconds on a Gen3 NVMe drive, and under 9 seconds at Gen4 rates, first load only, since the OS page cache makes reloads near-instant. Real-world loaders do not always sustain full drive speed, so we plan to publish our own measured model-load numbers across drive classes; watch the changelog. Our Storage Leaderboard covers the drives themselves.

Practical guidance: 2TB NVMe is the sensible floor for a dedicated AI workstation, 4TB if the machine hosts agents that accumulate state, and Gen4 or better if you swap between large models regularly.

Agentic AI Hardware FAQ

How much VRAM do I need to run a 70B model locally?

About 48 to 50GB at Q4 with moderate context, which is why this class belongs to 96GB unified memory machines and 96GB workstation GPUs rather than any consumer card. Push the context toward 128K and the total approaches 85GB. Run it at Q8 and the weights alone are 75GB.

Can I run the full DeepSeek-R1 locally?

Not on anything this site ranks. The full 671B model is a 404GB download at Q4 and wants roughly 450GB of combined memory to run acceptably. The practical local path is the official distills, and the 32B and 70B distills in our sizing table capture much of the capability at desktop-class requirements.

Do mixture-of-experts models change the hardware math?

Substantially. MoE models activate a fraction of their parameters per token, so memory scales with total parameters but speed scales closer to active parameters. GPT-OSS 120B (5.1B active) and Qwen3 30B-A3B (3.3B active) both run faster than their sizes suggest on the memory pools that fit them. You still need the memory for all the weights; you just get more speed per gigabyte.

How much storage does an AI workstation need?

Plan 2TB NVMe minimum, 4TB for agent hosts. A modest model library alone runs hundreds of gigabytes, agents accumulate logs and vector stores, and the largest single files (65GB and up) make slow drives painful every time you load or swap a model.

What hardware should I buy for local agents specifically?

Prioritize memory over raw speed. Agents hold long contexts and sometimes run concurrently, so the KV cache budget matters as much as the weight budget. A 96GB to 128GB unified memory system on our desktop Local AI leaderboard is the most economical entry, and RTX PRO 6000 towers on the workstation leaderboard are the answer when speed and concurrency both matter.

On the Horizon: GB300 Is in the Lab

Everything above reflects hardware we have tested. One system that will change the math is already on the bench: we have an NVIDIA GB300 Grace Blackwell Ultra deskside system in for review.

The reason it matters for this guide is memory. Every recommendation on this page works around a hard ceiling, splitting models across GPUs, leaning on quantization, or accepting CPU offload when weights and KV cache will not fit. NVIDIA specs the GB300 deskside systems with several hundred gigabytes of coherent memory addressable by the accelerator, which is a different order of magnitude from the 96GB-class cards that anchor the tables above. Model classes that currently require a multi-GPU tower, or that we list as impractical to run locally at all, move into single-box range.

We are not going to put numbers to that until we have run it. No figure on this page is based on GB300, and nothing here has been adjusted in anticipation of it. When the review publishes, this guide gets revised against measured results and the change will be noted in the log below.

How We Maintain This Guide

Model sizes and the popular-model roster change quarterly, faster than any hardware cycle. We re-verify the sizing table against official releases on a quarterly pass, and lab receipts are added as new systems clear our Local AI testing. Changes land in the dated note at the top of the page. Sizing figures marked as estimates use the published weight file plus measured KV cache costs and standard runtime overhead; vendor-stated figures are identified as such.

The post RAM, GPU, and Storage for Agentic AI: How Much You Actually Need appeared first on StorageReview.com.

Updated August 15, 2026: Initial publication with 27 tested Ubiquiti products. This page is refreshed every few weeks as new review units come through the lab.

Every product on this page has been through the StorageReview lab. Ubiquiti ships faster than almost any vendor we cover, and the catalog has grown to the point where picking between four NAS models or nine cameras is genuinely hard. This is the index: every UniFi and Ubiquiti product we have tested, grouped by what it does, with the specs that matter, our take, and a link to the full review behind each one.

Affiliate disclosure: the links in the Buy (affiliate) column of every table on this page are affiliate links to Ubiquiti’s store. If you purchase through one, StorageReview may earn a commission at no extra cost to you. Affiliate relationships never affect our testing, our conclusions, or which products we cover, and we do not list any product we have not tested in our lab.

A note on pricing: the price column is what the product cost at the time we reviewed it. Ubiquiti pricing is more stable than most of the hardware we cover, but check the store for current numbers before buying.

Where to Start

If you are building a UniFi setup from scratch, the order that matters is gateway, then switching, then cameras or storage. The Cloud Gateway Fiber at $279 is the cheapest sensible entry to 10G with real IDS/IPS throughput, and the Dream Router 7 folds WiFi 7 into the same box if you would rather not run separate access points.

For storage, the split is straightforward. The UNAS 2 is the $199 two-bay starting point, the UNAS Pro is the seven-bay enthusiast box, and the Enterprise NAS is where Ubiquiti stops competing with home NAS vendors and starts competing on real throughput, at 4,170 MB/s sequential read in our testing.

On cameras, nearly everything current is 4K with on-device AI, so the decision is form factor and field of view rather than image quality. The G6 Turret at $199 is the volume pick, the G6 Pro 360 replaces several cameras in one room, and you will need an NVR or a Protect-capable gateway to record any of it.

Our Picks

Every pick below is a product we have had in the lab. Full reviews are linked from each entry, and the complete list of everything we have tested follows.

Best UniFi NAS Overall: Ubiquiti Enterprise NAS

A 3U 16-bay ZFS box with dual 25GbE that measured 4,170 MB/s sequential NFS read and 90.9K IOPS 4K random read in our lab. Review: Ubiquiti Enterprise NAS Review

Best UniFi NAS for a Home or Small Office: UniFi UNAS 2

Two bays, PoE++ powered, up to 48TB raw, and 270.5 MB/s read over 2.5GbE. Review: UniFi UNAS 2 Review

Best UniFi Camera: UniFi G6 Turret

4K 8MP on a 1/1.8-inch sensor with on-device face and license plate recognition, at $199. Review: UniFi G6 Turret Review

Best UniFi Camera for Full-Room Coverage: UniFi G6 Pro 360

12MP panoramic sensor covering 180 degrees from a single ceiling mount. Review: UniFi G6 Pro 360 Review

Best UniFi Gateway: UniFi Cloud Gateway Fiber

5 Gbps of IDS/IPS throughput with 10GBASE-T and dual 10G SFP+, the cheapest route into 10G UniFi. Review: UniFi Cloud Gateway Fiber Review

Best UniFi Gateway with WiFi Built In: UniFi Dream Router 7

2.3 Gbps IDS/IPS throughput, WiFi 7, and a 10G SFP+ WAN in one box. Review: UniFi Dream Router 7 Review

Best UniFi Access Point: UniFi E7

Ten spatial streams of WiFi 7 rated at 11.5 Gbps on 6 GHz, with a 10GbE uplink. Review: UniFi E7 and E7-Campus Review

Best UniFi Switch for 10GbE: Switch Pro XG 10 PoE

Ten 10GbE RJ45 PoE ports, a 400W PoE budget, and 240 Gbps of switching capacity. Review: Switch Pro XG 8 PoE and Pro XG 10 PoE Review

Best UniFi Rack Accessory: UniFi Power Distribution Pro

A 2U PDU with 16 individually switchable outlets and 1,875W capacity for $279. Review: UniFi Power Distribution Pro Review

Everything We Have Tested

Network Storage

Product	What It Is	Our Take	Price at Review	Buy (affiliate)
Ubiquiti Enterprise NAS	3U 16-bay ZFS NAS, dual 25GbE SFP28, 64GB ECC, redundant 550W	Measured 4,170 MB/s sequential NFS read and 90.9K IOPS 4K random read	$3,999	Ubiquiti Store
UniFi UNAS Pro 8	2U 8-bay with NVMe cache, dual 10G SFP+, redundant power	Measured just over 2.2 GB/s sequential read	$799	Ubiquiti Store
UniFi UNAS Pro	7-bay desktop NAS, 10G SFP+, 154TB raw as tested	Measured 799 MB/s sequential read in RAID10	$499	Ubiquiti Store
UniFi UNAS 2	2-bay PoE++ NAS up to 48TB raw, 2.5GbE	Measured 270.5 MB/s read and 257.9 MB/s write	$199	Ubiquiti Store

Cameras and NVR

Product	What It Is	Our Take	Price at Review	Buy (affiliate)
UniFi G6 Turret	4K 8MP, 1/1.8-inch sensor, 30m IR, PoE	On-device face and license plate recognition at $199	$199	Ubiquiti Store
UniFi AI Dome	8MP 4K dome with on-device AI, PoE	40m IR, against 30m on the G6 Dome and 9m on the G5	$399	Ubiquiti Store
UniFi G6 Pro 360	12MP panoramic, 180 degree coverage, PoE+	Full-room coverage from a single ceiling camera	$499	Ubiquiti Store
UniFi G6 Dome	8MP 4K dome, 134 degree diagonal FoV, 9.25W max	Professional-grade build at a mid-range price	$279	Ubiquiti Store
UniFi G6 PTZ	Dual-lens 4K with 10x hybrid zoom, PoE+	AI motion tracking with smooth UniFi Protect integration	$399	Ubiquiti Store
UniFi G6 Bullet and G6 Instant	Both 4K 8MP; Bullet wired, Instant wireless	4K AI cameras under $200	$179 to $199	Ubiquiti Store
UniFi AI Camera Series	AI Pro 8MP with 3x optical zoom, AI Bullet, AI Theta	The high end of Protect, up to the $2,499 AI DSLR	$299 to $2,499	Ubiquiti Store
UniFi G5 Series	G5 Bullet, Turret Ultra, Dome Ultra, G5 Pro	The value tier, starting at $129	$129 to $1,800	Ubiquiti Store
UniFi Network Video Recorder Pro	7-bay NVR up to 168TB raw, 10G SFP+	Handles up to 24 4K cameras	$499	Ubiquiti Store

Switching and PoE

Product	What It Is	Our Take	Price at Review	Buy (affiliate)
UniFi Switch Flex 2.5G 8 PoE	8x 2.5GbE PoE++ plus 10GbE input, 60 Gbps capacity	196W of PoE budget with the optional 210W adapter	$199	Ubiquiti Store
Switch Pro XG 8 PoE and Pro XG 10 PoE	8 or 10 10GbE RJ45 PoE ports, dual 10G SFP+	XG 10 carries a 400W PoE budget and 240 Gbps switching	$499 to $699	Ubiquiti Store
Switch Pro Max 16 PoE	12x GbE PoE+ and 4x 2.5GbE PoE++, 10G uplinks	180W PoE in an affordable Layer 3 switch	$399	Ubiquiti Store
Switch Pro Max 48 PoE	48 ports, 4x 10G SFP+, 820W internal supply	720W total PoE and 224 Gbps switching capacity	$1,299	Ubiquiti Store

Gateways and Routers

Product	What It Is	Our Take	Price at Review	Buy (affiliate)
UniFi Cloud Gateway Fiber	10GBASE-T, dual 10G SFP+, NVMe slot to 2TB	5 Gbps IDS/IPS throughput, a cheap route to 10G	$279	Ubiquiti Store
UniFi Dream Router 7	WiFi 7 gateway, 10G SFP+ and 2.5GbE WAN	2.3 Gbps IDS/IPS throughput with WiFi 7 built in	$279	Ubiquiti Store
UniFi Dream Machine Pro Max	8x GbE LAN, dual 10G SFP+, RAID-capable storage	5 Gbps routing with full DPI and IPS enabled	$599	Ubiquiti Store
UniFi Express Gateway (UX)	WiFi 6 travel gateway, USB-C powered, 10W max	Full UniFi in a pocket-sized travel router	$149	Ubiquiti Store

WiFi Access Points

Product	What It Is	Our Take	Price at Review	Buy (affiliate)
UniFi E7 and E7-Campus	WiFi 7, 10 spatial streams, 10GbE uplink, PoE++	Rated 11.5 Gbps on 6 GHz	$499 to $799	Ubiquiti Store
UniFi U7 Pro XG and XGS	WiFi 7 with 10GbE uplink; XGS adds 4×4 on 5 GHz	XGS steps to PoE++ and 4×4 MU-MIMO	$199 to $299	Ubiquiti Store
UniFi U7 Outdoor	WiFi 7 outdoor AP, 2×2 MIMO, IPX6, PoE+	Rated 465m of open-space coverage with the directional antenna	$199	Ubiquiti Store

Power and Tools

Product	What It Is	Our Take	Price at Review	Buy (affiliate)
UniFi Power Distribution Pro	2U PDU, 16 switchable outlets, 1,875W max	Solves real rack headaches for $279	$279	Ubiquiti Store
UniFi UPS Tower	1000VA / 600W tower UPS, 11 outlets	Roughly 50 minutes of runtime at a 36 to 48W load	$159	Ubiquiti Store
UACC SFP Wizard	Pocket transceiver programmer for SFP through QSFP28	75 minutes of continuous SFP diagnostics per charge	$49	Ubiquiti Store

Ubiquiti FAQ

Is Ubiquiti gear worth it?

For the money, in our testing, generally yes, with a caveat. The value is in the ecosystem: one controller for network, cameras, storage, and power, with no per-camera or per-site licensing, which is where competing surveillance and networking stacks quietly get expensive. The caveat is that the value depends on staying inside the ecosystem, so mixing in third-party gear erases much of the advantage.

What is the best UniFi camera?

For most installations, the G6 Turret at $199, which brings 4K, a 1/1.8-inch sensor, 30m IR, and on-device face and license plate recognition at the lowest price in the current G6 line. Choose the AI Dome instead when you need longer night range, at 40m against the G6 Dome’s 30m, and the G6 Pro 360 when one 12MP panoramic camera can replace two or three fixed ones.

Which UniFi NAS should I buy?

The UNAS 2 at $199 for two drives and light duty, the UNAS Pro at $499 for seven bays and 10G, and the Enterprise NAS at $3,999 when you need dual 25GbE, ZFS, and redundant power. The performance gap is wide and worth understanding: we measured 270.5 MB/s on the UNAS 2 against 4,170 MB/s on the Enterprise NAS.

Do I need a UniFi gateway to use UniFi cameras?

You need something running UniFi Protect, which means either a Protect-capable console such as the Dream Machine Pro Max, or a dedicated recorder like the UNVR-Pro. Cameras alone will not record. Budget for the recorder and the storage in it when pricing a camera project, because that is where camera-only quotes tend to fall apart.

How often is this page updated?

Every few weeks. Ubiquiti releases frequently and we test most of what ships, so new entries are added as reviews publish and the dated note at the top records what changed. If a product is not on this page, we have not tested it, and we do not list hardware we have not put through the lab.

The post Ubiquiti Reviews: Every UniFi Product We Have Tested appeared first on StorageReview.com.

Dell has qualified KIOXIA’s 245.76TB NVMe SSDs for software-defined ObjectScale, putting 9.83PB of raw flash in a single 2U server. The density milestone lands on top of the data-path work Dell detailed earlier this year across ObjectScale 4.2 and 4.3, which targets high-throughput AI training, small-object pipeline efficiency, and inference acceleration through native S3 over RDMA, Key-Value (KV) Cache offload for LLMs, an overhauled chunk-store engine, and direct Apache Iceberg support via S3 Tables.

Density Milestone: 9.83 PB in a 2U Footprint via 245 TB KIOXIA SSDs

As enterprise AI data lakes expand across model training, fine-tuning, and compliance-retention tiers, rack density, thermal envelope management, and floor-space constraints have become primary design variables. To address raw capacity per rack unit, Dell has qualified KIOXIA’s 245.76 TB NVMe SSDs for software-defined ObjectScale running on Dell PowerEdge R7725 servers.

By populating 40 of the 245.76 TB drives in a single 2U chassis, ObjectScale reaches 9.83 PB of raw all-flash capacity; we reviewed the platform last summer. This density is aimed directly at enterprise AI factories, vector databases, large analytics repositories, and cyber-resilience platforms that require multi-petabyte unstructured storage footprints while minimizing power consumption, cooling load, and cabling complexity.

Throughput Gains and Small-Object Architecture

In software-defined deployments on qualified Dell PowerEdge servers, Dell’s internal testing indicates per-node read throughput of up to 40 GB/s, an 8x increase over previous-generation all-flash object platforms. Dell also points to a UK high-frequency trading customer sustaining roughly 280 GB/s of aggregate read throughput in production. Those figures line up with what we have measured on the hardware itself: in our R7725xd testing last November, the platform sustained more than 300 GB/s across the internal NVMe pool and 160 GB/s over NVMe-oF RDMA. That work characterized the server directly rather than running ObjectScale on top of it, but it establishes the ceiling the software is working against.

To eliminate metadata bottlenecks common in pipelines dealing with millions of small feature sets, vector chunks, and telemetry logs, ObjectScale uses an optimized chunk-store engine. The architecture aggregates small payloads into unified 128MB chunks before applying erasure coding across nodes. For typical 10KB files, more than 10,000 objects fit into a single chunk, substantially lowering metadata overhead.

This chunking approach also restructures drive failure rebuilds by reducing the total number of shards that must be recreated from billions to millions, cutting recovery times for dense NVMe drives from weeks to hours. Inline checksums verified at the stripe level reduce host CPU cycles spent on background integrity scans. Furthermore, the re-engineered Key-Value Store in ObjectScale 4.2 provides roughly 4x better memory efficiency and cuts metadata disk footprints by 30% to 60%.

S3 over RDMA and Inference Acceleration

ObjectScale 4.3 expands S3 over RDMA across Dell’s all-flash portfolio, including the XF960, EXF900, and software-defined PowerEdge R7725xd deployments. Utilizing RoCEv2 networking and Dell’s S3-over-RDMA SDK, the platform bypasses kernel TCP overhead and CPU interrupts to establish direct memory paths between host GPUs and backend NVMe media. Internal testing shows this delivers up to 230% higher throughput, an 80% reduction in latency, and up to a 98% reduction in client CPU utilization compared to standard S3 over TCP.

For inference architectures, ObjectScale functions as an external tier for KV Cache offload. Because long-context inference quickly exhausts on-GPU HBM and VRAM, integrating with vLLM, LMCache, and NVIDIA’s NIXL library allows inference frameworks to spill active attention states directly to ObjectScale over RDMA-accelerated S3. Dell reports up to a 19x improvement in Time to First Token against standard vLLM without cache offload, along with up to 5.3x higher token throughput and nearly 3x higher multi-turn throughput. Separately, Dell cites a TTFT of roughly 0.86 seconds in its own head-to-head testing against a competing object platform.

Native S3 Tables and Scale-Out Efficiency

ObjectScale 4.3 also brings native S3 Tables, which Dell lists as a tech preview, embedding Apache Iceberg open table formats directly within the storage bucket layer. Query engines including Trino, Starburst, Spark, and Flink can run analytics natively against S3 buckets without requiring dedicated staging or ETL data movement into separate warehouses. Dell reports that this architecture delivers up to 2x faster ingestion and up to 4.5x faster queries compared to traditional warehouse configurations.

Underlying storage efficiency includes 24+2 and 24+4 erasure coding profiles, which Dell says cut write amplification by up to 75%. On high-capacity HDD platforms like the EX500, the company reports up to 25% faster large-object ingest and up to 2x higher mid-size object write performance. Workload-specific bucket compression supports Snappy, LZ4, ZSTD, and Deflate algorithms with telemetry for FinOps cost tracking. Provisioning and orchestration integrate natively with cloud-native tooling, including Kubernetes Container Object Storage Interface (COSI) and Terraform providers.

The post Dell Puts 9.83PB in 2U With 245TB KIOXIA SSDs, Adds S3 over RDMA and KV Cache Offload to ObjectScale appeared first on StorageReview.com.

NVIDIA has moved Spectrum-X Ethernet Photonics, its co-packaged optics (CPO) Ethernet switch platform, into full production. The company puts the gains at 4x fewer lasers, 5x lower power consumption, and 10x higher mean time between incidents, and it named the five manufacturing partners building the switch at each stage of the supply chain. The milestone lands as the Vera Rubin compute platform it is designed to network moves toward its own production shipments this fall.

Inside the Co-Packaged Optics Switch

Conventional switches rely on pluggable optical transceivers, discrete modules that each carry their own laser and plug into the switch faceplate. Spectrum-X Ethernet Photonics instead integrates optical engines directly adjacent to the switching silicon, built on co-packaged optics with 200Gb/s SerDes. Removing the pluggable tier is what drives the laser count down by a factor of four, and fewer lasers mean fewer discrete components that can fail in a fabric that may span hundreds of thousands of links.

Those failure points are the reason the reliability figure matters more than the power figure for most operators. NVIDIA’s 10x higher mean time between incidents is a claim about how often something in the optical path takes a link down, not about raw throughput. Lowering the thermal and electrical footprint of the networking tier also frees power headroom within the data center envelope for accelerator compute, a concern that drives NVIDIA’s 800 VDC power architecture as well. When NVIDIA first detailed the platform, it framed the same advantages against networks using traditional transceivers as 5x better power efficiency, 5x longer AI uptime, and 1.3x faster time to deployment.

The CPO Supply Chain, Named Stage by Stage

The more unusual disclosure is the manufacturing chain itself, which NVIDIA broke out by stage. TSMC handles silicon photonics fabrication. SPIL performs chip-scale packaging and testing. Lumentum fabricates the laser chips, and TFC Communication builds the laser module subassemblies. Foxconn does the final switch system assembly. Co-packaged optics has spent years as a technology that looked good in demonstrations and proved difficult to yield at volume, so a named, five-vendor chain running in production is a concrete signal of maturity.

Cloud providers, including CoreWeave, Lambda, Oracle Cloud Infrastructure, Microsoft Azure, IBM Cloud, and Nebius, are among the early adopters integrating the CPO fabric.

Vera Rubin, the Platform Being Networked

The switch exists to feed Vera Rubin, which NVIDIA said at GTC Taipei in late May had entered full-scale manufacturing, with production shipments set to begin this fall. Representing the third generation of NVIDIA MGX rack-scale design, the platform is engineered to deliver up to 10x higher agentic AI throughput at scale compared to the previous-generation Grace Blackwell architecture. Systems manufacturing spans Tier 1 server vendors, including Dell Technologies, HPE, Lenovo, and Supermicro, alongside ASUS, ASRock Rack, Compal, Foxconn, GIGABYTE, Inventec, MSI, Pegatron, Quanta Cloud Technology, Wistron, and Wiwynn. On the storage and infrastructure software side, NVIDIA names Cloudian, DDN, Hitachi Vantara, IBM, MinIO, NetApp, Nutanix, VAST Data, and WEKA among partners in full-scale production on the platform.

At the rack level, the Vera Rubin NVL72 integrates Vera CPUs, Rubin GPUs, and sixth-generation NVLink networking into a unified compute envelope. To address multi-tenant compliance and enterprise data protection, the architecture incorporates full-stack NVIDIA Confidential Computing. This implementation establishes a rack-scale Trusted Execution Environment (TEE) featuring hardware-level attestation and end-to-end line-rate encryption across high-speed interconnects to prevent physical and firmware-level tampering.

“Agentic AI is a new kind of workload. One prompt can launch a thousand-step journey of reasoning, retrieval, tool use and response generation,” said Jensen Huang, founder and CEO of NVIDIA. “Vera Rubin was built for this moment,” he added, describing it as “an AI factory engine that delivers intelligence at scale, with the performance, efficiency and security needed to power the next industrial revolution.”

BlueField-4 DPU Integration for Multi-Tenant Isolation

The Vera Rubin architecture also incorporates NVIDIA BlueField-4 Data Processing Units (DPUs) to offload infrastructure workloads from primary host processors. Operating at software-defined networking line rates up to 800 Gb/s, BlueField-4 delivers hardware-isolated multi-tenant networking, telemetry, and storage virtualization. Utilizing the BlueField-4 Advanced Secure Trusted Resource Architecture, cluster operators can enforce granular traffic isolation, automate policy enforcement, and manage control planes across large-scale distributed deployments.

With the optics now in volume manufacturing and Rubin shipments slated to start this fall, the practical question shifts from whether the hardware is real to how quickly operators can prepare for it.

The post NVIDIA Spectrum-X Ethernet Photonics Enters Full Production With 4x Fewer Lasers and a Five-Vendor CPO Supply Chain appeared first on StorageReview.com.

Updated August 14, 2026: Initial publication. Ranked on our current enterprise bench (FIO, GDSIO, and DLIO); drives tested on the older suite are noted in Also Tested.

Every drive ranked on this page has been through the StorageReview lab. We do not rank announced products or vendor spec sheets, which matters more in enterprise storage than anywhere else we cover: the datasheet numbers are achievable, and the interesting question is what happens at sustained load, at low queue depth, and during AI checkpointing. Every pick links to the full review holding the data.

The enterprise field split into two distinct races this cycle. TLC drives are competing on mixed-use throughput and latency consistency for databases, analytics, and AI training. QLC drives are competing on capacity per watt, and they have run away with it: a single bay now holds 245.76TB at roughly 8.2TB per watt, against about 4.4TB per watt for the densest hard drives. The picks below cover both races, plus the form-factor question that increasingly decides deployments before performance does.

At a Glance

Category	Drive	Class	Standout Lab Result	Full Review
Best Overall Enterprise SSD	Micron 9550 MAX	Gen5 TLC, 3 DWPD, up to 25.6TB	10,957.9 MB/s 128K sequential write, the highest in our comparison group	9550 MAX Review
Best Read-Intensive	Kioxia CD9P-R	Gen5 TLC, 1 DWPD, up to 61.44TB	~30 us 4K random read latency at QD1 against 60 to 90 us for the field	CD9P-R Review
Best High-Capacity	Micron 6600 ION 245.76TB	Gen5 QLC, up to 245.76TB	A quarter petabyte per bay at roughly 8.2TB per watt	6600 ION Review
Best Capacity Efficiency	Solidigm D5-P5336 122.88TB	Gen4 QLC, 0.6 DWPD, 122.88TB	134.3PBW endurance in a 24W active envelope, our Editor’s Choice	D5-P5336 Review
Best for EDSFF Deployments	Micron 7600 MAX	Gen5 TLC, 3 DWPD, up to 12.8TB	1.78M IOPS peak 4K random write inside a 14W envelope	7600 MAX Review
Best High-Capacity Alternative	DapuStor R6060 122TB	Gen5 QLC, 0.6 DWPD, up to 245TB	13,274.8 MB/s 64K random read and dual-port support	R6060 Review

The Picks

Best Overall Enterprise SSD: Micron 9550 MAX

The 9550 MAX is the drive that wins on the numbers that matter for mixed-use deployments. In our FIO sweep it led 128K sequential write at 10,957.9 MB/s, roughly 2.5GB/s clear of the next drive in the group, and did it at the lowest write latency in the field at 182.2 microseconds. It was the only drive in that comparison to scale past 10GB/s in 64K random write, averaging 7.34GB/s with a 10.6GB/s peak, and it held the tightest latency curves across nearly every sweep we ran.

Built on Micron 232-layer TLC with a 3 DWPD rating, it comes in 3.2TB through 25.6TB, in both U.2 and E3.S, and our 12.8TB sample is rated for 70,080TBW random and 143,100TBW sequential. In DLIO checkpointing against LLAMA 3.1 405B it posted the lowest average completion times in the group. If one drive has to cover databases, analytics, and AI training pipelines, this is it.

Review: Micron 9550 MAX Review: Balanced Performance for AI, DB, and Analytics

Best Read-Intensive: Kioxia CD9P-R

Read-heavy fleets care about latency at low queue depth more than peak throughput, and that is where the CD9P-R separates itself. At QD1 it served 4K random reads in roughly 30 microseconds while the rest of the group sat between 60 and 90, and it delivered 32.3K IOPS at a single job and single queue depth, clearly ahead of the field. It also tied for the top 128K sequential read at 14,235.9 MB/s and posted the group’s highest 1M GPU Direct Storage read at about 6.2 GiB/s.

The generational jump is real: random write IOPS went from 200K on the CD8P-R at the same capacity to 450K, a 2.25x improvement, with sequential read up 23 percent. It runs BiCS FLASH TLC at 1 DWPD, reaches 61.44TB in 2.5-inch, and our 7.68TB E3.S sample drew 23W active. The tradeoff is honest: it finished last in the group on 128K sequential write at 6,912.4 MB/s. Buy it for reads.

Review: Kioxia CD9P-R Review: Read-Intensive Gen5 Up to 61.44TB

Best High-Capacity: Micron 6600 ION 245.76TB

A quarter petabyte in one bay is a category unto itself. The 6600 ION reached 12,729.8 MB/s in 128K sequential read and roughly 1.75 million IOPS in 4K random read in our testing, which is more than enough to feed read-centric workloads, but the real argument is density per watt. At a 30W ceiling it works out to about 8.2TB per watt against roughly 4.4TB per watt for the highest-capacity enterprise hard drives, and a 720-bay E3.L rack lands at 176.9PB versus 31.7PB filled with 44TB HDDs.

Read the fine print before deploying. Endurance is 1.0 SDWPD sequential but drops to 0.075 RDWPD on 4K random writes, the top capacity uses a 16K indirection unit, and it posted the slowest DLIO passes and the highest 1M GDS write latency in its group at 53.7 milliseconds. For object stores, AI data lakes, and content repositories, those are acceptable trades. For write-mixed tiers, they are not.

Review: Micron 6600 ION 245TB SSD Review: A Quarter Petabyte Per Drive Bay

Best Capacity Efficiency: Solidigm D5-P5336 122.88TB

Solidigm pioneered enterprise QLC and the 122.88TB D5-P5336 is still the efficiency benchmark, and the only drive on this page to earn our Editor’s Choice. It pairs 0.6 DWPD with a 134.3PBW endurance rating over a five-year warranty inside a 24W active envelope, and its 192-layer QLC with a 32K indirection unit delivered 3,152.5 MB/s in 128K sequential write, 25.9 percent faster than the 61.44TB model, with latency 20 percent lower.

It runs on Gen4 rather than Gen5, which caps sequential reads around 7GB/s, but that is often beside the point at this capacity. Know its limit: 16K random write showed no scaling at all between low and high queue depths, holding at about 35,145 IOPS, where the 61.44TB version scaled to 162,711. It is also broadly qualified on major OEM platforms already, which matters when deployment timelines run in quarters.

Review: Solidigm 122.88TB D5-P5336 Review: High-Capacity Storage Meets Operational Efficiency

Best for EDSFF Deployments: Micron 7600 MAX

The 7600 MAX is the drive to specify when the chassis, not the benchmark, drives the decision. It is the only pick here shipping in all three of U.2, E1.S, and E3.S, and it is built on Micron G9 TLC at 3 DWPD with a 14W sequential power envelope, the lowest on this page. In our testing it was the most aggressive scaler in 4K random write, peaking just over 1.78 million IOPS, and it held the tightest 16K sequential read latency in its group at 0.13 milliseconds average.

Peak throughput is not its story. It finished last in the group in both 128K sequential charts, at 6,960.6 MB/s write and 11,240.5 MB/s read, the only drive under 12GB/s. What you get instead is composure under sustained load in a small, power-disciplined package, which is exactly what dense EDSFF servers need. We tested the 6.4TB E3.S; the family runs 1.6TB to 12.8TB.

Review: Micron 7600 MAX Review: Mixed Use 3 DWPD SSD Built for Modern Apps

Best High-Capacity Alternative: DapuStor R6060 122TB

The R6060 is the answer when the 6600 ION is not available or dual-port is a requirement. Its 122.88TB E3.L sample led its comparison group in 64K random read at 13,274.8 MB/s, matched the group high in 1M GDS sequential read at 5.9 GiB/s, and turned in the fastest first DLIO pass at 465.33 seconds. It supports PCIe 5.0 x4 or dual-port 2×2, carries 0.6 DWPD, and a 245TB SKU is in the family.

Small-block writes are the documented weakness, the same profile as every high-density QLC drive here, and the E3.L 2T form factor needs a chassis check before you commit. For read-heavy capacity tiers where dual-port matters, it earns its place.

Review: DapuStor R6060 122TB Review: Read-Heavy Gen5 QLC at Scale

Also Tested

These enterprise drives went through the same lab process and are worth a shortlist spot for the right deployment, even though they do not hold a category slot today.

• Phison Pascari X200P: topped the group in 128K sequential read at 14,242.1 MB/s, but posted the slowest DLIO checkpoint times in the field.
• Solidigm D7-PS1010: a strong 1 DWPD Gen5 drive at 14,163.3 MB/s sequential read, with a documented GPU Direct write collapse to 1.6 GiB/s at high thread counts.
• Western Digital SN861: the 4K random read leader at peak concurrency with 2,555.6K IOPS, though with more run-to-run variability than the 9550 MAX.
• Kingston DC3000ME: a mainstream Gen5 option at 8,477.4 MB/s sequential write, 1 DWPD and 14,016TBW at 7.68TB, aimed at system integrators.
• DapuStor J5060: 61.44TB of Gen4 QLC at 1.69M IOPS 4K random read, with 31.9K IOPS random write showing the QLC write ceiling clearly.
• Micron 6550 ION: the predecessor to the 6600 ION and still the sequential leader in its group at 13,979.7 MB/s read. Note this was a sponsored early look, and its efficiency figures are Micron-supplied.
• Kioxia CM7-R E3.S: solid Gen5 results at 714,623 IOPS 4K random read, tested on our older VDBench suite, so it is not directly comparable to the drives above.
• DapuStor Haishen5 H5100: benchmarked as a 16-drive RAID5 array at 205GB/s read and 18.1M IOPS, an array result rather than a single-drive comparison.

On the Horizon

Nothing in this section is ranked, because none of it has been through our lab. Every figure below is the vendor’s own claim, and the status label matters as much as the specs: shipping, sampling, and show-floor demo are three very different things when you are planning a refresh. We will rank these drives when we can measure them.

PCIe Gen6 arrives: Micron 9650, Samsung PM1763

The Gen6 generation is real and shipping, which is the biggest change coming to this page. Micron’s 9650 entered mass production in February 2026 as the first PCIe Gen6 enterprise SSD, with Micron claiming 28,000 MB/s sequential reads, 14,000 MB/s writes, and up to 5.5 million random read IOPS from G9 TLC in E1.S and E3.S, including a 9.5mm variant built for direct liquid cooling. It comes as a 1 DWPD PRO from 7.68TB to 30.72TB and a 3 DWPD MAX from 6.4TB to 25.6TB, inside a 25W envelope. Samsung followed with the PM1763 in July 2026, claiming 28,400 MB/s reads and 6.8 million random read IOPS, shipping first at 4TB, 8TB, and 15.36TB.

For context on what that means in practice, the fastest Gen5 drive on this page measured 14,235.9 MB/s sequential read in our testing. Gen6 claims roughly double that on paper. Whether it holds up under sustained load, at low queue depth, and through AI checkpointing is exactly what our bench exists to answer.

The capacity race past 245TB

The 6600 ION’s 245.76TB is the ceiling we have actually tested, and vendors are already past it on paper. DapuStor showed a 512TB QLC drive at FMS 2026 in the R6060 family, in E3.L and E2, pitched as a petabyte of flash in two drives. It is a demonstration unit: no performance figures, no pricing, and no availability date have been published. Kioxia’s LC9 at 245.76TB uses a 32-die stack of 2Tb QLC dies over PCIe 5.0 in 2.5-inch and E3.L, and has been sampling since mid-2025 without a confirmed general availability date. SK hynix began sampling its PS1101 QLC enterprise drive to cloud providers in August 2026, claiming roughly 55 percent more performance than its previous generation.

Also worth watching

Kioxia’s CM10 is its first PCIe 6.0 enterprise SSD, sampling as of July 2026, built on 332-layer BiCS FLASH with cold-plate liquid cooling support on the E3.S and E1.S models; Kioxia claims about 92 percent higher sequential read than the CM9, though it has not published absolute figures. Further out, the GP Series uses XL-FLASH with 512-byte access granularity to extend GPU memory rather than serve as conventional storage, with evaluation samples expected at the end of 2026. Solidigm has said publicly that it will ship 245TB-class drives before the end of 2026 but has not named a model or published specs.

How We Rank

Every drive ranked here was tested in the StorageReview lab on our current enterprise bench: a Dell PowerEdge R760 with a Serial Cables Gen5 JBOF running Ubuntu 22.04.2 LTS. The suite is FIO for four-corners and mixed block sizes, GDSIO for GPU Direct Storage paths, and DLIO checkpointing against LLAMA 3.1 405B with 1,636GB checkpoints, which is where AI infrastructure buyers actually feel storage. We rank a drive only after it has been through that bench; announced drives and vendor-supplied numbers do not qualify for a slot, no matter how good the specs look.

One consequence worth stating plainly: our test suite changed in April 2025. Drives reviewed before that ran VDBench or an earlier four-corners methodology, so their numbers are not apples-to-apples with the current field, and they sit in Also Tested with that noted rather than being ranked against newer results.

Enterprise SSD pricing is almost never public and moves with contract, volume, and capacity, so this page carries no value slot and no dollar figures. Where a drive earns a spot, it earns it on measured performance, endurance, efficiency, and form-factor fit.

Enterprise SSD FAQ

What is the best enterprise SSD in 2026?

For mixed-use workloads, the Micron 9550 MAX. It led 128K sequential write at 10,957.9 MB/s with the lowest write latency in our comparison group, was the only drive to scale past 10GB/s in 64K random write, and posted the lowest DLIO checkpoint times, all at 3 DWPD. Read-intensive fleets should look at the Kioxia CD9P-R instead, and capacity-driven tiers at the Micron 6600 ION.

TLC or QLC for the data center?

It depends on the write pattern, and the gap is wider than the datasheets suggest. TLC drives here carry 1 to 3 DWPD and scale small-block writes cleanly. QLC buys enormous capacity per watt, up to 245.76TB in a bay, but every QLC drive we tested showed a small-block write ceiling: the D5-P5336 did not scale 16K random writes at all, and the 6600 ION is rated at just 0.075 RDWPD for 4K random. For read-dominated tiers, object stores, and AI data lakes, QLC is the right economics. For databases and mixed workloads, stay on TLC.

How much endurance do I actually need?

Match DWPD to the workload rather than buying the highest number available. Read-heavy serving tiers run comfortably at 0.6 to 1 DWPD, which is where the high-capacity QLC drives and the CD9P-R sit. Databases, analytics, and AI training pipelines that checkpoint frequently want 3 DWPD, which is the 9550 MAX and 7600 MAX class. Also read the fine print on how endurance is specified: several drives quote different DWPD figures for sequential, 16K random, and 4K random traffic, and the 4K number is often dramatically lower.

Is PCIe Gen5 worth it, and what about Gen6?

Gen5 is the mainstream enterprise choice now, roughly doubling sequential ceilings to the 14GB/s range against about 7GB/s on the Gen4 drives here. Gen6 enterprise SSDs have been announced, with vendor claims up to 28GB/s, but we have not benchmarked one, so nothing Gen6 is ranked on this page. When a Gen6 drive reaches our lab, it will be ranked on measured results like everything else.

U.2, E1.S, or E3.S?

The industry is moving to EDSFF, and form factor increasingly decides the shortlist before performance does. U.2 remains the safe choice for existing 2.5-inch bays and is still where the largest capacities land. E3.S is where new Gen5 server designs are going, and E1.S suits dense, power-constrained nodes. The Micron 7600 MAX is the only drive on this page shipping in all three, which is why it holds the EDSFF slot. Check chassis compatibility before committing, especially for the taller E3.L 2T drives used at the highest capacities.

What is the largest enterprise SSD available?

The Micron 6600 ION at 245.76TB, which we have tested at that capacity. It is available in U.2 and E3.L, and DapuStor lists a 245TB SKU in the R6060 family as well. The practical consideration is not whether the capacity exists but whether your rack, your rebuild times, and your failure domains are ready for a quarter petabyte behind a single drive connector.

The post Best Enterprise SSDs in 2026: Lab-Tested Leaderboard appeared first on StorageReview.com.

NVIDIA has transitioned its 800-VDC power architecture from a forward-looking design concept into an active production roadmap. The initial MGX-compatible 800-VDC power rack is scheduled to enter production in the second half of 2026, delivering high-voltage direct current to AI compute racks while the broader data center facility continues operating on existing alternating current distribution. This hybrid deployment model allows standard AC infrastructure to power storage arrays, network switches, cooling equipment, and management systems, isolating the DC transition to dense GPU clusters.

The push toward 800 VDC marks an expansion of full-stack co-design beyond silicon and networking into core electrical and mechanical data hall infrastructure. The upcoming NVIDIA Vera Rubin platform, specifically the Vera Rubin NVL72 rack design, serves as the primary system driving this reference architecture. As rack power escalates across hardware generations, traditional low-voltage AC delivery reaches severe physical limits. Supplying hundreds of kilowatts at 415 or 480 VAC drives operating current to extreme levels. Managing that amperage requires massive copper busbars, heavy cabling runs, and complex physical routing that escalate facility costs and add mechanical strain inside the rack. Doubling the distribution voltage to 800 VDC significantly reduces operating current, conductor mass, physical congestion, and resistive transmission losses.

“800 VDC unlocks the compute performance and power density required for AI at scale,” said Vladimir Troy, vice president of data center infrastructure at NVIDIA.

Why 800 VDC, and Why Now

Current data center power topologies convert electricity multiple times between utility substations, uninterruptible power supplies, floor power distribution units, and rack power supplies. Each conversion stage introduces efficiency losses and adds to the hardware footprint. NVIDIA’s hybrid architecture is designed to slot into existing AC infrastructure and deliver 800 VDC to compute racks within the row, with no changes to the building’s electrical system required. These building blocks are captured in NVIDIA DSX reference designs, which give operators a system-level blueprint connecting power architecture, rack-scale compute, and facility infrastructure.

Three Deployment Topologies

The joint architecture establishes a phased transition strategy across three primary deployment topologies. The first phase, designated Option A, uses rack-adjacent Power Racks that act as sidecars to perform localized AC-to-800-VDC conversion. Scheduled for production in Q3 2026, this topology ingests existing 415/480 VAC feeds and supplies 800 VDC to adjacent compute racks through dedicated DC whips, requiring no modifications to upstream building electrical infrastructure.

Subsequent phases shift power conversion further upstream to optimize white space. Option B, targeted for Q3 2027, implements centralized Power Centers that distribute 800 VDC across compute rows via overhead or underfloor busways, eliminating the physical footprint of sidecars at each rack. Option C centralizes conversion at the data hall level using 4.8 MW DC Power Blocks to supply native 800-VDC compute racks. The long-term vision for Option C targets 2029 and incorporates Solid-State Transformers to convert medium-voltage AC directly to 800 VDC. This direct conversion eliminates conventional step-down transformers, switchgear, and secondary distribution stages.

Rack Density and the Road to 1 MW

The generational power-density roadmap matches this electrical scaling. Generation 1 systems target 145 kW per rack, Generation 2 systems target 330 kW to support platforms like the Vera Rubin NVL72, and Generation 3 reaches 570 kW, with future native 800-VDC architectures designed to support up to 1 MW per rack.

Standards, Safety, and the Partner Ecosystem

Standardization for this architecture is advancing through the Open Compute Project, where NVIDIA, Google, and Microsoft have jointly driven the 800-VDC effort NVIDIA previewed at the OCP Global Summit in October. Following their March 2026 white paper, the working group published the LVDC Solid-State Transformer Specification version 0.3 in July 2026. More than 80 equipment manufacturers and infrastructure providers are building hardware to this open specification, including ABB, Eaton, Schneider Electric, and Vertiv. Vertiv has developed an aligned 800-VDC reference architecture supporting rack sidecars, pod-level systems, and centralized DC distribution models.

To ensure facility safety at higher direct-current voltages, the specification defines mandatory protection standards, including high-resistance grounding topologies, continuous insulation monitoring, active arc fault detection, and mechanical connector interlocks that prevent disconnects under electrical load.

The commercial significance of the MGX 800-VDC rollout lies in its immediate viability as a retrofit path. Rather than requiring greenfield facilities purpose-built for full direct-current topologies, the hybrid architecture enables operators to deploy high-density Vera Rubin and MGX accelerated systems in existing AC facilities without major upstream electrical overhauls.

The post NVIDIA Moves 800-VDC Power Architecture From Concept to Production, Just Don’t Turn Off AC Power Yet appeared first on StorageReview.com.

The Dell Pro Precision 7 16 is the higher-end 16-inch mobile workstation in Dell’s current Pro Precision lineup. Our review unit is built around significantly more GPU power than the 5 16s Intel tested alongside it. It pairs a Series 3 Intel Core Ultra 9 386H, a 16-core Panther Lake processor with a 50 TOPS NPU, with NVIDIA RTX PRO 3000 Blackwell graphics and 12GB of GDDR7, 64GB of LPDDR5x at 8533 MT/s, and two 1TB Gen5 SSDs in RAID 0. The panel is a 16-inch UHD+ Tandem OLED with touch and a 120Hz variable refresh rate, running at 3840 x 2400. In testing, the discrete GPU rendered the Blender Monster scene at 1,555.53 samples per minute, roughly 2.7 times the integrated Arc Pro B390 in the 5 16s Intel, while the same configuration recorded the shortest battery runtime of any Pro Precision we have tested.

Compared with the Pro Precision 5 16s Intel, the 7 16 trades some portability and battery life for substantially more GPU performance and a higher-end feature set. Our configuration adds the RTX PRO 3000, a 96Wh battery, a 165 W adapter, a second Gen5 drive bay, and two Thunderbolt 5 ports alongside Thunderbolt 4. It is a better fit for engineers, designers, visualization specialists, video professionals, and other users whose applications benefit from dedicated NVIDIA graphics. The 7 16 also features a large haptic trackpad, which immediately differentiates the keyboard deck from the 5 16s.

The RTX PRO 3000 Blackwell is the mobile professional card in the NVIDIA Blackwell generation, with 12GB of GDDR7 and certified drivers for professional applications. Combined with the 50 TOPS NPU in the Core Ultra 9 386H, the system can run local AI work across the CPU, the discrete GPU, or the dedicated accelerator, and our Procyon testing covers all three paths. Coverage also includes professional graphics, rendering, content creation, storage, and battery.

The Dell Pro Precision 7 16 starts at $3,293, while the hardware in our review unit costs approximately $8,245. Dell’s public configurator does not currently allow us to reproduce the two-drive Gen5 RAID 0 configuration exactly, even though the otherwise equivalent single-drive configuration is priced at $7,795 in our brief. Applying the same $440 price difference for the second Gen5 SSD brings the as-shipped equivalent to that figure. Commercial buyers may see different pricing through account agreements and volume purchases, so Dell.com’s single-unit pricing is best used as a reference. The system is now available on the Dell Pro Precision 7 Series 16 product page.

Dell Pro Precision 7 16 Specifications

Specification	Dell Pro Precision 7 16 (PW716260)
Model	Dell Pro Precision 7 Series 16 (PW716260)
Processor	Intel Core Ultra 9 386H vPro Enterprise (Series 3), 16 cores / 16 threads, up to 4.9GHz, 50 TOPS NPU
Graphics	NVIDIA RTX PRO 3000 Blackwell, 12GB GDDR7
Memory	64GB LPDDR5x, 8533 MT/s, dual-channel, onboard, non-ECC
Storage	Two 1TB Gen5 SED-ready SSDs in RAID 0
Display	16″ UHD+ Tandem OLED 3840 x 2400, touch, 120Hz VRR, 500 nits, 100% DCI-P3, VESA DisplayHDR True Black 1000, anti-reflection
Camera	8MP HDR RGB + IR with User Presence Detection and ExpressSign-In
Wireless	Intel Wi-Fi 7 BE211 2×2, Bluetooth 6.0
Keyboard	Zero-lattice spill-resistant with mini-LED backlighting
Security	TPM 2.0, FIPS 140-3, TCG certified, post-quantum cryptography, chassis intrusion detection, SED storage, Windows Hello facial recognition
Battery	6-cell, 96Wh Long Life Cycle
Power	165W USB-C AC adapter
Ports	Two Thunderbolt 5, one Thunderbolt 4, HDMI 2.1, headset, SD card slot
Operating System	Windows 11 Pro, Copilot+ PC
Chassis	Aluminum and magnesium, Magnetite
Systems Management	Intel vPro Enterprise
Certifications	ENERGY STAR, EPEAT Gold with Climate+, TCO Certified
Warranty	36 months ProSupport Next Business Day Onsite Service after Remote Diagnosis
Price	$3,293 base / approximately $8,245 as shipped

Build and Design

The Dell Pro Precision 7 16 has a noticeably more substantial design than the 5 Series models, featuring aluminum and magnesium construction with Dell’s dark Magnetite finish. Our configuration remains fairly portable for a 16-inch workstation with RTX PRO 3000 graphics and a 96Wh battery, with Dell listing a starting weight of 4.78 lb. The OLED configuration measures 13.93 x 9.46 inches and ranges from 0.80 to 0.83 inches thick. Next to the Pro Precision 5 16s, the differences are easy to see: the 7 16 drops the numeric keypad in favor of a centered keyboard, a large haptic trackpad, and speaker grilles running along both sides of the deck.

The keyboard uses Dell’s zero-lattice design with mini-LED backlighting, and the centered layout makes good use of the wider 16-inch chassis. Removing the numeric keypad gives the main keyboard and trackpad a more symmetrical position in front of the display, while the large speaker grilles fill the space along either side. Dell uses two 2.5W woofers and two 2.5W tweeters for a total peak output of 10W, a much larger audio configuration than the basic stereo setup in the 5 16s. The keyboard also includes the dedicated Copilot key used across Dell’s current commercial lineup.

Below the keyboard is one of the Pro Precision 7 16’s more distinctive features: a large haptic trackpad that takes up a substantial portion of the palm rest. Instead of using the hinged mechanism found in a conventional trackpad, the click response is generated electronically, allowing the surface to provide a similar click response across a much larger area. Combined with the centered keyboard layout, it gives the 7 16 a noticeably different feel from the Pro Precision 5 systems we have been testing.

Dell equips the Precision 7 16 with an 8MP HDR RGB and IR camera above the display, with User Presence Detection and ExpressSign-In available for automatic Windows locking and sign-in behavior. The camera supports Windows Hello facial recognition and is paired with dual-array microphones, although this camera configuration does not include a physical privacy shutter.

The 16-inch Tandem OLED display is a substantial upgrade over the IPS panels used in the Pro Precision 5 systems, particularly for creators and other users working with high-resolution visual content. Our configuration has a 3840 x 2400 resolution, 120Hz variable refresh rate, 500-nit brightness, 100% DCI-P3 coverage, anti-reflection treatment, and VESA DisplayHDR True Black 1000 certification. The combination of 4K-class resolution and a 16:10 aspect ratio gives applications a large working area, while the 120Hz refresh rate makes scrolling, window movement, and cursor motion noticeably smoother than on a conventional 60Hz workstation panel. OLED also delivers very deep blacks and strong contrast, which is useful when working with HDR media, photography, and other color-sensitive content.

Our OLED configuration also supports touch, which works particularly well on a display this large for quick navigation, selecting items, or moving through visual content. The conventional clamshell hinge limits the system’s usefulness for extended pen-style input compared with a convertible, since the display cannot fold flat against the keyboard. For occasional direct interaction, however, having touch available on a 16-inch workstation panel is a useful addition, especially with the high-resolution OLED display.

For I/O and connectivity, the right side has a full-size SD card slot, one Thunderbolt 4 USB-C port, the 3.5mm headset connection, and a wedge-shaped lock slot. Having two Thunderbolt 5 ports plus a third Thunderbolt 4 connection gives the 7 16 an unusually strong USB-C layout, although Dell leaves out both USB-A and built-in Ethernet. So, users working with older USB peripherals or wired networks will need an adapter or dock, which is an important difference from the Pro Precision 5 16s and its wider selection of legacy ports.

Moving over to the left side, you’ll see that connectivity is heavily centered on Thunderbolt, with HDMI 2.1 alongside two Thunderbolt 5 USB-C ports. Those Thunderbolt 5 connections support Power Delivery and DisplayPort 2.1, with bandwidth reaching up to 120 Gbps for supported devices, giving the 7 16 a lot more external I/O bandwidth for fast storage, docks, and high-resolution displays. The two ports are positioned beside the HDMI connection, keeping most desk-oriented display and docking cables together on the same side.

The underside of the Pro Precision 7 16 features a broad two-row intake grille that feeds the dual-fan cooling system directly above it. The long, narrow rubber feet maintain clearance beneath the chassis for airflow, while the rear edge provides additional space for the cooling system to exhaust heat behind the display. Dell keeps access to the internals relatively simple, with the entire bottom panel secured by just four T5 screws. Once those are removed, the cover can be released from the recesses near the hinges and lifted away.

Here, there is direct access to the 96Wh battery, both SSD positions, and the dual-fan cooling system. Our configuration uses two 1TB Gen5 SSDs in RAID 0, with SSD1 and SSD2 in separate positions on opposite sides of the motherboard. Dell classifies the battery, SSDs, cooling fans, wireless card, and speaker assembly as customer-replaceable components, while hardware such as the heatsink, I/O boards, display assembly, touchpad, and keyboard assembly is intended for technician service. The 64GB of LPDDR5x memory is integrated into the system board, so the memory capacity must be selected when the system is ordered.

The Pro Precision 7 16 uses a much beefier cooling setup than the single-fan design in the 5 16s, which is appropriate given the RTX PRO 3000 and higher power demands. Two large fans fill the rear corners and work with a wide thermal assembly covering the CPU and GPU, with hot air exhausted out the back of the chassis. Both SSDs are positioned outside that central cooling area and have their own covers, so either drive is easy to reach once the bottom panel is off. The large 96Wh battery takes up most of the lower half, but Dell still leaves the storage, cooling system, and other serviceable components easily accessible.

Dell Pro Precision 7 16 Performance

Our review unit runs the Core Ultra 9 386H with NVIDIA RTX PRO 3000 Blackwell graphics, 64GB of LPDDR5x at 8533 MT/s, and two 1TB Gen5 SSDs in RAID 0 on Windows 11 Pro, with benchmarks tested on the Best Performance power mode. For battery life testing, we configure systems into Balanced power mode and set the screen brightness to 50%.

For comparables, we included the Dell Pro Precision 5 16s Intel (Core Ultra X9 388H, Arc Pro B390, 64GB), the Dell Pro Precision 5 16s AMD (Ryzen AI 9 HX PRO 475, Radeon 890M, 64GB), and the Lenovo ThinkPad P14s Gen 7 (Core Ultra 7 366H, RTX PRO 1000, 64GB) as the external workstation reference.

PCMark 10

PCMark 10 measures general system performance across everyday work such as web browsing, video conferencing, spreadsheets, writing, photo editing, and rendering. The overall score is supported by the Essentials, Productivity, and Digital Content Creation subscores, which indicate where a system’s strengths lie. Higher scores are better.

PCMark 10	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
Overall Score	8,901	9,994	8,627	9,083
Essentials	9,671	12,106	10,744	10,686
Productivity	16,381	15,505	14,574	16,501
Digital Content Creation	12,081	14,433	11,127	11,534

 

PCMark 10 was one of the weaker results for the Pro Precision 7 16, with an overall score of 8,901, trailing both the 5 16s Intel and the ThinkPad P14s Gen 7. Essentials came in at 9,671, the lowest result of the four systems, while Productivity was much closer at 16,381 compared with 16,501 for the ThinkPad. These general productivity workloads do not take full advantage of the RTX PRO 3000, which becomes much more important in the GPU-focused tests later on.

PCMark 10 Modern Office Battery

The PCMark 10 Modern Office battery test repeatedly runs common office tasks until the battery reaches the test’s cutoff point, in Balanced mode at 50% display brightness. This is a rundown of the whole system rather than a synthetic idle drain, so it tracks closely with what a full day of productivity work does to the battery. Longer runtimes are better.

Modern Office Battery	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
Runtime (higher is better)	11 hours 43 minutes	24 hours 43 minutes	15 hours 5 minutes	15 hours 52 minutes

 

This is the bill for the configuration. Eleven hours and 43 minutes is the shortest run of any Pro Precision we have tested, less than half the 24 hours and 43 minutes the 5 16s Intel manages, and that is with a 96Wh pack against the 5 series 70Wh. A discrete GPU, a 4K OLED at 120Hz, and two Gen5 drives all draw power from the same battery. It still clears a working day, but all-day-and-then-some belongs to the 5 series.

Geekbench 6

Geekbench 6 measures processor performance using a mix of common tasks, with separate scores for single-core and multi-core workloads, plus GPU compute scores through OpenCL and Vulkan. Higher scores are better. Our review unit’s CPU run was flagged as invalid by the benchmark’s tamper detection. We reviewed it and treated the flag as a false positive, so the scores are included below.

Geekbench 6	Dell Pro Precision 7 16 (RTX PRO 3000)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000)
CPU Single-Core	2,915	2,969	2,854	2,808
CPU Multi-Core	17,280	17,401	14,169	16,319
GPU OpenCL	131,492	57,998	37,520	87,537
GPU Vulkan	104,344	63,015	53,969	73,204

 

The two Intel Precisions are effectively tied on CPU, with the 5 16s Intel a fraction ahead at 2,969 and 17,401 against 2,915 and 17,280. Everything that separates them is on the GPU side, where the RTX PRO 3000 returns 131,492 in OpenCL, 2.3 times the Arc Pro B390 and 50% clear of the RTX PRO 1000 in the ThinkPad.

Geekbench 7

Geekbench 7 joins the suite alongside Geekbench 6 as comparison data builds. Its CPU scores are calibrated against a baseline of 2,500, set by the AMD Ryzen 7700, while GPU scores are calibrated against a baseline of 100,000, set by the NVIDIA GeForce RTX 4060. Higher scores are better, and double the score indicates double the performance. Because Geekbench 7 uses new workloads and new baselines, its scores are not comparable to Geekbench 6 results.

Geekbench 7	Dell Pro Precision 7 16 (RTX PRO 3000)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000)
CPU Single-Core	2,652	2,733	2,659	2,533
CPU Multi-Core	18,648	18,965	16,641	17,651
GPU OpenCL	96,919	54,247	31,874	75,340
GPU Vulkan	104,229	47,436	*	70,067
GPU CUDA	174,559	N/A	N/A	114,469

 

Geekbench 7 repeats the pattern. The 386H trails its own 5 16s sibling slightly on both CPU metrics, then leads every GPU metric, topping out at 174,559 in CUDA against 114,469 for the RTX PRO 1000. Neither integrated system can run CUDA at all, which is the practical argument for the card.

Cinebench 2026

Cinebench 2026 is the current release in the Cinebench line and the only version we report. It tests CPU and GPU performance using Maxon’s Redshift render engine. It is built on the latest Cinema 4D 2026 code and is designed to show whether a machine is stable under high CPU load, whether a notebook’s cooling can sustain longer render tasks, and how it handles demanding real-world 3D work. Because code and compiler changes accelerated scene rendering, Cinebench 2026 scores use an adjusted range and should not be compared to scores from previous Cinebench versions. Its GPU test supports current NVIDIA and AMD hardware but does not yet run on Intel integrated graphics, and our review unit’s CPU results were confirmed by a repeat run.

Cinebench 2026	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
CPU Single Thread	518	535	471	502
CPU Multiple Threads	3,873	4,618	4,767	4,492
GPU	51,517	N/A	5,667	34,437

 

The GPU result is the headline here: 51,517 for the Radeon 890M versus 5,667, with the Arc Pro B390 unable to run the test. The CPU side is less flattering. Single thread at 518 sits mid-pack, and the multi-thread 3,873 is the lowest of the four, 16% behind the 5 16s Intel on a comparable 16-core, 16-thread part. We repeated the run and reproduced the same figure.

7-Zip Compression

The built-in 7-Zip benchmark measures how quickly the processor can compress and decompress data using multiple threads, run with a 128MB dictionary across ten passes. Decompression tends to scale with thread count while compression leans on memory latency, so the two halves often tell different stories. Higher GIPS scores are better.

7-Zip 24.09 (GIPS)	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
Compressing	96.202	96.281	85.627	90.364
Decompressing	95.896	97.246	117.863	90.526
Total Rating	96.049	96.764	101.745	90.445

 

Compression was essentially tied between the two Intel Precisions, with the 7 16 scoring 96.202 GIPS compared with 96.281 GIPS from the 5 16s Intel. Decompression favors the 24-thread AMD system, which reached 117.863 GIPS and finished with the highest total rating at 101.745. The 7 16 placed second overall at 96.049, ahead of the ThinkPad but behind both the 5 16s Intel and AMD systems in total performance.

y-cruncher

y-cruncher measures how quickly the processor can calculate large numbers of digits of Pi, placing a heavy load on the CPU and memory subsystem. At the same time, the BBP runs to extract hexadecimal digits of Pi. Results are in seconds, so lower times are better. The AMD unit could not complete the 5-billion- and 10-billion-digit runs because its memory reservation for the integrated GPU reduces the available pool below the amount required for those sizes.

y-cruncher (seconds, lower is better)	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
Pi 1B	31.557	26.422	22.840	26.685
Pi 2.5B	101.524	80.463	64.242	76.787
Pi 5B	226.255	183.822	N/A	172.994
Pi 10B	500.875	407.658	N/A	392.083
Pi BBP 1B	1.639	1.635	1.100	1.621
Pi BBP 10B	21.512	19.036	12.277	18.200
Pi BBP 100B	291.628	234.797	140.284	219.969

 

y-cruncher is the clearest CPU loss in the review. The 7 16 is the slowest of the four on every computation, taking 500.875 seconds at Pi 10B, compared with 392.083 for the ThinkPad and 407.658 for the 5 16s Intel, and the BBP runs repeat it exactly. There is no GPU component here, so the discrete card contributes nothing, and the sustained-clock behavior is left exposed.

Blender

The Blender benchmark measures rendering performance using three different 3D scenes: Monster, Junkshop, and Classroom. Results are reported in samples per minute, so higher scores are better. We test on both the CPU and GPU. Scores are not comparable across Blender versions, so we have trimmed the older releases from the suite and report the current Blender 5.2 results here. The GPU figures represent each system’s fastest renderer: the discrete card on the two NVIDIA systems and the integrated GPU on the 5 16s pair.

Blender 5.2 (samples/min)	Dell Pro Precision 7 16 (RTX PRO 3000)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000)
GPU
Monster	1,555.53	572.84	123.00	922.95
Junkshop	1,257.01	468.85	101.21	795.65
Classroom	1,000.62	420.48	82.84	625.62
CPU
Monster	137.49	138.88	130.86	131.08
Junkshop	101.34	94.97	99.74	97.76
Classroom	69.26	65.39	74.18	68.45

 

This is what the card is for. The RTX PRO 3000 renders Monster at 1,555.53 samples per minute, 2.7 times the Arc Pro B390 and 69% ahead of the RTX PRO 1000, and the margin holds across Junkshop and Classroom. CPU rendering is a different picture, with all four within a few percent of each other; the 7 16 takes Junkshop at 101.34 and loses to the 24-thread AMD part in Classroom.

LuxMark

LuxMark measures GPU compute performance by rendering complex scenes through OpenCL, based on LuxCoreRender. We run the Food and Hall scenes on all available OpenCL devices in each system, so single-GPU systems are scored on that GPU while the dual-GPU systems render on the discrete and integrated GPUs together, as noted in the column headers. Higher scores are better.

LuxMark v4	Dell Pro Precision 7 16 (RTX PRO 3000 + iGPU)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000 + iGPU)
Hall	19,300	3,505	2,058	11,342
Food	7,624	1,713	1,034	4,103

 

LuxMark scales the same way. Hall at 19,300 is 5.5 times the Arc Pro B390 and 70% ahead of the ThinkPad, and Food widens that to 4.5 times and 86%. These are the margins that justify the chassis for anyone running OpenCL renderers.

V-Ray

Chaos V-Ray measures ray-traced rendering throughput, reported in vpaths, where higher is better. We run the CUDA-compatible engine on every system so results remain comparable to notebooks with only integrated graphics; on systems without a discrete GPU, that path executes on the integrated graphics, even though V-Ray reports the processor name in its device field. On dedicated NVIDIA hardware, we also capture V-Ray’s RTX engine, which engages the card’s ray tracing cores and is reported separately.

V-Ray GPU (vpaths)	Dell Pro Precision 7 16 (RTX PRO 3000)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000)
CUDA Engine	2,469	946	1,039	1,568
RTX Engine	3,891	N/A	N/A	2,589

 

V-Ray shows a significant advantage for the two systems with dedicated NVIDIA graphics, especially when the RTX engine is used. The Pro Precision 7 16 reached 3,891 vpaths in the RTX test, compared with 2,589 for the RTX PRO 1000, while its CUDA result of 2,469 was 57% higher than that of the ThinkPad. The integrated systems can use only the CUDA path here, with both finishing at less than half of the 7 16’s score.

3DMark CPU Profile

The 3DMark CPU Profile benchmark measures CPU performance at fixed thread counts, from a single thread up to the maximum available, showing how performance scales as more cores are engaged. Higher scores are better.

3DMark CPU Profile	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
Max Threads	10,557	10,748	9,267	10,500
8 Threads	5,996	6,818	6,573	6,550
4 Threads	4,261	4,483	4,238	4,219
1 Thread	1,185	1,210	1,178	1,165

 

The three Intel systems finish within about 2% of each other at max threads, with the 7 16 second at 10,557. The eight-thread result is the outlier, with 5,996 the lowest in the group by a clear margin, even behind the ThinkPad. Single-thread is a four-way tie inside 4%.

3DMark Storage and Blackmagic Disk Speed Test

3DMark Storage measures how an SSD performs during gaming-related tasks such as loading games, installing software, saving progress, and moving game files. Blackmagic Disk Speed Test measures an SSD’s sequential read and write speeds using large media files. Higher is better in both.

Storage	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
3DMark Storage	2,189	2,804	2,325	3,094
Blackmagic Write (MB/s)	7,387.8	7,660.2	5,070.5	8,262.3
Blackmagic Read (MB/s)	6,776.2	8,418.1	5,071.2	8,511.5

 

On paper, the RAID 0 array does not deliver what the configuration implies. 3DMark Storage returns 2,189, the lowest of the four comparison systems and well behind the ThinkPad’s single drive at 3,094, and SPECworkstation independently puts Storage at 1.30, second lowest. The sequential numbers are healthier in absolute terms, with Blackmagic writes of 7,387.8 MB/s and reads of 6,776.2 MB/s, though both still trail the single-drive 5 16s Intel and ThinkPad. Those results are consistent with how striping behaves rather than a sign of a faulty array. Lightly threaded storage tests split a single queue across two drives, which adds overhead without adding parallelism, and RAID 0’s benefits only appear as demand climbs. We verified that the array is configured and performing as intended.

As always, we tested the system as it shipped, and this one shipped in RAID 0, which is a somewhat unusual factory choice. Buyers are not locked into it. The two Gen5 drives can be split into separate OS and data volumes, mirrored in RAID 1 for redundancy, or simply run as independent drives, and several of those layouts may net better storage performance in scenarios like the ones tested here. The catch is that moving away from the factory RAID 0 means reinstalling the operating system, so the storage layout is worth deciding at deployment time rather than after the machine is in service.

Blackmagic RAW Speed Test

The Blackmagic RAW Speed Test measures how many frames per second a system can decode Blackmagic RAW video on the CPU and on the GPU. We quote the 8K results at 12:1 compression, and higher is better.

Blackmagic RAW Speed Test	Dell Pro Precision 7 16 (RTX PRO 3000)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000)
8K 12:1 CPU (fps)	77	77	74	77
8K 12:1 GPU (fps)	137	87	49	96

 

CPU decode is identical across the three Intel systems at 77fps. The GPU path is where the card shows up, reaching 137fps, 57% ahead of the Arc Pro B390 and 43% ahead of the RTX PRO 1000.

Topaz Video AI

The Topaz Video AI benchmark measures AI video upscaling and frame-interpolation performance in frames per second across the application’s enhancement models, run here at 1080p input, where higher is better. The 16X Slowmo Aion model failed to complete on our review unit and on the 5 16s Intel, while the 5 16s AMD ran it without issue.

Topaz Video AI (fps)	Dell Pro Precision 7 16 (RTX PRO 3000)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000)
Artemis 1X / 2X / 4X	12.38 / 8.94 / 3.15	6.24 / 5.21 / 1.90	3.72 / 2.21 / 0.78	5.17 / 3.19 / 1.12
Iris 1X / 2X / 4X	13.06 / 7.49 / 2.32	5.29 / 3.19 / 0.96	4.91 / 2.72 / 0.92	5.91 / 3.12 / 1.01
Proteus 1X / 2X / 4X	12.06 / 8.60 / 2.66	6.45 / 6.09 / 2.47	3.99 / 2.70 / 1.18	4.78 / 3.14 / 1.05
Gaia 1X / 2X / 4X	3.73 / 2.66 / 2.02	3.30 / 2.26 / 1.51	1.87 / 1.34 / 0.96	1.62 / 1.13 / 0.79
Nyx 1X / 2X	3.59 / 3.08	1.56 / 1.57	1.87 / 1.53	2.40 / 2.09
Hyperion HDR 1X	15.58	3.23	11.48	14.56
4X Slowmo Apollo / APFast	18.04 / 30.60	8.48 / 22.00	6.08 / 17.59	10.39 / 29.94
16X Slowmo Aion	DNF	DNF	9.05	N/A

 

The 7 16 leads every model it completed, roughly doubling the 5 16s Intel on Artemis and Iris and better than doubling the ThinkPad. The one gap in the row is the 16X Slowmo Aion model, which did not finish (the same failure we recorded on the 5 16s Intel).

UL Procyon AI Text Generation

The Procyon AI Text Generation Benchmark streamlines LLM performance testing by providing a concise, consistent evaluation method. It enables repeated testing across four local models (Phi, Mistral, Llama3, and Llama2) while minimizing the complexity of large models and the number of variables. Developed with AI hardware leaders, it optimizes the use of local AI accelerators to deliver more reliable, efficient performance assessments. All four systems ran the models through ONNX Runtime with DirectML on their GPUs.

Procyon AI Text Generation	Dell Pro Precision 7 16 (RTX PRO 3000)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000)
Phi	2,244	893	434	1,618
Mistral	2,062	646	403	1,397
Llama3	1,859	669	357	1,252
Llama2	1,990	750	390	DNF

 

Local LLM inference is the widest AI margin in the review. The 7 16 leads every model, 2,244 on Phi against 893 for the Arc Pro B390, and it still holds a 39% lead over the ThinkPad on the same test. The ThinkPad did not complete Llama2.

UL Procyon AI Computer Vision

The Procyon AI Computer Vision Benchmark measures AI inference performance across CPUs, GPUs, and dedicated accelerators using a range of neural networks, evaluating tasks such as image classification, object detection, segmentation, and super-resolution with models including MobileNet V3, Inception V4, YOLO V3, DeepLab V3, Real ESRGAN, and ResNet 50. The WinML runs use float32 on CPU and GPU, giving a like-for-like view across vendors. We also run the newer Computer Vision 2 suite through each vendor’s native path: OpenVINO in int8 on the Intel NPUs and fp16 on their iGPUs, TensorRT in fp16 on the RTX PRO 3000, and Ryzen AI on the AMD NPU; those results are listed separately since precision and runtime differ by platform. Higher scores are better.

Procyon AI Computer Vision (WinML)	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
CPU	122	143	114	134
GPU	553	410	245	426

 

Procyon AI Computer Vision 2 (native runtimes)	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD
NPU (int8)	1,614	1,630	1,189
iGPU (fp16)	837	1,529	N/A
dGPU (TensorRT fp16)	3,470	N/A	N/A

 

The WinML float32 path puts the 7 16 last on CPU at 122 and first on GPU at 553, 35% ahead of the Arc Pro B390 and 30% ahead of the ThinkPad. Running each vendor native runtime instead, the discrete card more than doubles anything the integrated silicon manages at 3,470 through TensorRT. The NPU result of 1,614 is a virtual tie with the 5 16s Intel, which is expected since both use the same 50 TOPS engine, and the 837 on integrated graphics reflects the smaller iGPU in the 386H rather than the Arc Pro B390.

UL Procyon AI Image Generation

The Procyon AI Image Generation Benchmark provides a consistent method for measuring AI inference performance from low-power NPUs to high-end GPUs, with three tests: Stable Diffusion XL FP16 for high-end GPUs, Stable Diffusion 1.5 FP16 for moderately powerful GPUs, and Stable Diffusion 1.5 INT8 for low-power devices. The benchmark uses the optimal inference path for each platform: OpenVINO on Intel systems, the AMD-optimized DirectML pipeline on Radeon GPUs, and TensorRT on NVIDIA GPUs. The INT8 test uses Intel’s quantized SD 1.5 model where supported; the AMD pipeline does not offer a comparable quantized run. Our review unit runs this suite through TensorRT on the RTX PRO 3000. New with this round, we also ran the INT8 workload on the Intel NPU.

Procyon AI Image Generation	Dell Pro Precision 7 16 (RTX PRO 3000)	Dell Pro Precision 5 16s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Lenovo ThinkPad P14s Gen 7 (RTX PRO 1000)
SD 1.5 FP16	1,424	638	316	943
SD 1.5 INT8	17,287	7,778	2,855	12,403
SDXL FP16	1,273	738	200	765
SD 1.5 INT8 NPU	2,870	2,881	N/A	N/A

 

TensorRT leads every row. The INT8 workload returns 17,287 against 12,403 for the ThinkPad and 7,778 for the Arc Pro B390, and SDXL comes in at 1,273, where the two integrated systems sit at 738 and 200. The NPU score of 2,870 is effectively identical to the 5 16s Intel score of 2,881, again, the same Intel accelerator doing the same work.

SPECviewperf 15

SPECviewperf 15 measures graphics performance using viewsets derived from professional applications in CAD, 3D modeling, rendering, engineering, and medical visualization. Our review unit ran the suite at its native 4K and produced scores in all 11 viewsets. The comparison systems ran at 1080p, which is not comparable to a 4K run, so this table includes only the review unit.

SPECviewperf 15 (4K)	Dell Pro Precision 7 16 (RTX PRO 3000)
3dsmax-08	32.38
blender-01	30.20
catia-07	42.71
creo-04	114.70
energy-04	26.28
enscape-01	11.65
maya-07	71.18
medical-04	50.48
snx-05	82.27
solidworks-08	63.99
unreal_engine-01	50.42

 

Run at the panel native 4K, the 7 16 completed all 11 viewsets. creo-04 at 114.70 and snx-05 at 82.27 are the strongest results, and enscape-01 at 11.65 is the weakest. Because the comparison systems ran these viewsets at 1080p, these figures should not be read against the other tables in this review.

SPECworkstation 4

SPECworkstation 4 measures workstation performance across CPU, graphics, storage, AI, product design, engineering, financial services, and other professional workloads, using real applications grouped into seven industry verticals. Higher scores are better, and N/A means the system did not complete every workload required for that category. The standalone 7-Zip workload inside SPECworkstation failed on our review unit and on the 5 16s AMD, which suppresses the CPU subsystem and Productivity and Development scores for both; reruns reproduced the same failure on each, and our separate 7-Zip section above covers that ground.

SPECworkstation 4	Dell Pro Precision 7 16	Dell Pro Precision 5 16s Intel	Dell Pro Precision 5 16s AMD	Lenovo ThinkPad P14s Gen 7
Hardware Subsystems
Graphics	6.54	2.67	2.65	4.51
Accelerator	3.88	2.25	2.31	3.26
Storage	1.30	1.80	1.00	1.67
CPU	N/A	1.34	N/A	1.35
Industry Verticals
AI & Machine Learning	1.76	1.47	1.45	1.65
Energy	1.87	1.55	1.38	1.69
Financial Services	1.03	0.94	1.29	0.96
Life Sciences	2.03	1.84	1.47	1.82
Media & Entertainment	1.87	1.60	1.48	1.81
Product Design	1.87	1.75	1.41	1.89
Productivity & Development	N/A	1.35	N/A	1.34

 

The Graphics subsystem score of 6.54 is 2.4 times that of the integrated system and 45% higher than the ThinkPad, while the 7 16 leads four of the six industry verticals it completed. Product Design goes to the ThinkPad by a narrow 1.89 to 1.87 margin, while the AMD system leads Financial Services. The CPU and Productivity and Development scores are N/A because the standalone 7-Zip workload failed, the same failure recorded on the 5 16s AMD. Storage finished at 1.30, the second-lowest result of the group, which also follows the relatively weak storage performance recorded in 3DMark Storage.

Conclusion

The Dell Pro Precision 7 16 is built around a very different set of priorities than the Pro Precision 5 16s Intel, and the RTX PRO 3000 is the reason to spend the extra money. GPU rendering was where the gap became especially large, with the 7 16 producing 1,555.53 samples per minute in Blender Monster compared with 572.84 from the Arc Pro B390, with similarly large gains in Junkshop and Classroom. V-Ray, local AI text generation, Topaz Video AI, and Blackmagic RAW also benefited heavily from the discrete GPU, including 137 fps in the BRAW GPU test compared with 87 fps from the 5 16s Intel. CPU performance does not provide the same advantage; however, the less expensive 5 16s Intel was as fast or faster in several processor-heavy tests, including Geekbench, Cinebench, and y-cruncher.

The additional hardware also comes with a substantial battery penalty, even with the 7 16 carrying a much larger 96Wh brick. Its 11 hours and 43 minutes in the Modern Office test is less than half the 24 hours and 43 minutes delivered by the 5 16s Intel, which is a major difference for anyone regularly working away from a desk. There are some excellent upgrades beyond the RTX PRO 3000, including the 3840 x 2400 Tandem OLED with 120Hz VRR, the large haptic trackpad, two Thunderbolt 5 ports, a full-size SD card reader, and room for two Gen5 SSDs. Dell also keeps the battery, SSDs, and cooling fans accessible once the bottom cover is removed, although the 64GB of LPDDR5x memory is soldered and cannot be upgraded later. Our RAID 0 configuration was also light in storage testing, as the two Gen5 drives did not outperform the single-drive systems, contrary to what their configuration would suggest.

Pricing ultimately makes the choice between these two systems fairly easy. Our exact dual-drive configuration works out to approximately $8,245 based on the hardware Dell shipped, putting the 7 16 roughly $1,900 to $2,300 above the 5 16s Intel, depending on how the configurations are compared. Engineers, 3D artists, visualization professionals, video editors, and users running GPU-accelerated local AI workloads can see enormous performance gains from the RTX PRO 3000, and those workloads give the added cost a valuable purpose. Users spending most of their time in CPU-heavy applications, development work, office workloads, or anything that does not benefit substantially from NVIDIA graphics should look closely at the 5 16s Intel instead. It offers similar or better CPU performance, more than twice the battery life, and a considerably lower price, while the Pro Precision 7 16 earns its premium specifically for users who can put its much faster GPU to work.

For configuration options and current pricing, visit the Dell Pro Precision 7 Series 16 product page.

Leaderboard: The Dell Pro Precision 7 16 Intel ranks #19 on our Laptop Battery Life Leaderboard and appears in the field on our Best Mobile Workstations and Best Laptops for Local AI pages.

The post Dell Pro Precision 7 16 Intel Review: RTX PRO 3000 Blackwell in a Tandem OLED Workstation appeared first on StorageReview.com.

Updated August 14, 2026: Initial publication. Prices noted are as-tested, single-unit prices at review time; volume pricing differs, and this market moves quickly.

Every laptop ranked here has been through the StorageReview lab. Business laptops get judged on a different rubric than consumer machines: manageability, firmware security, serviceability, and battery endurance count alongside benchmark speed, and every system on this page was tested on all of it. Nothing is ranked from a spec sheet, and every pick links to the review holding the data.

The 2026 field is the strongest we have covered. Two machines cleared 26 hours of measured battery life, all three silicon vendors (Intel, AMD, and Qualcomm) now ship credible business platforms, and every current system carries an NPU. The picks below cover the seven roles corporate buyers actually hire laptops for.

At a Glance

Category	System	Platform	Standout Result	Full Review
Best Overall Business Laptop	Dell Pro 7 14 Intel	Core Ultra 7 366H, 64GB, vPro	PCMark 10 of 8,438 with 26 hr 18 min of battery at 2.80 lb	Pro 7 14 Intel Review
Best Battery Life	Dell Pro 5 14 Intel	Core Ultra X7 368H, Arc B390, 64GB LPCAMM2	26 hr 48 min, the longest laptop battery life we have measured	Pro 5 14 Intel Review
Best 2-in-1	HP EliteBook X Flip G1i	Core Ultra 7 258V, 32GB	24 hr 34 min, the longest convertible battery result in our lab	EliteBook X G1i Review
Best 16-inch	Dell Pro 5 16 AMD	Ryzen AI 9 HX PRO 470, 64GB SODIMM	103.9 GIPS in 7-Zip, the strongest sustained CPU in the field	Pro 5 16 AMD Review
Best AMD Business Laptop	HP EliteBook X G1a	Ryzen AI 9 HX 375, 64GB	Cinebench R23 multi-core of 21,013 with a 2.8K OLED touch display	EliteBook X G1a Review
Best Executive Thin-and-Light	Dell Pro 14 Premium	Core Ultra 7 268V, 32GB	2.52 lb with a Tandem OLED QHD+ panel and 13 hr 55 min of battery	Pro 14 Premium Review
Best ARM Business Laptop	HP EliteBook 6 G1q	Snapdragon X Plus, 45 TOPS NPU, 32GB	19 hr 35 min of battery at 3.17 lb	EliteBook 6 G1q Review

The Picks

Best Overall Business Laptop: Dell Pro 7 14 Intel

The Pro 7 14 Intel posted the highest PCMark 10 overall score of any business laptop through our lab at 8,438, and it did it in Dell’s thinnest Pro chassis at 2.80 pounds with 26 hours 18 minutes of measured battery life. The Core Ultra 7 366H and 64GB of LPDDR5x keep productivity workloads quick (Geekbench 6 multi-core of 16,787), and the Gen5 SSD turned in 8.4GB/s reads in Blackmagic.

The manageability column is fully checked: Intel vPro with AMT, Hardware Shield, Dell SafeBIOS, and testing beyond standard MIL-STD requirements. Our review’s one caution is that it trades sustained CPU performance for portability, so heavy compile or render jobs belong on a bigger machine. It listed at $5,600 single-unit as tested at review time.

Review: Dell Pro 7 14 Intel Review: 26 Hours of Battery in Dell’s Thinnest Pro Laptop

Best Battery Life: Dell Pro 5 14 Intel

The Pro 5 14 Intel holds the longest battery result we have ever measured in a laptop, 26 hours 48 minutes in PCMark 10 Modern Office, and it tops our Laptop Battery Life Leaderboard outright. It is no efficiency-only machine either: the Core Ultra X7 368H pairs with Arc B390 graphics that put up real GPU numbers (Blender Monster at 366.6 samples per minute), and 7-Zip hit 89.4 GIPS.

It is also the serviceable one: 64GB of LPCAMM2 memory is replaceable rather than soldered, and the Gen5 SSD is user-accessible. With vPro, FIPS 140-3 TPM, ControlVault 3+, and quantum-resistant BIOS verification, this is the fleet machine that runs three shifts. $5,492 single-unit as tested at review time.

Review: Dell Pro 5 14 Intel Review

Best 2-in-1: HP EliteBook X Flip G1i

The Flip G1i is the longest-running convertible we have tested at 24 hours 34 minutes, and its clamshell sibling, the EliteBook X G1i, is right behind at 23 hours 31 minutes. Lunar Lake’s Core Ultra 7 258V delivers strong single-core response (Geekbench 6 at 2,840) at very low power, and the 48 TOPS NPU qualifies the pair for Copilot+ features.

HP’s security story is the differentiator: a built-in Endpoint Security Controller, automatic threat isolation, remote locate and lock, and a three-year HP Wolf Pro Security license in the box. Base pricing at review was $1,499 for the Flip, which made it one of the most accessible machines on this page.

Review: HP EliteBook X G1i and Flip G1i Review

Best 16-inch: Dell Pro 5 16 AMD

For desk-first buyers, the Pro 5 16 AMD is the strongest sustained performer here: 103.9 GIPS in 7-Zip and a Cinebench R23 multi-core of 18,764, both the best in this field, from the Ryzen AI 9 HX PRO 470. The 16-inch WQXGA panel at 500 nits brings a dedicated numeric keypad, and 64GB of standard SODIMM memory keeps upgrades cheap.

Battery is still respectable at 15 hours 22 minutes, and AMD PRO manageability plus Dell SafeBIOS and SafeID cover the IT checklist. Our review called it the best fit for buyers who want a larger screen and stronger performance during long CPU-heavy workloads, and the data backs that up. $4,648 single-unit as tested at review time.

Review: Dell Pro 5 16 AMD Review

Best AMD Business Laptop: HP EliteBook X G1a

The EliteBook X G1a is the strongest 14-inch AMD machine we have tested in this class: its Ryzen AI 9 HX 375 posted a Cinebench R23 multi-core of 21,013, the best multi-core number on this page, alongside a 55 TOPS NPU and 64GB of memory. The 2880×1800 OLED touch display is the nicest panel in the field, and the whole package starts at 3.3 pounds.

Battery lands mid-pack at 10 hours 48 minutes, the tradeoff for that OLED and sustained CPU headroom. HP Wolf Pro Security Edition and a full Windows Hello stack cover the business requirements. At $2,749 as tested at review time, it undercut most of this page while outrunning it in multi-core work.

Review: HP EliteBook X G1a Review

Best Executive Thin-and-Light: Dell Pro 14 Premium

The Pro 14 Premium is the boardroom pick: 2.52 pounds in its lightest configuration, a Tandem OLED QHD+ display, and enough performance to stay out of the way, with a PCMark 10 overall of 7,175 and 13 hours 55 minutes of battery. Lunar Lake’s Core Ultra 7 268V favors responsiveness and efficiency over sustained grunt, which matches how executive machines actually get used.

vPro Enterprise support keeps it manageable in the fleet. Our review positioned it as a direct alternative to Lenovo’s ThinkPad X9 14 Aura Edition, and notably, the Dell keeps the business features the Lenovo dropped. Around $2,500 as tested at review time.

Review: Dell Pro 14 Premium Review: High-End Performance Meets Executive Style

Best ARM Business Laptop: HP EliteBook 6 G1q

The EliteBook 6 G1q is the ARM option that finally makes business sense: 19 hours 35 minutes of battery from a 3.17-pound chassis, with the Snapdragon X Plus turning in a Geekbench 6 multi-core of 11,408. Its 45 TOPS Hexagon NPU also earned it the Best Thin-and-Light NPU System spot on our Best Laptops for Local AI leaderboard, where it ran 3B and 4B models locally at usable speeds.

The considerations are the usual ARM ones: check your application compatibility list before a fleet rollout, and note that PCMark 10 does not run on this platform, so cross-fleet comparison leans on Geekbench and 7-Zip. HP Sure Click and Sure Sense cover endpoint security. About $3,100 as configured at review time.

Review: HP EliteBook 6 G1q Review: All-Day Power in a Lightweight Laptop

Also Tested

These systems went through the same lab process. Some sit just outside our trailing two-year data window, and some fall short of a category spot for reasons the reviews spell out.

• Dell Pro 7 14 AMD: the portable AMD PRO option at 2.80 pounds with 19 hours 28 minutes of battery; edged out here by the EliteBook X G1a’s multi-core lead and lower price.
• Dell Pro Rugged 14: MIL-STD-810H and IP53 in a semi-rugged chassis with an 1,100-nit display; it skips our standard battery test, so it is not ranked, but it is the field-work answer in this lineup.
• Lenovo ThinkPad X9 14 Aura Edition: a sleek 15 hour 10 minute OLED machine, but our review found it reads as a consumer laptop; no vPro, SmartCard, or WWAN options.
• HP EliteBook 1040 G11: our former top pick for a high-end 14-inch business laptop (22 hours 8 minutes of battery); its August 2024 review has now aged past our two-year data window.
• Lenovo ThinkPad X1 Carbon Gen 12: the 2.42-pound benchmark for premium business ultralights; July 2024 review, now outside the data window.
• Lenovo ThinkPad X1 2-in-1 Gen 9: our previous convertible recommendation with ThinkShield and MIL-SPEC 810H; July 2024 review, now outside the data window.
• Dell Latitude 7450 Ultralight: 2.33 pounds, the lightest business laptop we have tested, with noted chassis flex; July 2024 review, now outside the data window.

How We Rank

Every laptop on this page ran the same lab suite: PCMark 10 for overall productivity, Geekbench 6, Cinebench, 7-Zip, and y-cruncher for compute, storage benchmarks on the shipping SSD, and PCMark 10 Modern Office for battery, measured to shutdown. Business qualification matters as much as speed: we weight manageability platforms (Intel vPro, AMD PRO), firmware security, and serviceability alongside the numbers, and machines without those features do not rank here regardless of performance.

Rankings draw on a trailing two-year window of reviews so the field reflects current silicon; older systems move to Also Tested with their review dates noted. Prices move quickly, so any dollar figure here is the as-tested, single-unit price at review time, which is also not what volume buyers pay. We do not make value claims without checking current vendor configurator pricing, and when we do, we date the check.

Business Laptop FAQ

What is the best business laptop in 2026?

The Dell Pro 7 14 Intel is our Best Overall pick: the highest PCMark 10 score we have measured in a business laptop (8,438), 26 hours 18 minutes of battery, and a full vPro and SafeBIOS manageability stack at 2.80 pounds. If battery is the priority, its sibling Pro 5 14 Intel runs 30 minutes longer and adds replaceable memory; if sustained CPU work is the priority, the Dell Pro 5 16 AMD leads the field.

What actually makes a laptop a business laptop?

Manageability and security, not the badge. Fleet tools like Intel vPro with AMT or AMD PRO with DASH let IT departments deploy, patch, and remotely disable machines at scale; firmware protections like Dell SafeBIOS and HP’s Endpoint Security Controller guard the boot chain; and TPM 2.0, smart card, and WWAN options round out the checklist. Lenovo’s ThinkPad X9 14 is the cautionary example: a fine laptop our review ultimately classified as consumer because it dropped those features.

Which business laptop has the longest battery life?

The Dell Pro 5 14 Intel, at 26 hours 48 minutes in PCMark 10 Modern Office, the longest laptop battery result in StorageReview lab history. Business machines dominate our Laptop Battery Life Leaderboard generally: efficiency-first silicon and big batteries are what corporate buyers reward.

Intel, AMD, or Snapdragon for a business fleet?

All three now field credible options, which was not true two years ago. Intel brings the deepest vPro management install base and, in Lunar Lake and the Core Ultra Series 3, exceptional battery results. AMD PRO machines counter with multi-core throughput (the two Ryzen AI 9 HX systems here post the best Cinebench numbers on the page). Snapdragon delivers the best performance-per-watt in the thin-and-light class, with the standard caveat of app-compatibility validation before rollout.

Can business laptops run local AI?

Increasingly, yes. Every 2025-2026 machine on this page carries an NPU between 45 and 55 TOPS, enough for Copilot+ features and small local models, and the EliteBook 6 G1q crosses over to our Best Laptops for Local AI leaderboard outright. For larger models you want discrete GPU or big unified memory machines, which that page covers in depth.

The post Best Business Laptops in 2026: Lab-Tested Leaderboard appeared first on StorageReview.com.

Updated August 14, 2026: Dell Pro Precision 7 16 Intel added to Also Tested following its review. Originally published earlier today. Prices noted are as tested at review time; this market moves quickly, so check vendor configurators before buying.

Every laptop ranked here has been through the StorageReview lab. We measure local AI performance directly: UL Procyon AI Text Generation runs the same models on every system that can hold them, and where the hardware warrants we go deeper with LM Studio and Ollama. Nothing on this page is ranked from a spec sheet, and every pick links to the review holding the data.

There are two roads to running AI models on a laptop in 2026. Discrete NVIDIA RTX PRO GPUs deliver the fastest tokens per second, but the model has to fit inside the card’s VRAM, 8GB to 24GB across this field. Unified and shared memory designs, led by AMD’s Ryzen AI Max+ (Strix Halo) and Intel’s Core Ultra shared LPCAMM2 platforms, trade raw speed for capacity: the GPU can borrow most of system memory, so far larger models load at all. The picks below cover both roads, plus the NPU-first efficiency class.

At a Glance

Category	System	AI Engine	Standout Result	Full Review
Best Overall Laptop for Local AI	Dell Pro Max 18 Plus	RTX PRO 5000 Blackwell 24GB / 128GB CAMM2	185 tok/s (Phi, Procyon), fastest laptop we have tested	Pro Max 18 Plus Review
Best for Large Models	HP ZBook Ultra G1a 14	Ryzen AI Max+ PRO 395, up to 96GB assignable unified memory	Loaded DeepSeek-R1 70B; Gemma 3 27B at 9 tok/s	ZBook Ultra G1a Review
Best 16-inch Balance	Dell Pro Max 16 Plus	RTX PRO 5000 Blackwell 24GB / 128GB CAMM2	179 tok/s (Phi) with 6 hr 21 min of battery	Pro Max 16 Plus Review
Best Ultraportable	Lenovo ThinkPad P14s Gen 7	RTX PRO 1000 Blackwell 8GB / 64GB LPCAMM2	55 tok/s (Phi) at 3.59 lb	P14s Gen 7 Review
Best Without a Discrete GPU	Dell Pro Precision 5 14s Intel	Intel Arc Pro B390 iGPU, 64GB shared LPCAMM2	Ran models that exceed 8GB VRAM, 23 hr 50 min of battery	Pro Precision 5 14s Review
Best Thin-and-Light NPU System	HP EliteBook 6 G1q	Snapdragon X Plus, 45 TOPS Hexagon NPU, 32GB	Llama 3.2 3B at 37 tok/s in LM Studio	EliteBook 6 G1q Review

The Picks

Best Overall Laptop for Local AI: Dell Pro Max 18 Plus

The Pro Max 18 Plus posted the fastest local AI numbers of any laptop through the StorageReview lab. In UL Procyon AI Text Generation, its RTX PRO 5000 Blackwell (24GB GDDR7) pushed Phi to 185.1 tokens per second, Mistral to 140.5, and Llama 3 to 119.7, with time to first token under 0.35 seconds on every model. The Core Ultra 9 285HX and 128GB of CAMM2 memory keep the rest of the pipeline out of the way, and the 24GB of VRAM comfortably holds 20B-class quantized models.

The tradeoffs are exactly what the chassis suggests: 7.17 pounds and 3 hours 39 minutes of measured battery life, the shortest runtime in our laptop dataset. This is a deskside machine that happens to fold. It listed at $9,245 as tested at review time. If local inference speed is the whole question, this is the answer.

Review: Dell Pro Max 18 Plus: Blackwell RTX PRO 5000 Performance To Go

Best for Large Models: HP ZBook Ultra G1a 14

Speed is one axis; capacity is the other. The ZBook Ultra G1a pairs AMD’s Ryzen AI Max+ PRO 395 with 128GB of LPDDR5X unified memory, up to 96GB of it assignable to the Radeon 8060S GPU. That pool let us load models no discrete-GPU laptop can touch: DeepSeek-R1 70B and QwQ 32B both ran locally in LM Studio and Ollama, and Gemma 3 27B generated at 8.96 tokens per second with prompt processing at 73 tokens per second.

Its Procyon numbers trail every RTX PRO machine here (Phi at 65 tokens per second), so this is not the pick for fast chat on small models. It is the pick when the model itself is the point, and it does that at 3.3 pounds with 10 hours 35 minutes of battery. This is the same unified-memory story AMD’s Strix Halo tells on our desktop side, folded into a 14-inch chassis.

Review: HP ZBook Ultra G1a 14 Review

Best 16-inch Balance: Dell Pro Max 16 Plus

The Pro Max 16 Plus runs the same RTX PRO 5000 Blackwell 24GB as the 18 Plus and gives up almost nothing: Phi at 178.6 tokens per second, Mistral at 134.2, Llama 3 at 114.7, roughly 96 percent of the flagship’s throughput. In exchange it starts at 5.63 pounds instead of 7.17 and nearly doubles the battery result at 6 hours 21 minutes.

For most people who want serious local inference in a bag, this is the better buy of the two. It also holds the Best Overall Mobile Workstation spot on our Best Mobile Workstations leaderboard, so the AI speed comes with the full professional application stack already validated.

Review: Dell Pro Max 16 Plus Review: Built for Long Days and Big Jobs

Best Ultraportable: Lenovo ThinkPad P14s Gen 7

At 3.59 pounds, the P14s Gen 7 is the lightest laptop here with a discrete Blackwell GPU. The RTX PRO 1000 (8GB GDDR7) turned in Phi at 55.1 tokens per second and Mistral at 40.3, real interactive speeds for 7B-class models, and Panther Lake’s 50 TOPS NPU makes it a Copilot+ machine. The battery result of 15 hours 52 minutes means the AI capability does not cost you the workday.

Know the ceiling before you buy: our Llama 2 13B run did not finish because that model wants about 12GB of graphics memory through the Procyon path and the GPU has 8GB. For 7B-and-under quantized models this is a terrific carry; for bigger ones, look up the page.

Review: Lenovo ThinkPad P14s Gen 7 Review

Best Without a Discrete GPU: Dell Pro Precision 5 14s Intel

No discrete GPU, no problem, within reason. The Pro Precision 5 14s Intel pairs the Core Ultra X9 388H (50 TOPS NPU) with Intel Arc Pro B390 integrated graphics drawing on a 64GB LPCAMM2 shared memory pool. In our testing that pool ran larger local AI workloads than 8GB discrete cards could hold, completing the full Procyon AI Text Generation suite (scores of 887 on Phi and 786 on Llama 2) where VRAM-limited systems posted DNFs.

Generation speed is modest next to RTX PRO silicon, so treat it as a capable background assistant rather than a speed demon. The rest of the package is remarkable: 3.12 pounds and 23 hours 50 minutes of measured battery, second-longest in our entire dataset. It also holds the Best Ultraportable Workstation spot on our Best Mobile Workstations board.

Review: Dell Pro Precision 5 14s Intel Review

Best Thin-and-Light NPU System: HP EliteBook 6 G1q

The EliteBook 6 G1q answers a different question: how little machine do you need for useful on-device AI? Its Snapdragon X Plus with a 45 TOPS Hexagon NPU ran Llama 3.2 3B at 36.7 tokens per second and Gemma 3 4B at 29.3 in LM Studio, with 1B models topping 62 tokens per second. Those are usable chat speeds from a 3.17-pound machine that measured 19 hours 35 minutes of battery.

The ceiling is low, 4B parameters was the largest model we tested, so this is for local chat, summarization, and offline assistants rather than heavy models. At around $3,100 as configured at review, it was also the least expensive system on this page.

Review: HP EliteBook 6 G1q Review: All-Day Power in a Lightweight Laptop

Also Tested

These systems went through the same lab process and are worth a look for the right buyer, even though they do not hold a category spot today.

• Lenovo ThinkPad P16 Gen 3: the third RTX PRO 5000 machine in the fleet (Phi at 141.2 tok/s); our unit shipped with 32GB of RAM, which held back the rest of the workflow story.
• HP ZBook Fury G1i 18: same 24GB Blackwell GPU class as the Pro Max 18 Plus, a step behind on throughput (Phi at 157.4 tok/s) and $11,687 as tested at review.
• Dell Pro Precision 7 16 Intel: RTX PRO 3000 Blackwell with 12GB of VRAM, the strongest midrange GPU result in this field (Procyon text generation score of 2,244 on Phi via DirectML); its 12GB opens the 14B class that 8GB cards cannot hold.
• Lenovo ThinkPad P1 Gen 8: RTX PRO 2000 8GB in a 4.06-pound chassis with 12 hours 59 minutes of battery; Phi at 77.3 tok/s.
• Dell Pro Max 14 Premium: RTX PRO 2000 8GB at 3.55 pounds; Phi at 54.1 tok/s.
• Dell Pro Max 16 (Ryzen): RTX PRO 1000 8GB with a 16 hour 2 minute battery result; Phi at 61.9 tok/s.
• Lenovo ThinkPad P14s Gen 6: RTX PRO 500 6GB; entry Blackwell AI at 45.7 tok/s on Phi.
• Dell Pro Precision 5 16s Intel: the 16-inch sibling of our no-dGPU pick on the same shared-memory platform, with the longest workstation battery we have measured at 24 hours 43 minutes.
• Dell Pro Precision 5 14s AMD: 60 TOPS NPU and a 64GB shared pool; AMD’s Procyon INT8 image path was not yet available at test time, which limits comparison.
• HP EliteBook X G1a: Ryzen AI 9 HX 375 with a 55 TOPS NPU; we have not yet run our LLM suite on it.

How We Rank

UL Procyon AI Text Generation is our cross-fleet yardstick: the same four models (Phi, Mistral, Llama 3, and Llama 2 where it fits) on every laptop that can hold them, so tokens-per-second numbers here are directly comparable. Where the hardware makes it interesting, we go deeper with LM Studio and Ollama runs on larger models. Each laptop is ranked once per measurement basis, and a system only appears on this page if it produced usable local AI results in our lab.

Prices move quickly in this market, so any dollar figure on this page is the as-tested price at the time the review published. We do not make value claims without checking current vendor configurator pricing, and when we do, we date the check.

Laptop Local AI FAQ

What matters more for local AI on a laptop, VRAM or total memory?

Both, for different reasons. A model must fit in the memory the GPU can reach: on discrete cards that is VRAM (8GB to 24GB in this field), and our Llama 2 13B run DNF’d on 8GB cards because it wanted about 12GB. Unified and shared memory designs flip the equation: AMD’s Ryzen AI Max+ can assign up to 96GB of system RAM to the GPU, and Intel’s shared LPCAMM2 pools reach 64GB, so far larger models load at all, just at lower speeds. Fit determines whether a model runs; the silicon determines how fast.

What is the largest model StorageReview has run on a laptop?

DeepSeek-R1 70B, loaded locally on the HP ZBook Ultra G1a 14 through its 96GB assignable unified memory pool. For sustained generation we measured Gemma 3 27B at 8.96 tokens per second on the same machine. On our desktop side the bar is higher; the Best Desktops for Local AI leaderboard covers systems that serve much larger models.

Do NPU TOPS ratings matter for running LLMs?

Less than the marketing suggests, today. Most local LLM runtimes lean on the GPU, and every tokens-per-second number on this page came from GPU inference except the Snapdragon EliteBook, where the platform’s AI stack is the point. NPUs currently earn their keep on efficiency and on INT8 image generation paths, and they are why several of these machines qualify as Copilot+ PCs. Buy for the GPU and memory first.

What happens to battery life when you run models locally?

Inference is one of the heaviest sustained loads a laptop can run, so plan on wall power for real sessions. The battery numbers on this page are PCMark 10 Modern Office results, a productivity measure, and the spread is enormous: 3 hours 39 minutes on our fastest AI laptop versus nearly 24 hours on the shared-memory Dell. Our Laptop Battery Life Leaderboard ranks the full field.

Should I just buy a desktop for local AI instead?

If the machine will live on a desk, yes, probably. Deskside systems offer more memory per dollar, better sustained thermals, and no battery compromise; our Best Desktops for Local AI leaderboard starts at roughly the price of the midrange laptops here. The laptops on this page are for people whose AI workload has to travel.

The post Best Laptops for Local AI in 2026: Lab-Tested Leaderboard appeared first on StorageReview.com.

Updated August 14, 2026: Initial publication; pricing claims verified against vendor configurators August 14, 2026. Desktop workstations are ranked here on SPECworkstation, SPECviewperf, rendering, and compute results; several of these towers are ranked separately on our Best Desktops for Local AI page by inference throughput, because those are different questions answered by different data.

Every desktop workstation ranked on this page has been through the StorageReview lab. We benchmark the full professional stack: SPECworkstation, SPECviewperf viewsets, Blender and V-Ray rendering, LuxMark, y-cruncher, 7-Zip, and Geekbench, on configurations we disclose. No system is ranked from a spec sheet, and every pick links to the review holding the data.

The field runs wider than any other category we cover: from 1-liter systems that mount behind a monitor to a four-GPU tower that configured past $74,000 as tested. Class and form factor decide more than brand here, so the picks below are organized by what you can physically put on, under, or behind the desk.

At a Glance

Category	System	CPU / GPU (as tested)	GPUs (Tested / Max)	Full Review
Best Overall Desktop Workstation	HP Z8 Fury G6i	Xeon 696X / 2× RTX PRO 6000 Blackwell Max-Q	2 / 4 (up to 384GB VRAM)	Z8 Fury G6i Review
Best for CPU-Heavy Workloads	Dell Precision 7875	Threadripper PRO 9995WX / 2× RTX PRO 6000 Blackwell	2 / 2	Precision 7875 Review
Best Single-GPU Workstation	Dell Pro Max Tower T2	Core Ultra 9 285K / RTX PRO 6000 Blackwell 600W	1 / 2 on select configs	Pro Max Tower T2 Review
Best Mid-Range Tower	Lenovo ThinkStation P3 Tower Gen 2	Core Ultra 9 285 / RTX 5000 Ada	1 / 1	P3 Tower Gen 2 Review
Best Small Form Factor	Lenovo ThinkStation P3 Ultra SFF Gen 2	Core Ultra 9 285 / RTX 4000 SFF Ada	1 / 1	P3 Ultra SFF Gen 2 Review
Best Mini Workstation	HP Z2 Mini G1a	Ryzen AI Max+ PRO 395 / integrated Radeon 8060S	Integrated (fixed)	Z2 Mini G1a Review

The Picks

Best Overall Desktop Workstation: HP Z8 Fury G6i

No other workstation we have tested has this much headroom. The Z8 Fury G6i pairs Intel’s 64-core Xeon 696X with support for up to four RTX PRO 6000 Blackwell Max-Q cards and 384GB of aggregate VRAM, fed by dual power supplies. Our dual-GPU review build posted a SPECworkstation Graphics score of 11.17 and 7,351 samples per minute in Blender GPU rendering, and the chassis takes two more cards without modification. Configurations started around $7,900 at review time and our loaded review unit priced out at $74,878, which tells you exactly who this is for.

Read the full HP Z8 Fury G6i review

Best for CPU-Heavy Workloads: Dell Precision 7875

When the workload is core-bound, the 96-core Threadripper PRO 9995WX is the strongest silicon we have benchmarked in a workstation. In our head-to-head testing the 7875 out-rendered the Xeon-based Z8 Fury by 50 to 66 percent in Blender CPU workloads and led 3DMark CPU by 43 percent, posting 1,039 samples per minute in Blender Monster on the CPU alone. Dual RTX PRO 6000 Blackwell cards give it 192GB of VRAM, and the same machine holds the Best Tower for Local AI spot on our AI leaderboard by inference data. As our review put it, this is not a machine you buy speculatively.

Read the full Dell Precision 7875 review

Best Single-GPU Workstation: Dell Pro Max Tower T2

The most interesting result in this category: one full-power GPU beat two throttled ones. The T2’s single 600W RTX PRO 6000 Blackwell posted 8,025 samples per minute in Blender GPU rendering, ahead of both dual Max-Q Blackwell systems above it in this list. In a 32-liter mid-tower with a Core Ultra 9 285K, it is also the value entry into the current-generation field: Dell’s configurator starts at $1,474 (verified August 14, 2026), while our loaded review unit came to $12,713 at review time. If your renderer does not scale across cards, this is the smarter buy than a multi-GPU flagship.

Read the full Dell Pro Max Tower T2 review

Best Mid-Range Tower: Lenovo ThinkStation P3 Tower Gen 2

At $7,939 as tested, the P3 Tower Gen 2 is the clean answer for professional workloads that do not need HEDT silicon. The Core Ultra 9 285 with an RTX 5000 Ada delivered 30,760 in Cinebench R23 multi-core and 3,884 samples per minute in Blender GPU, and in our testing it beat Lenovo’s own Xeon-based ThinkStation PX in AI inference, GPU rendering, and single-core work. Our review framed it as bridging high-end performance and affordability, and the benchmark spread supports that.

Read the full ThinkStation P3 Tower Gen 2 review

Best Small Form Factor: Lenovo ThinkStation P3 Ultra SFF Gen 2

A 24-core Core Ultra 9 and an RTX 4000 SFF Ada in under four liters. The P3 Ultra SFF Gen 2 posted a Geekbench 6 multi-core of 20,334 and beat its own full-tower sibling in CPU-bound tests like Blender CPU and y-cruncher, at 4,128 dollars as tested. Its GPU ceiling is the honest trade: the 70W RTX 4000 SFF renders at less than half the rate of the Tower’s RTX 5000 Ada. For desk-constrained professional work, it is the best balance we have measured.

Read the full P3 Ultra SFF Gen 2 review

Best Mini Workstation: HP Z2 Mini G1a

The Z2 Mini G1a earned our Editor’s Choice award, and the numbers explain why. AMD’s 16-core Ryzen AI Max+ PRO 395 posted 37,156 in Cinebench R23 multi-core, the strongest desktop-class CPU result in this group, from a 2.8-liter chassis with no discrete GPU at all. Its 128GB unified memory pool also earned it the Best Without a Discrete GPU spot on our Local AI leaderboard, making it the rare system that leads two categories for two different reasons.

Read the full HP Z2 Mini G1a review

Also Tested

These systems have been through the same lab process and are solid choices that did not take a category slot: the Lenovo ThinkStation P8 (its 112,974 Cinebench R23 multi-core remains the highest we have ever recorded in a workstation, on the prior-generation Threadripper PRO 7995WX), the HP Z6 G5 A (our top tower pick of 2023, still formidable but two GPU generations back), the original Dell Precision 7875 (the 2024 dual RTX 6000 Ada build, since superseded by the Blackwell configuration above), the ThinkStation P3 Ultra Gen 1 ($985 base, still a strong compact value), the ThinkStation P3 Tiny Gen 2 (a true 1-liter workstation that hits, per our review, a very attractive balance of size, performance, and serviceability), and the P3 Tiny Gen 1 (from $799, the budget entry into workstation-class tiny PCs), and the Dell Precision 3680 (our 2024 value standout at a $1,029 base; it has since left Dell’s current lineup and now sells mainly through reseller and refurbished channels, so we no longer make the value claim for it).

How We Rank

Three rules govern every StorageReview leaderboard. First, only lab-tested systems are ranked; anything we have not benchmarked can be mentioned, but it cannot hold a category. Second, systems are ranked once per measurement basis: several towers here also appear on our Best Desktops for Local AI page, ranked there by vLLM inference throughput and memory ceiling, ranked here by SPEC, rendering, and compute results. Different question, different data, sometimes a different winner. Third, rankings derive from our standardized suite plus street price, with editorial judgment breaking ties inside scoring bands. A note on pricing: workstation prices move constantly, so as-tested figures are labeled as review-time numbers, and any value claim on this page is checked against the vendor’s configurator on the date in the changelog above. Vendors do not see rankings before publication, and no placement is paid.

Desktop Workstation FAQ

What is the best desktop workstation in 2026?

For maximum capability, the HP Z8 Fury G6i: no other chassis we have tested scales to four RTX PRO 6000 Blackwell GPUs and 384GB of VRAM. For core-bound work, the 96-core Dell Precision 7875 posted the strongest CPU rendering results we have measured. For most professional workloads at rational budgets, the Dell Pro Max Tower T2 and Lenovo P3 Tower Gen 2 cover the single-GPU middle of the market.

Do more GPUs mean faster rendering?

Not automatically, and our data makes the case: the Pro Max Tower T2’s single 600W RTX PRO 6000 rendered 8,025 samples per minute in Blender, beating dual Max-Q configurations of the same silicon in the Z8 Fury (7,351) and Precision 7875 (7,259). Density-optimized Max-Q cards trade clocks for thermals, and not every renderer scales cleanly across cards. Multi-GPU wins on capacity, memory pool, and parallel batch work; a single full-power card often wins a single job.

Threadripper PRO or Xeon?

Our Z8 Fury and Precision 7875 head-to-head is the current answer: Threadripper PRO 9995WX won Blender CPU rendering by 50 to 66 percent and 3DMark CPU by 43 percent, while the Xeon 696X platform won 7-Zip by 38 percent, small-model inference, and PCIe-bound multi-GPU scaling. Pick by workload, not by badge.

How much does a desktop workstation cost?

Treat exact figures as moving targets; these are the patterns from our review history, with as-tested prices reflecting their review dates. The tested field spans a $799-base 1-liter ThinkStation P3 Tiny (2024) to a $74,878 as-tested Z8 Fury G6i (2026). Verified today: the Dell Pro Max Tower T2 configurator opens at $1,474 (August 14, 2026). Broadly, serious single-GPU builds have landed between $4,000 and $13,000 as tested, and multi-GPU flagship configurations run $30,000 and up, with prices trending higher across the industry.

The post Best Desktop Workstations in 2026: Lab-Tested Leaderboard appeared first on StorageReview.com.

Updated August 14, 2026: Dell Pro Precision 7 16 Intel added to Also Tested following its review. Originally published earlier today. In the lab now: additional current-generation mobile workstations as review units land. Battery data on this page cross-references our Laptop Battery Life Leaderboard.

Every mobile workstation ranked on this page has been through the StorageReview lab. We benchmark the full professional stack: SPECworkstation 4.0, SPECviewperf 15 viewsets, Blender rendering, UL Procyon AI, LuxMark, and a PCMark 10 Modern Office battery rundown on every unit. No system is ranked from a spec sheet, and every pick links to the review holding the data.

The defining split in 2026 is that a mobile workstation no longer requires a discrete GPU. Dell's Pro Precision 5 series delivers ISV-certified workstation graphics from integrated silicon while posting the best battery numbers we have ever measured in the class, and RTX PRO 5000 desktop-replacements own the other end of the spectrum with near-desktop rendering throughput at three times the weight. This page ranks both, by what each is actually for.

At a Glance

Category	System	CPU / GPU (as tested)	Weight	Full Review
Best Overall Mobile Workstation	Dell Pro Max 16 Plus	Core Ultra 9 285HX / RTX PRO 5000 24GB	5.63 lb	Pro Max 16 Plus Review
Best Desktop Replacement (18-inch)	Dell Pro Max 18 Plus	Core Ultra 9 285HX / RTX PRO 5000 24GB	7.17 lb	Pro Max 18 Plus Review
Best Ultraportable Workstation	Dell Pro Precision 5 14s Intel	Core Ultra X7 / Arc Pro B390 integrated	3.12 lb	Pro Precision 5 14s Intel Review
Best 14-inch with a Discrete GPU	Lenovo ThinkPad P14s Gen 7	Core Ultra 7 366H / RTX PRO 1000 8GB	3.59 lb	ThinkPad P14s Gen 7 Review
Best Premium Thin-and-Light	Lenovo ThinkPad P1 Gen 8	Core Ultra 7 255H / RTX PRO 2000 8GB	4.06 lb	ThinkPad P1 Gen 8 Review
Best Value	Dell Pro Max 16 (AMD)	Ryzen AI 9 HX 370 / RTX PRO 1000	4.59 lb	Pro Max 16 Review
Best Battery Life	Dell Pro Precision 5 16s Intel	Core Ultra X7 / Arc Pro B390 integrated	4.20 lb	Pro Precision 5 16s Intel Review

The Picks

Best Overall Mobile Workstation: Dell Pro Max 16 Plus

The Pro Max 16 Plus delivers desktop-replacement performance without desktop-replacement weight. Its 175W RTX PRO 5000 pushed Blender GPU rendering to 3,875 samples per minute and posted a SPECworkstation AI and ML score of 2.49, matching its own 18-inch sibling while weighing a pound and a half less and lasting nearly three hours longer on battery. Our review called it a new benchmark for 16-inch mobile workstations, and the numbers back the sentiment. Configurations run from $2,779 to roughly $8,900 as tested.

Read the full Dell Pro Max 16 Plus review

Best Desktop Replacement: Dell Pro Max 18 Plus

If raw throughput is the requirement, this is the fastest mobile workstation we have benchmarked. In an identical-silicon shootout against the HP ZBook Fury G1i 18, the Pro Max 18 Plus won nearly every benchmark we ran: SPECviewperf across all ten viewsets (solidworks-08 at 145.56, a 31 percent lead), Blender GPU at 3,928 samples per minute, plus LuxMark, V-Ray, and Geekbench multi-core. The trade is battery life, at 3 hr 39 min the shortest we have ever measured. Know what you are buying: a desk-to-desk machine at 7.17 pounds, from $3,488 to $9,245 as tested.

Read the full Dell Pro Max 18 Plus review

Best Ultraportable Workstation: Dell Pro Precision 5 14s Intel

The strongest 14-inch productivity system we have reviewed, and it does it without a discrete GPU. The 5 14s Intel was the first laptop in its group to cross 10,000 points in SPECworkstation 4, with a Digital Content Creation score of 14,147 that ran 23 percent ahead of the ThinkPad P14s Gen 7, and it still delivered 23 hr 50 min of battery, fifth on our battery leaderboard. ISV-certified graphics from Intel's Arc Pro B390 integrated silicon make it a real workstation at 3.12 pounds, from $2,228.

Read the full Pro Precision 5 14s Intel review

Best 14-inch with a Discrete GPU: Lenovo ThinkPad P14s Gen 7

When the workload demands CUDA, this is the smallest system that carries it well. The RTX PRO 1000 swept the SPECviewperf table against every integrated-graphics rival, Cinebench 2024 multicore came in 64 percent ahead of the comparable Dell, and the whole package runs 15 hr 52 min on battery at 3.59 pounds, four hours longer than its predecessor.

Read the full ThinkPad P14s Gen 7 review

Best Premium Thin-and-Light: Lenovo ThinkPad P1 Gen 8

The P1 Gen 8 remains the premier choice for workstation power that travels like an ultrabook. A 3.2K Tandem OLED touch display, RTX PRO 2000 graphics that scored 31.52 in the 3ds Max viewset (nearly four times the integrated competition), and 12 hr 59 min of battery in a 4.06-pound chassis. At roughly $4,169 it is priced like the engineering exercise it is, and our review judged the expense justified.

Read the full ThinkPad P1 Gen 8 review

Best Value: Dell Pro Max 16 (AMD)

The value math in this class starts at $1,349, and the tested config holds its own at less than half the flagship price. The Ryzen AI 9 HX 370 with RTX PRO 1000 delivered performance comparable to Intel H-class premium models in our testing, and its 16 hr 2 min battery result is the longest we have measured in any laptop carrying a discrete GPU.

Read the full Dell Pro Max 16 review

Best Battery Life: Dell Pro Precision 5 16s Intel

24 hr 43 min, the longest runtime we have ever recorded in a workstation. The 5 16s Intel also posted the best Cinebench 2026 single-thread score we have measured from any laptop at 535, and led its comparison group in SPECworkstation at 9,994 overall, 12 percent ahead of a far more expensive discrete-GPU stablemate. It holds the workstation battery record on our Laptop Battery Life Leaderboard.

Read the full Pro Precision 5 16s Intel review

Also Tested

These systems have been through the same lab process and are solid choices that did not take a category slot: the Lenovo ThinkPad P16 Gen 3 (RTX PRO 5000 in a 16-inch chassis, consistently in the same performance tier as the winners), ThinkPad P16v Gen 3 (the balanced 4.6-pound middle ground), ThinkPad P16s Gen 4 (its Ultra 7 outruns pricier silicon in 7-Zip), ThinkPad P14s Gen 6 (last year's 14-inch pick, still strong viewport performance), Dell Pro Max 16 Premium and Pro Max 14 Premium (premium OLED builds that trade benchmark ceilings for refinement), Dell Pro Precision 5 14s AMD (24 threads at 3.08 pounds, the R23 multicore leader in its group), the HP ZBook Fury G1i 18 (loses the benchmark race to the Pro Max 18 Plus but wins on battery, serviceability, and four M.2 bays), and the HP ZBook Ultra G1a (exceptional CPU compute from Strix Halo, though our review found its AI performance did not meet the marketing), and the Dell Pro Precision 7 16 Intel (RTX PRO 3000 Blackwell under a 4K Tandem OLED with Thunderbolt 5; big GPU gains over the 5 16s, though CPU-heavy work and battery favor its cheaper sibling).

How We Rank

Three rules govern every StorageReview leaderboard. First, only lab-tested systems are ranked; anything we have not benchmarked can be mentioned, but it cannot hold a category. Second, systems are ranked once per measurement basis: battery endurance is ranked on our Laptop Battery Life Leaderboard and referenced here, and local AI capability will be ranked separately on our upcoming Best Laptops for Local AI page. Third, rankings derive from our standardized suite (SPECworkstation 4.0, SPECviewperf 15, Blender, Procyon AI, PCMark 10) plus street price, with editorial judgment breaking ties inside scoring bands. Vendors do not see rankings before publication, and no placement is paid.

Mobile Workstation FAQ

What is the best mobile workstation in 2026?

For most professionals, the Dell Pro Max 16 Plus: RTX PRO 5000 performance that matches 18-inch desktop replacements in our benchmarks, at 5.63 pounds and with usable battery life. If maximum throughput outranks portability, the Pro Max 18 Plus is the fastest system we have tested; if portability outranks everything, the Pro Precision 5 14s Intel is the strongest 14-inch system we have reviewed.

Do I still need a discrete GPU in a workstation laptop?

Less than you used to. Dell's Pro Precision 5 series carries ISV-certified drivers on integrated graphics and won seven of eleven SPECviewperf viewsets against AMD's best integrated silicon in our testing, while doubling the battery life of comparable discrete systems. The line still matters for CUDA-dependent work and heavy rendering, where an RTX PRO system sweeps the table; the difference is that skipping the dGPU no longer means leaving the workstation category.

Why do the 18-inch desktop replacements rank low on battery?

Because they are portable between desks, not between meetings. The two 18-inch systems we have tested, the Pro Max 18 Plus and ZBook Fury G1i, hold the two shortest runtimes on our battery leaderboard at 3 hr 39 min and 4 hr 48 min while holding the two highest sustained-performance results in this class. That is the trade, and it is the right one for their buyers.

How much does a good mobile workstation cost?

The field we tested spans $1,349 for an entry Pro Max 16 AMD to $11,687 for a maxed ZBook Fury G1i 18. The broad pattern in our data: strong integrated-graphics workstations start around $2,200, discrete RTX PRO 1000-class systems land in the $3,000 to $5,000 range as tested, and RTX PRO 5000 desktop-replacements run $8,000 and up in review configurations.

The post Best Mobile Workstations in 2026: Lab-Tested Leaderboard appeared first on StorageReview.com.

The Dell Pro Precision 5 14s AMD is the Ryzen half of Dell’s newest 14-inch mobile workstation. Our review unit runs the Ryzen AI 9 HX PRO 475, a 12-core, 24-thread processor with a 60 TOPS NPU, alongside Radeon 890M integrated graphics, 64GB of LPDDR5x memory, a 1TB Gen4 SSD, and a 14-inch QHD+ display running at 120Hz. It is the same chassis Dell sells with Intel silicon, but the two configurations diverge more than the shared shell suggests.

The headline for this one is threads. Where most 14-inch workstations in this class ship 16 cores without simultaneous multithreading, the HX PRO 475 brings 24 threads to bear, and it shows up everywhere rendering and compression workloads scale. The panel is the other differentiator: this configuration gets a 2560 x 1600 120Hz screen, whereas the Intel build we tested shipped a 1920 x 1200 panel with no high-refresh option listed.

The Dell Pro Precision 5 14s AMD starts at $2,253, and our review configuration prices out at $5,404 as a single-unit purchase on Dell.com, with the 16GB to 64GB LPDDR5x jump being the largest line item at $1,700. As with all commercial systems, most business buyers purchase through an account team at volume discounts, so the web price is best viewed as a reference ceiling rather than a typical fleet cost. The system is available now on the Dell Pro Precision 5 Series 14S product page.

Dell Pro Precision 5 14s AMD Specifications

Specification	Dell Pro Precision 5 14s AMD (PW514265)
Processor	AMD Ryzen AI 9 HX PRO 475 (12 cores/24 threads, up to 5.2GHz, 36MB cache, 60 TOPS NPU)
Graphics	AMD Radeon 890M (integrated)
Memory	64GB LPDDR5x, 8533 MT/s rated, dual-channel, non-ECC
Storage	1TB SSD, PCIe Gen4
Display	14-inch QHD+/WQXGA (2560 x 1600), non-touch, 120Hz, 500 nits, IPS, 100% sRGB, ComfortView Plus, anti-glare
Camera	8MP HDR RGB + IR with User Presence Detection
Wireless	MediaTek Wi-Fi 7 MT7925, Bluetooth 5.4
Keyboard	English US mini-LED backlit with Copilot key
Security	Fingerprint reader, smart card reader, ControlVault 3+, TPM 2.0, FIPS 140-3 certified, post-quantum cryptography, chassis intrusion detection
Battery	3-cell, 70Wh Long Lifecycle, ExpressCharge and ExpressCharge Boost
Power	100W USB-C adapter
Operating System	Windows 11 Pro (Copilot+ PC)
Chassis	Aluminum alloy
Certifications	ENERGY STAR, EPEAT Gold with Climate+
Warranty	36 months Basic Onsite Service after Remote Diagnosis
Price	$2,253 starting; $5,404 as tested (Dell.com single-unit)

Build and Design

The Dell Pro 5 14 AMD arrives in a dark gray aluminum-alloy chassis with a sleek matte finish and a subtle Dell logo centered on the lid. Measuring just 0.75 inches thick and weighing approximately 3.08 pounds, it offers a portable design without sacrificing performance. Processor options range from the AMD Ryzen AI 5 PRO to the Ryzen AI 9 HX PRO, paired with integrated AMD Radeon 840M through 890M graphics, depending on the configuration. Our review unit features the Ryzen AI 9 HX PRO 475 processor with integrated Radeon 890M graphics.

The unit features a 14-inch non-touch QHD+ IPS display with a 120Hz refresh rate and a rated 35ms response time. Its 500-nit brightness and anti-glare finish produce a bright, easily visible image under a variety of lighting conditions, while 100% sRGB coverage provides vibrant and accurate colors. AMD FreeSync support helps deliver smoother motion, and Dell ComfortView Plus reduces potentially harmful blue-light emissions without significantly affecting color accuracy.

The keyboard deck features a mini-LED backlit keyboard with a dedicated Copilot key. As expected for a compact 14-inch system, there is no numeric keypad, leaving enough room for a comfortable, well-spaced layout. A square power button with an integrated fingerprint reader sits at the upper-right corner, providing convenient biometric authentication.

Below the keyboard is a large, centered glass touchpad that offers plenty of room for navigation and multitouch gestures.

The right side of the Dell Pro 5 includes a 1GbE RJ-45 port, a USB 3.2 Gen 1 Type-A port, and a universal audio jack. It also features a wedge-shaped lock slot for physically securing the laptop, along with optional eSIM support for mobile broadband connectivity.

The left side provides the remainder of the laptop’s connectivity, including two Thunderbolt 4 USB-C ports with Power Delivery and DisplayPort 1.4 support. It also includes a USB 3.2 Gen 1 Type-A port with PowerShare and a full-size HDMI 2.1 output. An optional Smart Card reader is available for organizations that require card-based authentication.

Above the display is an 8MP HDR RGB+IR camera with user-presence detection, enabling high-resolution video conferencing, Windows Hello facial authentication, and presence-aware security features. Dual-array microphones help capture clear audio during calls, while an integrated sliding privacy shutter provides a simple physical way to block the camera when it is not in use.

Removing the bottom panel provides a clear view of the Dell Pro 5’s internal layout. The cooling solution uses a single fan and heat-pipe assembly to manage the processor’s heat, while two 2W speakers are positioned along the lower corners of the chassis. Powering the system is a three-cell, 70Wh lithium-ion battery pack that occupies much of the lower portion of the unit.

Wireless connectivity is handled by a replaceable MediaTek Wi-Fi 7 MT7925 card supporting 2×2 802.11be MIMO and Bluetooth 5.4. Our configuration also includes a replaceable 1TB Kioxia TLC PCIe Gen4 SSD. The system memory is soldered to the motherboard and cannot be upgraded by the user, making the initial configuration an important purchasing decision. Our review unit shipped with 64GB of dual-channel LPDDR5 memory operating at 8,533MT/s.

The bottom cover features a generously sized ventilation section that provides the cooling system with a consistent supply of fresh air. Long rubber feet run along the front and rear edges, slightly elevating the laptop to improve airflow while keeping it stable and preventing it from sliding across a desk.

Dell Pro Precision 5 14s AMD Performance

Our review unit runs the Ryzen AI 9 HX PRO 475 with Radeon 890M graphics, 64GB of LPDDR5x, and a 1TB Gen4 SSD on Windows 11 Pro, with benchmarks tested in the Best Performance power mode. For battery life testing, we configure systems into Balanced power mode and set the screen brightness to 50%. One note on the memory: Dell rates this configuration at 8533 MT/s, but Task Manager, the system BIOS, and our SPEC runs all report 8000 MT/s on the review unit.

For comparables, we included the Intel version of the same machine, the Dell Pro Precision 5 14s Intel (Core Ultra X9 388H, Arc Pro B390, 64GB), the 16-inch Dell Pro Precision 5 16s AMD (Ryzen AI 9 HX PRO 475, Radeon 890M, 64GB), and the previous-generation Dell Pro 5 16 AMD (Ryzen AI 9 HX PRO 470, Radeon 890M, 64GB) as the generational reference. There is no 14-inch AMD system in the prior Pro 5 lineup, so both AMD comparisons are 16-inch machines; the comparisons here are about silicon and platform rather than chassis size.

PCMark 10

PCMark 10 measures general system performance across everyday work such as web browsing, video conferencing, spreadsheets, writing, photo editing, and rendering. The overall score is supported by Essentials, Productivity, and Digital Content Creation subscores that show where a system’s strengths sit. Higher scores are better.

PCMark 10	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
Overall Score	8,762	10,019	8,627	8,268
Essentials	11,442	12,188	10,744	10,870
Productivity	14,272	15,827	14,574	14,322
Digital Content Creation	11,178	14,147	11,127	9,852

The review unit scored 8,762 overall, a 6% gain on the previous-generation Pro 5 16 AMD and effectively a tie with its own 16-inch sibling. The Intel twin is 14% clear of it here, with the gap concentrated in Digital Content Creation, where Arc Pro B390 graphics do most of the work.

PCMark 10 Modern Office Battery

The PCMark 10 Modern Office battery test repeatedly runs common office tasks until the battery reaches the test’s cutoff point, in Balanced mode at 50% display brightness. This is a rundown of the whole system rather than a synthetic idle drain, so it tracks closely with what a full day of productivity work does to the battery. Longer runtimes are better.

Modern Office Battery	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
Runtime (higher is better)	14 hours 26 minutes	23 hours 50 minutes	15 hours 5 minutes	15 hours 22 minutes

This is the review unit’s weakest showing. At 14 hours and 26 minutes, it trails the previous-generation Pro 5 16 AMD by nearly an hour, gives up 39 minutes to its own 16-inch sibling, and concedes more than nine hours to the Intel version of the same laptop. A 120Hz QHD+ panel accounts for part of that against the Intel unit’s lower-resolution screen, but not nine hours of it. Buyers who value all-day unplugged runtime should look hard at the Intel build.

Geekbench 6

Geekbench 6 measures processor performance using a mix of common tasks, with separate scores for single-core and multi-core workloads, plus GPU compute scores through OpenCL and Vulkan. Higher scores are better.

Geekbench 6	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
CPU Single-Core	2,831	3,010	2,854	2,989
CPU Multi-Core	14,611	17,396	14,169	14,348
GPU OpenCL	36,322	56,745	37,520	33,449
GPU Vulkan	49,868	55,581	53,969	48,348

Geekbench 6 is the one CPU test in this review where the Intel twin’s 16 cores beat 24 threads outright, taking multi-core by 19%. The generational gain over the HX PRO 470 is modest at 2% multi-core, and single-core goes to the older chip. On GPU compute, the Radeon 890M is roughly two-thirds of the Arc Pro B390 in OpenCL and close to 90% in Vulkan.

Geekbench 7

Geekbench 7 joins the suite alongside Geekbench 6 as comparison data builds. Its CPU scores are calibrated against a baseline of 2,500, set by the AMD Ryzen 7700, while GPU scores are calibrated against a baseline of 100,000, set by the NVIDIA GeForce RTX 4060. Higher scores are better, and double the score indicates double the performance. Because Geekbench 7 uses new workloads and new baselines, its scores are not comparable to Geekbench 6 results.

Geekbench 7	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
CPU Single-Core	2,571	2,702	2,659	2,602
CPU Multi-Core	15,514	18,788	16,641	14,860
GPU OpenCL	31,423	54,854	31,874	29,645
GPU Vulkan	*	46,906	*	14,451

*Both Precision AMD units failed workload validation in the Geekbench 7 Vulkan run, which assigns a score of zero to the affected subtest. We traced it to the fluid simulation workload and have discarded both results pending clean runs.

The newer suite tells the same story as Geekbench 6, with the Intel twin 21% ahead in multi-core. Worth noting the 16-inch AMD sibling lands 7% above this unit on the same silicon, the clearest thermal-headroom signal in the review.

Cinebench 2026

Cinebench 2026 is the current release in the Cinebench line and the only version we report. It tests CPU and GPU performance using Maxon’s Redshift render engine. It is built on the latest Cinema 4D 2026 code and is designed to show whether a machine is stable under high CPU load, whether a notebook’s cooling can sustain longer render tasks, and how it handles demanding real-world 3D work. Because code and compiler changes accelerated scene rendering, Cinebench 2026 scores use an adjusted range and should not be compared to scores from previous Cinebench versions. Its GPU test supports current AMD and NVIDIA hardware but does not yet run on Intel integrated graphics, and the Dell Pro 5 16 AMD has not completed a GPU run.

Cinebench 2026	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
CPU Single Thread	466	513	471	485
CPU Multiple Threads	4,230	4,513	4,767	4,407
GPU	5,313	N/A	5,667	N/A

The Radeon 890 M’s ability to run the GPU test at all is a practical advantage over the Arc Pro B390 here, since Cinebench 2026 does not yet support Intel integrated graphics. On the CPU side, the 16-inch AMD sibling pulls 13% ahead of the review unit, again pointing to sustained-load headroom rather than silicon differences.

7-Zip Compression

The built-in 7-Zip benchmark measures how quickly the processor can compress and decompress data using multiple threads, run with a 128MB dictionary across ten passes. Decompression tends to scale with thread count, while compression leans on memory latency, so the two halves often tell different stories. Higher GIPS scores are better.

7-Zip 24.09 (GIPS)	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
Compressing	89.099	92.836	85.627	92.443
Decompressing	117.545	93.089	117.863	115.362
Total Rating	103.322	92.963	101.745	103.903

Decompression scales almost perfectly with thread count, and all three 24-thread AMD systems land near 116 GIPS, while the 16-thread Intel unit manages 93. That carries the total rating: 103.322 for the review unit against 92.963 for the Intel twin, an 11% win. Compression, which leans more on memory latency, goes the other way.

y-cruncher

y-cruncher measures how quickly the processor can calculate large numbers of digits of Pi, placing a heavy load on the CPU and memory subsystem. At the same time, the BBP runs to extract hexadecimal digits of Pi. Results are in seconds, so lower times are better. Neither Pro Precision AMD unit could complete the 5-billion or 10-billion digit runs, because the memory these platforms reserve for the integrated GPU leaves less available than those problem sizes require; the previous-generation Pro 5 16 AMD completed 5 billion but not 10 billion.

y-cruncher (seconds, lower is better)	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
Pi 1B	24.648	28.404	22.840	25.160
Pi 2.5B	71.084	83.968	64.242	73.320
Pi 5B	N/A	190.356	N/A	163.768
Pi 10B	N/A	417.224	N/A	N/A
Pi BBP 1B	1.104	1.684	1.100	1.249
Pi BBP 10B	14.333	20.028	12.277	14.141
Pi BBP 100B	166.762	241.409	140.284	161.401

At the sizes it can run, the review unit is quicker than the Intel twin across the board, finishing Pi to 2.5 billion digits 15% faster and the 100 billion BBP extraction 31% faster. The asterisk is that the Intel machine finishes the two largest runs at all, which matters for anyone whose workloads scale past what the shared memory pool allows.

Blender

The Blender benchmark measures rendering performance using three different 3D scenes: Monster, Junkshop, and Classroom. Results are reported in samples per minute, so higher scores are better, and we test on both the CPU and GPU. Scores are not comparable across Blender versions, so we report the current 5.2 release here.

Blender 5.2 (samples/min)	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
GPU
Monster	124.50	567.52	123.00	118.77
Junkshop	101.06	463.30	101.21	95.75
Classroom	82.95	417.38	82.84	78.42
CPU
Monster	127.02	133.46	130.86	127.45
Junkshop	95.84	94.27	99.74	98.89
Classroom	71.63	64.26	74.18	69.52

Blender is the clearest illustration of the tradeoff in this review. The Radeon 890M renders Monster at 124.50 samples per minute against 567.52 for the Arc Pro B390, more than four and a half times faster on the Intel twin. CPU rendering runs the other direction but by far smaller margins, with the review unit taking Junkshop and Classroom from the Intel twin on thread count. Against the previous generation, the GPU gain is 5%, so the 890M in this platform performs essentially as it did last round.

LuxMark

LuxMark measures GPU compute performance by rendering complex scenes through OpenCL, based on LuxCoreRender. We run the Food and Hall scenes on all available OpenCL devices in each system, so the GPU listed in each column header did the rendering. Higher scores are better.

LuxMark v4	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
Hall	2,077	3,494	2,058	2,125
Food	1,041	1,705	1,034	982

The Arc Pro B390 is 68% faster in Hall and 64% faster in Food. Against its own predecessor, the review unit is flat in Hall and 6% up in Food, which matches the Blender picture: this is the same GPU generation, not a step forward.

V-Ray

Chaos V-Ray measures ray-traced rendering throughput, reported in vpaths, where higher is better. We run the CUDA-compatible engine on every system so results remain comparable to notebooks with discrete graphics; on systems without a discrete GPU, that path executes on the integrated graphics, even though V-Ray reports the processor name in its device field.

V-Ray GPU (vpaths)	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
CUDA Engine	897	890	1,039	861

The four systems land within 20% of each other, and the review unit edges the Intel twin by less than 1%. That tight grouping is consistent with what we saw across the Precision family. This compatibility path leans on the CPU and memory subsystem as much as the GPU, so it does not separate these platforms the way LuxMark and Blender do.

3DMark CPU Profile

The 3DMark CPU Profile benchmark measures CPU performance at fixed thread counts, from a single thread up to the maximum available, showing how performance scales as more cores are engaged. Higher scores are better.

3DMark CPU Profile	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
Max Threads	8,778	10,277	9,267	6,696
8 Threads	6,270	6,441	6,573	5,153
4 Threads	4,047	4,176	4,238	3,513
1 Thread	1,160	1,174	1,178	1,176

All four are within 2% on one thread, and the field only separates as threads are added. The generational gain is the story here: 8,778 at max threads is 31% ahead of the previous-generation Pro 5 16 AMD, the largest generational CPU gain in this review. The Intel twin still leads at 10,277.

3DMark Storage and Blackmagic Disk Speed Test

3DMark Storage measures how an SSD performs during gaming-related tasks such as loading games, installing software, saving progress, and moving game files. Blackmagic Disk Speed Test measures an SSD’s sequential read and write speeds using large media files. Higher is better in both.

Storage	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
3DMark Storage	2,424	3,093	2,325	2,477
Blackmagic Write (MB/s)	4,772.0	9,000.8	5,070.5	5,166.5
Blackmagic Read (MB/s)	4,900.4	9,809.2	5,071.2	4,758.0

Our review unit shipped with a Gen4 Kioxia BG7, and it shows: sequential throughput sits near 4.8 GB/s in both directions, whereas the Intel configuration’s Gen5 drive reads at nearly 9.8 GB/s, roughly twice as fast. Dell’s configurator lists this AMD build with Gen4 storage, while the Intel build is Gen5, so this is a platform difference rather than a bad sample, and it is worth attention from anyone moving large media files.

Blackmagic RAW Speed Test

The Blackmagic RAW Speed Test measures how many frames per second a system can decode Blackmagic RAW video on the CPU and on the GPU. We quote the 8K results at 12:1 compression, and higher is better.

Blackmagic RAW Speed Test	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
8K 12:1 CPU (fps)	77	79	74	72
8K 12:1 GPU (fps)	49	85	49	45

CPU decode is close across the group at 72 to 79 fps. The GPU path is where the Radeon 890M struggles, at 49 fps against 85 for the Arc Pro B390, which is the difference between falling short of real-time 8K playback and clearing it comfortably.

Topaz Video AI

The Topaz Video AI benchmark measures AI video upscaling and frame-interpolation performance in frames per second across the application’s enhancement models, run here at 1080p input, where higher is better. The previous-generation Pro 5 16 AMD was not tested with Topaz.

Topaz Video AI (fps)	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
Artemis 1X / 2X / 4X	3.75 / 2.24 / 0.79	6.22 / 5.28 / 1.89	3.72 / 2.21 / 0.78	3.25 / 1.94 / 0.68
Iris 1X / 2X / 4X	4.86 / 2.70 / 0.86	5.18 / 3.14 / 0.94	4.91 / 2.72 / 0.92	4.20 / 2.38 / 0.78
Proteus 1X / 2X / 4X	3.93 / 2.69 / 1.19	6.43 / 6.08 / 2.46	3.99 / 2.70 / 1.18	3.42 / 2.35 / 1.03
Gaia 1X / 2X / 4X	1.90 / 1.35 / 0.94	3.24 / 2.25 / 1.50	1.87 / 1.34 / 0.96	1.66 / 1.15 / 0.83
Nyx 1X / 2X	1.86 / 1.56	1.57 / 1.54	1.87 / 1.53	1.56 / 1.34
Hyperion HDR 1X	11.34	3.22	11.48	9.49
4X Slowmo Apollo / APFast	6.05 / 17.72	8.57 / 22.13	6.08 / 17.59	5.72 / 16.11
16X Slowmo Aion	9.13	DNF	9.05	7.98

The Intel twin wins most of the upscaling models, in some cases by wide margins. Still, the Radeon 890M owns the two places it matters for AMD buyers: Hyperion HDR at 11.34 fps against 3.22, more than triple, and the 16X Slowmo Aion model, which completed here and on the 16-inch sibling but failed on the Intel twin. If your pipeline depends on either of those, the AMD build is the safer machine.

UL Procyon AI Text Generation

The Procyon AI Text Generation Benchmark streamlines LLM performance testing by providing a concise, consistent evaluation method. It enables repeated testing across four local models, Phi, Mistral, Llama3, and Llama2, while minimizing the complexity of large models and the number of variables. Developed with AI hardware leaders, it optimizes the use of local AI accelerators to deliver more reliable, efficient performance assessments. All four systems ran the models through ONNX Runtime with DirectML on their GPUs.

Procyon AI Text Generation	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
Phi	424	887	434	371
Mistral	396	646	403	346
Llama3	352	674	357	306
Llama2	379	786	390	329

The Arc Pro B390 roughly doubles the Radeon 890M in Phi, Llama2, and Llama3, and leads Mistral by 63%, the widest AI gap in the review. Generationally, the review unit gains 14 to 15% on the HX PRO 470, and all four systems complete Llama2 thanks to their 64GB shared memory pools, which is the practical advantage integrated platforms hold over small-VRAM discrete cards for local inference.

UL Procyon AI Computer Vision

The Procyon AI Computer Vision Benchmark measures AI inference performance across CPUs, GPUs, and dedicated accelerators using a range of neural networks, evaluating tasks such as image classification, object detection, segmentation, and super-resolution with models including MobileNet V3, Inception V4, YOLO V3, DeepLab V3, Real ESRGAN, and ResNet 50. The WinML runs use float32 on CPU and GPU, giving a like-for-like view across vendors. We also run the newer Computer Vision 2 suite through each vendor’s native path, Ryzen AI in int8 on the AMD NPUs and OpenVINO on the Intel NPU and iGPU; those results are listed separately since precision and runtime differ by platform. Higher scores are better.

Procyon AI Computer Vision (WinML)	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
CPU	106	141	114	91
GPU	247	404	245	214

 

Procyon AI Computer Vision 2 (native runtimes)	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
NPU	1,176	1,647	1,189	611
iGPU	N/A	1,517	N/A	N/A

The generational NPU gain is the standout number: 1,176 against 611 for the HX PRO 470, nearly double, reflecting the move from a 50 TOPS to a 60 TOPS engine plus a newer Ryzen AI runtime. It is worth noting that the Intel twin still scores higher at 1,647 despite its 50 TOPS rating, a reminder that vendor TOPS figures do not translate directly into benchmark throughput.

UL Procyon AI Image Generation

The Procyon AI Image Generation Benchmark provides a consistent method for measuring AI inference performance from low-power NPUs to high-end GPUs, with three tests: Stable Diffusion XL FP16 for high-end GPUs, Stable Diffusion 1.5 FP16 for moderately powerful GPUs, and Stable Diffusion 1.5 INT8 for low-power devices. The benchmark uses the optimal inference path for each platform, meaning the AMD-optimized DirectML pipeline on the Radeon systems and OpenVINO on the Intel unit. The INT8 test runs on the NPU where a supported quantized model exists; the current Pro Precision AMD units do not yet have one available through Procyon, so those cells are blank pending a supported path.

Procyon AI Image Generation	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
SD 1.5 FP16	281	632	316	255
SDXL FP16	197	731	200	173
SD 1.5 INT8 (iGPU)	N/A	7,873	N/A	N/A
SD 1.5 INT8 (NPU)	N/A	3,003	N/A	3,598

On the FP16 runs, the Arc Pro B390 is more than twice as fast in SD 1.5 and nearly four times as fast in SDXL, which is a real gap for anyone generating images locally. The interesting cell is the NPU row: the previous-generation Pro 5 16 AMD has a genuine quantized NPU running at 3,598 using AMD’s Ryzen AI path, ahead of the Intel twin’s 3,003, and we would expect this unit’s newer 60 TOPS NPU to beat it once Procyon exposes a supported quantized model for the platform.

SPECviewperf 15

SPECviewperf 15 measures graphics performance using viewsets derived from professional applications in CAD, 3D modeling, rendering, engineering, and medical visualization, replayed here at 1080p. Higher scores are better, although performance can vary considerably between applications and graphics architectures. The enscape-01 viewset failed to complete on the review unit across multiple runs, and catia-07 failed on the Intel twin.

SPECviewperf 15 (FHD)	Dell Pro Precision 5 14s AMD (Radeon 890M)	Dell Pro Precision 5 14s Intel (Arc Pro B390)	Dell Pro Precision 5 16s AMD (Radeon 890M)	Dell Pro 5 16 AMD (Radeon 890M)
3dsmax-08	26.30	19.51	26.23	23.21
blender-01	23.01	21.55	23.17	19.89
catia-07	22.70	DNF	22.70	20.16
creo-04	49.27	63.57	49.42	44.27
energy-04	29.51	38.07	29.69	25.39
enscape-01	DNF	14.92	8.45	8.02
maya-07	53.30	83.86	53.48	48.67
medical-04	73.42	69.41	73.31	65.45
snx-05	59.58	78.74	61.33	51.67
solidworks-08	36.66	33.38	36.40	33.12
unreal_engine-01	27.31	41.38	27.38	26.61

This table is more competitive than the raw GPU compute results suggest. The review unit beats the Intel twin in four of the nine viewsets, both completed, taking 3dsmax, blender, medical, and solidworks. At the same time, Intel’s certified driver advantage shows up in the CAD and engineering traces, where it leads by 29 to 57%. Against its own predecessor, the review unit gains in all ten viewsets it completed, by about 12% on average.

SPECworkstation 4

SPECworkstation 4 measures workstation performance across CPU, graphics, storage, AI, product design, engineering, financial services, and other professional workloads, using real applications grouped into seven industry verticals. Higher scores are better, and N/A means the system did not complete every workload required for that category.

SPECworkstation 4	Dell Pro Precision 5 14s AMD	Dell Pro Precision 5 14s Intel	Dell Pro Precision 5 16s AMD	Dell Pro 5 16 AMD
Hardware Subsystems
CPU	1.15	N/A	N/A	1.09
Graphics	2.57	2.70	2.65	2.36
Accelerator	2.22	2.29	2.31	N/A
Storage	0.93	1.70	1.00	0.89
Industry Verticals
AI & Machine Learning	1.38	1.45	1.45	1.37
Energy	1.28	1.67	1.38	1.24
Financial Services	1.07	0.93	1.29	0.98
Life Sciences	1.41	N/A	1.47	1.34
Media & Entertainment	1.39	1.53	1.48	N/A
Product Design	1.34	1.80	1.41	1.34
Productivity & Development	1.03	1.32	N/A	0.78

The review unit beats the Intel twin in Financial Services, where the 16-inch AMD sibling leads the group outright, and trails Intel in the graphics-weighted verticals, while its 0.93 Storage subsystem score tracks with the Gen4 drive.

Conclusion

The Dell Pro Precision 5 14s AMD packs substantial multi-threaded performance into a compact 3.08-pound chassis. Its Ryzen AI 9 HX PRO 475 delivered its strongest results in CPU-intensive workloads, beating the Intel version by 11% in the overall 7-Zip benchmark. It also completed every y-cruncher workload that fit within its available memory. Compared with the previous-generation AMD platform, the 31% improvement in the 3DMark CPU Profile and near doubling of NPU performance represent meaningful progress.

The trade-offs are equally clear. Radeon 890M graphics fall well behind the Intel model’s Arc Pro B390 in most GPU compute, rendering, and AI workloads. At the same time, the Gen4 Kioxia SSD provides roughly half the sequential throughput of the Intel configuration’s Gen5 drive. Battery life reached a respectable 14 hours and 26 minutes, but that remains more than nine hours behind the Intel version. The AMD system does have a few workload-specific advantages, including support for the Cinebench 2026 GPU test and successful completion of Topaz Video AI’s 16X Slowmo Aion model.

Outside of performance, the AMD configuration benefits from an excellent 14-inch QHD+ 120Hz display, a durable aluminium-alloy chassis, strong port selection, Wi-Fi 7, and a comprehensive set of business security features. The soldered memory makes choosing the correct capacity upfront important, especially considering the steep $1,700 upgrade to 64GB in Dell’s online configurator.

Configurations on the Dell Store start at $2,253 and reach $5,404 as tested. The Dell Pro Precision 5 14s AMD carries a high single-unit price, but commercial customers will typically buy through negotiated volume agreements. It is best suited to mobile professionals whose workloads benefit from 24 CPU threads, ample shared memory, and a high-resolution display. Buyers prioritizing GPU acceleration, maximum battery life, or faster storage will find the Intel configuration more compelling, but for CPU-heavy workstation tasks in a highly portable form factor, the AMD version makes a strong case.

For configuration options and current pricing, visit the Dell Pro Precision 5 Series 14S product page.

Leaderboard: The Dell Pro Precision 5 14s AMD ranks #13 on our Laptop Battery Life Leaderboard at 14 hours 26 minutes.

The post Dell Pro Precision 5 14s AMD Review: 24 Threads in a 3.08-Pound Workstation appeared first on StorageReview.com.

Updated August 14, 2026: Dell Pro Precision 7 16 Intel enters at #19 with 11 hr 43 min; the HP ZBook Ultra G1a 14 rotates off at the 20-system cap. New results are added as laptop reviews publish, and each entry links to the full review behind the number.

Every runtime on this page was measured in the StorageReview lab using the PCMark 10 Modern Office battery rundown, which simulates a full day of document work, web browsing, and video conferencing until the battery gives out. Same test, same conditions, every laptop. Nothing here is a manufacturer claim.

Across two years of reviews we have measured everything from under 4 hours to nearly 27, and only the 20 longest-running systems make this page. Class matters more than battery size: the table below shows why a 70Wh business laptop can nearly quadruple the runtime of a 99Wh desktop-replacement workstation.

Battery Life Standouts

Category	System	Runtime
Longest runtime we have recorded	Dell Pro 5 14 Intel	26 hr 48 min
Longest in a mobile workstation	Dell Pro Precision 5 16s Intel	24 hr 43 min
Longest in a convertible	HP EliteBook X Flip G1i	24 hr 34 min
Longest with a discrete GPU	Dell Pro Max 16 (AMD)	16 hr 2 min

The full field below spans business laptops, mobile workstations, convertibles, and rugged tablets, ranked strictly by measured runtime.

The Picks

Best Overall Battery Life: Dell Pro 5 14 Intel

26 hr 48 min on PCMark 10 Modern Office from a 70Wh battery, the longest runtime we have ever recorded. Review: Dell Pro 5 14 Intel Review

Best Battery Life in a Thin Business Laptop: Dell Pro 7 14 Intel

26 hr 18 min, within half an hour of the leader, in Dell’s thinnest Pro chassis. Review: Dell Pro 7 14 Intel Review

Best Mobile Workstation Battery Life: Dell Pro Precision 5 16s Intel

24 hr 43 min, the only 16-inch mobile workstation we have measured past 24 hours. Review: Dell Pro Precision 5 16s Intel Review

Best Convertible Battery Life: HP EliteBook X Flip G1i

24 hr 34 min from a 68Wh battery, with a 360-degree hinge and pen support. Review: HP EliteBook X Flip G1i Review

Best Battery Life with a Discrete GPU: Dell Pro Max 16 (AMD)

16 hr 2 min with an RTX PRO 1000 on board, the longest runtime of any discrete-GPU system we have tested. Review: Dell Pro Max 16 (AMD) Review

Best Small-Battery Runtime: HP EliteBook 6 G1q

19 hr 35 min from just 56Wh, the only sub-60Wh system in our top ten. Review: HP EliteBook 6 G1q Review

The Top 20

Rank	System	Class	Battery	PCMark 10 Runtime
1	Dell Pro 5 14 Intel	Business laptop	70Wh	26 hr 48 min
2	Dell Pro 7 14 Intel	Business laptop	70Wh	26 hr 18 min
3	Dell Pro Precision 5 16s Intel	Mobile workstation	70Wh	24 hr 43 min
4	HP EliteBook X Flip G1i	Business convertible	68Wh	24 hr 34 min
5	Dell Pro Precision 5 14s Intel	Mobile workstation	70Wh	23 hr 50 min
6	HP EliteBook X G1i	Business laptop	68Wh	23 hr 31 min
7	HP EliteBook 6 G1q	Business laptop	56Wh	19 hr 35 min
8	Dell Pro 7 14 AMD	Business laptop	70Wh	19 hr 28 min
9	Dell Pro Max 16 (AMD)	Mobile workstation	96Wh	16 hr 2 min
10	Lenovo ThinkPad P14s Gen 7	Mobile workstation	75Wh	15 hr 52 min
11	Dell Pro 5 16 AMD	Business laptop	70Wh	15 hr 22 min
12	Lenovo ThinkPad X9 14 Aura Edition	Ultraportable	55Wh	15 hr 10 min
13	Dell Pro Precision 5 14s AMD	Mobile workstation	70Wh	14 hr 26 min
14	Dell Pro 14 Premium	Business laptop	Up to 60Wh	13 hr 55 min
15	Dell Pro Rugged 12	Rugged tablet	71.2Wh	13 hr 14 min
16	Lenovo ThinkPad P1 Gen 8	Mobile workstation	90Wh	12 hr 59 min
17	Lenovo ThinkPad P16v Gen 3	Mobile workstation	90Wh	12 hr 0 min
18	Lenovo ThinkPad P14s Gen 6	Mobile workstation	75Wh	11 hr 48 min
19	Dell Pro Precision 7 16 Intel	Mobile workstation	96Wh	11 hr 43 min
20	HP EliteBook X G1a	Business laptop	74.5Wh	10 hr 48 min

How We Test

Every result comes from the PCMark 10 Modern Office battery rundown, run on the configuration we reviewed with our standardized settings. Runtimes are specific to the tested configuration; a different battery option, display panel, or GPU in the same chassis will change the result. Laptops we have reviewed without completing a battery test are not listed, and a handful of units could not finish the benchmark reliably; they are excluded rather than estimated. Only lab-tested systems appear on this page. The leaderboard carries the top 20 runtimes from the trailing two years of reviews; systems below the cut, mostly discrete-GPU desktop-replacement workstations, stay in their reviews, and results retire as hardware ages out of the window.

Battery Life FAQ

Does a bigger battery mean longer battery life?

No, and this leaderboard is the proof. The 70Wh Dell Pro 5 14 Intel ran 26 hours 48 minutes, while the ThinkPad P16 Gen 3 with a 99.9Wh pack, the largest battery allowed on an airplane, reached roughly 7 hours. Silicon efficiency, discrete graphics, and display power dominate the outcome; capacity just sets the ceiling.

What is good battery life for a business laptop in 2026?

Based on our data, the current top tier of business laptops exceeds 24 hours in PCMark 10 Modern Office, comfortably two working days. Anything above 15 hours is a true all-day system. Under 8 hours generally means the machine is a desk-first workstation carrying a discrete GPU and a high-power CPU.

What is the shortest battery life we have ever measured?

The Dell Pro Max 18 Plus, at 3 hours and 39 minutes, with the HP ZBook Fury G1i 18 close behind at 4 hours and 48 minutes. Neither number is a flaw; both are 18-inch desktop-replacement workstations hauling RTX PRO 5000 Blackwell GPUs, 96 to 99Wh batteries, and seven-plus pounds of hardware. Machines in this class are portable between desks, not between meetings, and the battery mostly exists to survive the walk. That is also why they sit below the cut line of this leaderboard rather than on it.

Why do mobile workstations rank lower?

Discrete GPUs, HX-class processors, and high-resolution displays all draw power even at idle. The interesting exception is the certified-graphics-without-a-dGPU approach: the Dell Pro Precision 5 series delivers ISV-certified workstation graphics from integrated silicon and posts nearly 25 hours, holding two of the top five spots on this page.

The post Laptop Battery Life Leaderboard 2026: Top 20 Lab-Tested Laptops, Ranked appeared first on StorageReview.com.

Updated August 13, 2026: Initial publication. On the roadmap as hardware lands: GB300-class systems. Pick order is provisional pending final composite scoring.

Every system ranked on this page has been through the StorageReview lab. We benchmark local AI performance directly: vLLM online serving throughput, time-to-first-token, and time-per-output-token across models including GPT-OSS-120B, Llama 3.1 8B, Mistral Small 3.1 24B, and Qwen3 Coder 30B, plus MAMF compute efficiency and GDSIO storage testing. No system is ranked from a spec sheet.

The defining question for a local AI desktop in 2026 is unified memory versus discrete VRAM. Deskside appliances such as NVIDIA’s GB10-based DGX Spark and AMD’s Ryzen AI Max (Strix Halo) systems put 128GB of unified memory behind a single chip at appliance prices, holding models that would otherwise require multiple discrete GPUs. Workstation towers answer with raw throughput: RTX PRO 6000 Blackwell cards deliver far higher tokens per second, at several times the cost and power draw. This page ranks both, in separate tiers, from one consistent test suite, because the right answer depends on the largest model you intend to run and how fast you need it to respond.

At a Glance

Category	System	Memory	GPUs (Tested / Max)	Full Review
Best Overall Deskside AI System	NVIDIA DGX Spark	128GB unified LPDDR5X	GB10 integrated (fixed)	DGX Spark Review
Best GB10 Implementation	Acer Veriton GN100	128GB unified LPDDR5X	GB10 integrated (fixed)	Veriton GN100 Review
Best x86 Alternative	AMD Ryzen AI Halo (Strix Halo)	Up to 128GB unified LPDDR5X	Radeon 8060S integrated (fixed)	Ryzen AI Halo Review
Best Without a Discrete GPU	HP Z2 Mini G1a	Unified LPDDR5X (ran GPT-OSS 120B)	Integrated, no dGPU (fixed)	Z2 Mini G1a Review
Best Tower for Local AI	Dell Precision 7875	192GB GDDR7 (2× 96GB)	2× RTX PRO 6000 / 2 max	Precision 7875 Review
Best Multi-GPU Platform	HP Z8 Fury G6i	192GB GDDR7 as tested / up to 384GB	2× RTX PRO 6000 tested / 4 max	Z8 Fury G6i Review
The Extreme Pick	Comino Grando RTX PRO 6000	768GB GDDR7 (8× 96GB)	8× RTX PRO 6000 / 8 max	Comino Grando Review

Deskside AI Appliances

The appliance tier, what Dell calls deskside AI and NVIDIA calls the personal AI supercomputer, trades peak throughput for model capacity, power efficiency, and price. With 128GB of unified memory, these systems comfortably hold 70B-class models at high quantization and can stretch to 120B-class, workloads that would demand multiple discrete GPUs in a tower.

Essential lab reading for this tier: our DGX Spark thermal test compares OEM cooling designs across the GB10 systems below and applies to every unit in this class until these platforms see a revision.

Best Overall Deskside AI System: NVIDIA DGX Spark

The DGX Spark is the reference point every other deskside AI box is measured against. The GB10 Grace Blackwell superchip pairs a 20-core Arm CPU with 128GB of unified LPDDR5X, and dual ConnectX-7 200GbE ports make it the only appliance class we’ve tested that clusters out of the box: our two-node distributed inference testing ran pipeline-parallel workloads across Dell, GIGABYTE, and HP nodes over 200GbE. The CUDA software stack remains the deepest in the segment.

Read the full DGX Spark review

Best GB10 Implementation: Acer Veriton GN100

Among the GB10 OEM systems, thermal design is the real differentiator, and the Veriton GN100 stood out in our testing. All GB10 boxes share the same silicon and memory configuration, so sustained performance comes down to cooling. Our multi-OEM thermal comparison is, to our knowledge, the only one of its kind published.

Read the full Veriton GN100 review

Best x86 Alternative: AMD Ryzen AI Halo

If you need Windows or a standard x86 software stack, Strix Halo is the deskside answer. AMD’s Ryzen AI Max+ 395 platform pairs 128GB of unified memory with a dual-OS setup, and in our testing it handled 200B-parameter-class models, a direct shot at the DGX Spark without the Arm/DGX OS commitment.

Read the full Ryzen AI Halo review

Best Without a Discrete GPU: HP Z2 Mini G1a

The Z2 Mini G1a ran GPT-OSS 120B with no discrete GPU at all. HP’s mini workstation puts AMD’s Ryzen AI Max+ PRO silicon in a compact, quiet, IT-friendly chassis, and it remains the clearest demonstration that unified-memory x86 systems have changed what a small office box can do with large models.

Read the full Z2 Mini G1a review

Also Tested: Deskside AI Appliances

These systems have been through the same lab process and are solid choices that did not take a category slot: Dell Pro Max with GB10, ASUS Ascent GX10, GIGABYTE AI TOP ATOM, HP ZGX Nano G1n, and HP EliteDesk 8 Mini G1a.

Workstation Towers for Local AI

When response time matters more than acquisition cost (interactive coding assistants, multi-user serving, agentic pipelines with long tool-call chains), discrete VRAM still rules. These towers are ranked here on inference throughput and memory ceiling, and for each we list the GPU configuration we tested alongside the chassis maximum, since what a chassis can ultimately hold matters as much as what shipped in our build; they are ranked separately on our Best Desktop Workstations page against SPECworkstation and rendering workloads, because they answer two different questions.

Best Tower for Local AI: Dell Precision 7875

Dual RTX PRO 6000 Blackwell GPUs make the Precision 7875 the fastest standard-form-factor system we’ve tested for local inference. With a Threadripper PRO 9995WX and 192GB of combined VRAM across two cards, it holds 100B-class models entirely in GPU memory while delivering interactive-grade time-to-first-token that no unified-memory appliance approaches. Know the ceiling, though: the 7875 chassis supports a maximum of two dual-width cards, so our dual-GPU build is the maxed-out configuration; there is no adding a third later.

Read the full Precision 7875 review

Best Multi-GPU Platform: HP Z8 Fury G6i

The Z8 Fury G6i is the tower you buy when you plan to grow into more GPUs. Our review build ran two RTX PRO 6000 Max-Q cards for 192GB of combined VRAM, but the chassis, fed by dual power supplies totaling up to 2700W, supports up to four Blackwell cards and 384GB of VRAM. That gap between as-tested and maximum is the point: no standard OEM tower we have tested offers more GPU headroom, and the path from two cards to four requires no chassis change.

Read the full Z8 Fury G6i review

The Extreme Pick: Comino Grando RTX PRO 6000

768GB of VRAM in a liquid-cooled 4U chassis: the Grando exists for the buyer whose model does not fit anywhere else. Our review unit shipped with eight RTX PRO 6000 Blackwell cards at 96GB each, the chassis maximum, and Comino’s liquid cooling sustains all eight at full TDP around the clock without throttling. That is more GPU memory than many rack servers, in something that can still live beside a desk. It is loud on price, not on acoustics, and it is deliberately the outlier on this list: proof of where the deskside ceiling actually is.

Read the full Comino Grando review

How We Rank

Three rules govern every StorageReview leaderboard. First, only lab-tested systems are ranked. If we haven’t benchmarked it, it can be mentioned, but it cannot hold a category. Second, systems are ranked once per measurement basis. The towers above also appear on our desktop workstation leaderboard, ranked there by SPECworkstation and rendering performance, ranked here by inference throughput and memory ceiling. Different question, different data, sometimes a different winner. Third, there is a viability bar: a system must run a 30B-class model at interactive speeds, or offer at least 96GB of model-accessible memory, to be ranked on this page.

Rankings are derived from a composite of vLLM online serving throughput, time-to-first-token, time-per-output-token, MAMF compute efficiency, GDSIO storage performance, and street price. Editorial judgment breaks ties within scoring bands. Vendors do not see rankings before publication, and no placement on this page is paid.

Local AI Desktop FAQ

What is the best desktop for agentic AI?

Agentic workloads such as coding agents, tool-calling pipelines, and multi-step autonomous tasks are throughput- and latency-sensitive in a way single-chat use is not, because agents chain many model calls with large context. That favors the tower tier: the Dell Precision 7875’s discrete VRAM delivers the sustained time-to-first-token that keeps long agent chains responsive. For budget-conscious agentic experimentation, a GB10-class appliance runs the same stacks at lower speed. Our full sizing guidance is in RAM, GPU & Storage for Agentic AI (coming soon).

How much memory do I need to run a 70B model locally?

As a working rule, a 70B model at 4-bit quantization needs roughly 40-48GB of model-accessible memory before context; comfortable interactive use with meaningful context wants more. That is why 128GB unified-memory appliances handle 70B-class models well, and why 24-32GB single-GPU systems do not make this page.

Do I need special power to run these systems?

For the appliance tier, no: GB10-class boxes and Strix Halo systems run comfortably on a standard office outlet. The towers deserve a real conversation with whoever owns the building. A fully configured HP Z8 Fury G6i can carry dual power supplies totaling up to 2700W, more than a standard 15-amp, 120V circuit can deliver, and the Comino Grando specifies 2000W hot-swap supplies that require 180-264V input, dropping to 1000W units on 110V service. Before buying from the top of this page, check what the wall can actually feed; a dedicated circuit or 208/240V service may be part of the true cost of a full GPU loadout. Heat follows the same math, since every watt drawn ends up in the room.

What about a Mac Studio?

Not right now, and not only because we have yet to lab-test one. Apple has stopped selling the high-memory Mac Studio configurations, which were the machine’s one real advantage for local AI: enough unified memory to hold very large models. Without those configurations, the current lineup is not a serious contender for this page. Apple is expected to refresh the Mac Studio this year; if high-memory options return, we will test one and reconsider.

The post Best Desktops for Local AI in 2026: Lab-Tested Leaderboard appeared first on StorageReview.com.

Updated August 15, 2026: Ubiquiti Reviews joins the family, indexing all 27 UniFi and Ubiquiti products we have tested, alongside the new Best Enterprise SSDs leaderboard.

StorageReview Best is our family of living leaderboards for storage and complete systems, built on one rule most buyer’s guides skip: if we have not benchmarked it in our lab, we do not rank it. Every pick links to a full review containing the data behind the call, and every leaderboard carries a dated changelog so you can see exactly what changed and when. These pages are updated as new hardware completes testing, not on a publishing calendar.

How These Leaderboards Work

Three rules govern every StorageReview leaderboard. First, only lab-tested products are ranked. Products we have not measured can be mentioned, but they cannot hold a category. Second, products are ranked once per measurement basis. The same workstation tower can lead one leaderboard on SPECworkstation and rendering results and another on AI inference throughput, because those are different questions answered by different data. Third, rankings are built from a composite of our standardized benchmark suites plus street price, with editorial judgment breaking ties inside scoring bands. Vendors do not see rankings before publication, and no placement on any leaderboard is paid.

The Leaderboards

Leaderboard	What It Covers	Updated
Storage Leaderboard	The best SSDs, hard drives, portable storage, and memory cards, ranked from our storage test suite	Aug 13, 2026
Best Enterprise SSDs	Data center SSDs ranked on FIO, GPU Direct Storage, and DLIO checkpointing	Aug 15, 2026
Best Desktops for Local AI	GB10 deskside AI appliances, AMD Strix Halo systems, and RTX PRO 6000 workstation towers, ranked on vLLM serving throughput and memory ceiling	Aug 13, 2026
Best Laptops for Local AI	Unified-memory, discrete-GPU, and NPU laptops ranked on lab-measured LLM performance	Aug 14, 2026
Best Mobile Workstations	Workstation laptops ranked on SPECworkstation, SPECviewperf, rendering, and battery data	Aug 14, 2026
Best Desktop Workstations	Workstation towers ranked on SPECworkstation, SPECviewperf, and rendering workloads	Aug 14, 2026
Best Business Laptops	Business laptops ranked for performance, battery life, and fleet manageability across Intel, AMD, and Snapdragon	Aug 14, 2026
Laptop Battery Life Leaderboard	The top 20 lab-measured PCMark 10 Modern Office runtimes from the past two years of reviews	Aug 14, 2026
RAM, GPU & Storage for Agentic AI	A sizing guide for local and agentic AI hardware, from 8B to 200B-class models	Aug 16, 2026
Ubiquiti Reviews	Every Ubiquiti and UniFi product we have tested, indexed by category	Aug 15, 2026

Current Leaders

A few of the systems currently holding category spots. Each leaderboard page carries the full field, the data, and the reasoning.

Storage

Category	Current Leader	Leaderboard
Best Overall Enterprise SSD	Micron 9550 MAX	Enterprise SSDs
Highest-Capacity SSD Tested	Micron 6600 ION 245.76TB	Enterprise SSDs

Local AI

Category	Current Leader	Leaderboard
Best Overall Deskside AI System	NVIDIA DGX Spark	Desktops for Local AI
Best Tower for Local AI	Dell Precision 7875	Desktops for Local AI
Best Without a Discrete GPU	HP Z2 Mini G1a	Desktops for Local AI
Best Laptop for Local AI	Dell Pro Max 18 Plus	Laptops for Local AI

Workstations

Category	Current Leader	Leaderboard
Best Overall Desktop Workstation	HP Z8 Fury G6i	Desktop Workstations
Best Overall Mobile Workstation	Dell Pro Max 16 Plus	Mobile Workstations
Fastest Mobile Workstation Benchmarked	Dell Pro Max 18 Plus	Mobile Workstations

Business Laptops

Category	Current Leader	Leaderboard
Best Overall Business Laptop	Dell Pro 7 14 Intel	Business Laptops

Battery Life

Category	Current Leader	Leaderboard
Longest Battery Life (26 hr 48 min)	Dell Pro 5 14 Intel	Laptop Battery Life
Longest Battery Life in a Workstation (24 hr 43 min)	Dell Pro Precision 5 16s Intel	Laptop Battery Life

In the Lab

Leaderboards grow as reviews publish. Current testing that will feed upcoming rankings includes the mobile workstation field (ThinkPad P-series, Dell Pro Max, HP ZBook), current business laptops from Dell, HP, and Lenovo, and GB300-class rack systems as hardware lands. If a category you care about is not listed, it is worth checking the workstation review archive; leaderboards start from published reviews.

Leaderboard FAQ

How often are these pages updated?

Whenever a new review changes a ranking, not on a calendar. Every leaderboard carries a dated changelog at the top listing what was added, what moved, and what is in the lab now.

Can vendors submit products for testing?

Yes. Review units go through the same benchmark suite as everything else in the category, and testing a product does not guarantee a leaderboard spot. Vendors can reach the lab through our contact page.

Are any placements paid?

No. Rankings are set by benchmark data and editorial judgment, vendors do not see them before publication, and sponsorships never buy a spot on any leaderboard.

The post StorageReview Best: Lab-Tested Buyer’s Guides appeared first on StorageReview.com.

RAVEL and Strata Expanse have named RT-One as the first commercial customer of the AMPHIX AI Infrastructure Platform, a deployment model designed to validate enterprise AI infrastructure before permanent production sites are completed.

What AMPHIX Is

AMPHIX combines site-ready land, resilient power, cybersecure edge networking, certified compute stacks, and orchestration software. Its production-grade Centers of Excellence, or COEs, allow organizations to test AI infrastructure under expected power, thermal, and performance conditions before making larger capital commitments. The companies say traditional production-grade pilots can take up to nine months to deploy and cost up to $4 million, while rarely testing the configuration that actually reaches production. The launch lands in the same wave of AI infrastructure plays as Volta’s stealth exit and NVIDIA’s buildout financing programs, all aimed at the gap between capital commitment and production revenue.

“The pace of innovation, unprecedented demand for GPUs, and complexity of AI infrastructure stacks have created an environment with more capital at risk, more operational exposure, and longer timelines before a single workload runs in production. The AMPHIX Centers of Excellence were designed to change that,” said Denise Muyco, CEO at RAVEL. “Customers like RT-One can validate, operationalize, and prove their infrastructure against real workloads before making major infrastructure investments. That’s the model the industry needs to move towards, and RT-One is the first to run it.”

What RT-One Will Do at Colusa

RT-One will use the AMPHIX COE near Colusa, California, to validate and optimize its high-performance AI technology stack. The company will test compute, networking, thermal, and orchestration configurations; refine policies governing performance, energy consumption, and cost; train operations personnel; and establish customer environments ahead of campus availability. RAVEL’s Orchestrate AI software will manage and optimize workloads across the environment.

The validated configuration will become part of RT-One’s standardized deployment block for its federated AI infrastructure model across the Americas. The reference design can then be replicated at future sites, with adaptations at layers affected by local regulatory requirements. This approach is intended to reduce repeated validation work, lower rollout risk, and move contracted capacity into service sooner.

“Our customers buy contracted capacity with a firm delivery date and defined performance standards; we pay a price if we miss either,” said Fernando Palamone, CEO of RT-One. “Every validation hour we complete in the COE is an hour we do not spend proving the same point again in Brazil, in Paraguay, or in Colombia. That is what compresses the interval between capital commitment and revenue service: not speed on any single site, but certifying once and deploying many times.”

Who Does What Inside AMPHIX

The Colusa COE is one of the largest planned AMPHIX sites and will be powered by a microgrid developed and managed by Colusa Indian Energy. Within the platform, Strata Expanse supplies the physical foundation, including land, power, cooling, and secure connectivity. RAVEL is responsible for preparing the COE for production use, including technology ecosystem certification, deployment validation, and operator training.

Strata Expanse said the platform is designed to provide a repeatable blueprint for AI infrastructure deployments by separating the work of validating infrastructure configurations from the longer construction cycle for permanent campuses. For RT-One, this enables infrastructure and operating policies to be certified in advance, then carried into subsequent deployments.

The AMPHIX model positions the COE as both a validation environment and an interim production platform, allowing organizations to tune models, architectures, orchestration policies, and governance on the same configuration intended for production. Access is priced on a flexible consumption basis.

The post RT-One Becomes First AMPHIX Customer, Certifying One AI Stack to Deploy Across the Americas appeared first on StorageReview.com.

At FMS 2026, Silicon Motion introduced the MonTitan SSD Reference Design Kit (RDK), a platform built around the company’s next-generation PerformaShape technology. The design targets Agentic AI infrastructure, where enterprise SSDs can provide a persistent memory layer for KV cache offload and autonomous AI agents.

Agentic AI workloads place different demands on storage than conventional AI applications. Agents continuously reason, execute actions, retain context, and interact with external tools. These operations generate varied data types and produce rapidly changing access patterns. As AI data centers support longer multi-step inference processes and larger KV caches, storage must maintain high throughput, consistent latency, predictable quality of service, and sufficient endurance for sustained write-intensive activity.

“AI agents require storage to evolve into a persistent memory layer that retains context and supports continuous autonomous operations,” said Jason Chien, Senior Director of Enterprise Product Marketing at Silicon Motion.

Inside the Updated PerformaShape

The latest PerformaShape architecture adds hardware support for Multi-Dimensional Shaping, enabling more precise workload management. Integrated performance monitoring and support for the NVMe TP4176 API are intended to help enterprise SSDs maintain predictable QoS across complex, dynamic environments, including multi-tenant and multi-agent deployments.

Silicon Motion said the MonTitan RDK is designed to help storage developers build SSDs that preserve context for AI agents while managing data movement across multiple agents. The architecture is intended to reduce resource contention and provide more consistent performance as workloads change.

The SM8366 and SM8466 Controllers

PerformaShape is integrated into Silicon Motion’s SM8366 PCIe 5.0 and SM8466 PCIe 6.0 enterprise SSD controllers. The MonTitan SSD RDK uses these controller platforms to provide SSD manufacturers with a scalable foundation for storage products targeting AI servers and data centers. Silicon Motion said the reference design can also help reduce development time for Agentic AI storage solutions.

The post Silicon Motion MonTitan RDK Targets Agentic AI Storage With Next-Gen PerformaShape and PCIe 6.0 Support appeared first on StorageReview.com.

In late July, Liqid unveiled a scale-up AI platform built around AMD Instinct MI350P PCIe GPUs. The UltraStack 30 pairs Liqid’s composable GPU pooling technology with AMD accelerators to address enterprise and cloud AI inference deployments that need to scale beyond the accelerator count and memory capacity of conventional servers.

UltraStack 30: Pooling 30 AMD MI350P GPUs

The proposed architecture enables up to 30 AMD Instinct MI350P GPUs to be pooled within a single server environment. Liqid states that this can provide 4.3TB of aggregate HBM3E capacity for large models, retrieval-augmented generation deployments, and long-context inference workloads. The platform is intended to allow GPU resources to be allocated dynamically, rather than fixed to individual servers, with native Kubernetes support for parallel model deployments.

The collaboration targets a growing infrastructure challenge in AI inference: maximizing useful accelerator utilization while controlling the cost of generated tokens. Liqid claims its pooled GPU approach can deliver up to 3.7x higher token throughput, 2.1x more tokens per dollar, and 1.8x better tokens per watt compared to traditional server configurations. The company also targets up to a 65% reduction in deployment cost and a 50% reduction in power consumption for large-model deployments, figures it says are based on internal projections that vary by configuration and workload.

“Liqid has established itself as the leader in GPU pooling and scaling. The AMD Instinct MI350P Series gives us the ideal PCIe-based GPU to build the next generation of AI infrastructure, allowing us to deliver solutions tuned for enterprise AI inference where utilization and cost per token decide the economics,” said Rick Hegberg, CEO of Liqid.

Liqid UltraStack 30: AMD MI350P
Specification	Configuration
Server	AMD EPYC 9005 Series CPU, Dual-Socket
GPUs	30× AMD Instinct MI350P PCIe
AI Performance	69 PFLOPS (FP8)
GPU HBM Memory	4.3 TB
Total System Power	~22 kW

Key Metric	Value
Performance per kW	~3.14 PFLOPS/kW
HBM per kW	~195 GB/kW
HBM per PFLOP	~62 GB/PFLOP

Inside the MI350P

AMD positions the MI350P as a PCIe-based accelerator designed to deliver AI inference performance without requiring specialized cooling or a broader data center redesign. AMD’s spec sheet lists the card at 144GB of HBM3E with 4TB/s of memory bandwidth, 2.3 PFLOPS of dense FP8 compute from its CDNA 4 architecture, and a 600W typical board power (450W configurable) in a passively cooled, double-slot PCIe 5.0 form factor. Thirty of those cards is exactly where the UltraStack’s 4.3TB and 69 PFLOPS aggregate figures come from. Combined with Liqid’s fabric, the GPUs can be added and scaled on demand to support changing inference demand, potentially reducing stranded GPU capacity and enabling higher model density per system.

“The AMD Instinct MI350P PCIe GPU delivers exceptional AI inference performance without requiring specialized infrastructure. Together with Liqid’s GPU pooling solutions, customers can scale GPU resources on demand to achieve higher utilization, lower infrastructure costs, and industry-leading AI inference economics,” said Suresh Andani, corporate vice president, Compute and Enterprise AI Group at AMD.

Who the Platform Targets

The solution is aimed at enterprise AI teams, NeoCloud providers, AI service providers, and HPC and research organizations. Identified use cases include private, on-premises enterprise assistants for functions such as sales, HR, marketing, legal, and coding; RAG and long-context inference; scientific workloads including drug discovery and materials science; and mixed-model fleets that serve multiple model sizes from a shared accelerator pool.

From GPU Pooling to CXL Memory Pooling

The CXL side of that story is no longer theoretical. In early August, Liqid launched the EX-5410C Memory Platform, which the company calls the industry’s first and only fully disaggregated, software-defined memory pooling solution. Built on CXL 2.0 and managed through Liqid Matrix software, the EX-5410C pools up to 40TB of DRAM per chassis, scales to a unified pool exceeding 160TB, and dynamically allocates memory across as many as 16 server nodes with what Liqid describes as zero stranded capacity.

Liqid claims up to 30x faster processing for graph analytics workloads and up to 7x more tokens per second for certain KV cache workloads on the platform, and an early testbed is already running at Pacific Northwest National Laboratory, available to DOE-funded researchers through the AMAIS initiative. For the UltraStack, the takeaway is directional: the same composability model Liqid applies to GPU capacity is extending to system memory, which would let future deployments allocate shared memory for KV cache and other memory-intensive AI workloads alongside pooled accelerators.

Further details on joint Liqid and AMD solutions are expected as the companies advance development and customer deployment activities.

The post Liqid Pools 30 AMD MI350P GPUs in One Server: 4.3TB of HBM3E and 69 PFLOPS for AI Inference appeared first on StorageReview.com.

When we reviewed the NVIDIA DGX Spark, storage was the platform’s clearest design flaw, and it is a form-factor problem before it is anything else. The Spark’s internal slots take short M.2 drives, the 2230 and 2242 class, where packaging wins, and capacity loses. The high-capacity end of the client SSD market lives in full-size 2280 drives, where 8TB models ship today, and the Spark simply has nowhere to put one. That leaves a machine built for serious AI work with a storage ceiling better suited to a thin-and-light laptop, and no internal path around it. The Luisuantech GP Spark is a solution for exactly that problem: a 0.58-liter, four-bay box for full-size M.2 drives that cables to the Spark’s 100GbE port, shows up as native NVMe devices with no drivers or formatting, and serves GPU Direct Storage traffic at close to line rate.

The pitch is simplicity with client-drive economics. The GP Spark’s four bays take ordinary M.2 2280 or 22110 NVMe SSDs, the form factors the Spark itself locks out, and presents them over NVMe-oF RDMA through a hardware offload engine on a dedicated chip. Luisuantech’s spec sheet validates drives up to 4TB today, 16TB per enclosure, though these are the same slots where 8TB client drives already ship, so the practical ceiling is a validation question rather than a mechanical one. There is no enterprise array here, no licensing, and no storage OS to learn. Plug a DAC or AOC cable between the GP Spark and the DGX Spark, run modprobe nvme-rdma on the Spark’s Ubuntu base, and the drives appear as /dev/nvme devices ready for GDS access.

Design and Build

The GP Spark is a 150mm x 150mm x 26mm box, a smaller footprint than the Spark itself, wrapped in a perforated metal chassis with a single power button that doubles as a status light: green for normal, red for fault. Power comes over USB-C PD from a 20V/5.4A external adapter, with the whole unit rated under 100W, including drives. The rear panel carries exactly three connectors: the USB-C power input, a USB-C factory debug port, and the QSFP28 100GbE data port, which accepts copper DACs or optical modules. Inside, a dedicated data processor and coprocessor handle the NVMe-oF offload, and the four M.2 bays sit under the top cover.

The rear panel is all business: USB-C power and debug ports on the left, the QSFP28 cage on the right, and a row of copper fin stacks visible through the vents between them. Cooling is entirely passive.

Pop the top cover, and the four M.2 bays sit in a row, here populated with our KIOXIA XG8 test drives. The lid itself is the drive cooler: blue thermal pads on its underside couple each SSD to the finned heatsink that forms the top of the chassis, a clean passive solution for client drives that never see sustained enterprise duty cycles.

A disclosure before the numbers: our unit is a prototype. The bottom label reads GP-Spark-1000, marks the device Prototype, Not for Resale, and carries a February 2026 build date under the Swingsoon brand Luisuantech uses on hardware. Production units may differ in fit and finish, though the platform behavior we tested is what Luisuantech is shipping to reviewers today. Two further notes on the out-of-box experience. Our unit shipped with two printed manuals entirely in Chinese, and initial setup appears to route through Wi-Fi onboarding. Neither is a blocker for the audience this box targets, but a Western launch will need English documentation.

Setup and Architecture

There is no RAID controller and no storage abstraction onboard: the GP Spark is a JBOF in the literal sense, exposing each installed SSD as its own NVMe-oF namespace. In our configuration, four drives appeared as four /dev/nvme devices on the host. Redundancy or striping is the host’s job. The vendor spec sheet lists a single 100GbE port at 10GB/s and 2.7M IOPS; the product report separately references 2x100GbE configurations and up to 24GB/s, a figure Luisuantech confirmed is aggregate read plus write. Our unit and testing used the single-port configuration.

Luisuantech GP Spark Specifications

Specification	Luisuantech GP Spark
Platform Overview
Drive Bays	4 x M.2 NVMe (2280 / 22110) Mixed capacities supported, up to 4TB per drive
Network	QSFP28 100GbE (DAC or optical) RDMA required
Protocols	NVMe-oF RDMA GPU Direct Storage (GDS)
Performance (Vendor-Stated)
Throughput	10GB/s per 100GbE port Up to 24GB/s aggregate read plus write
IOPS	2.7M
Access Latency	Under 20 microseconds
Power and Physical
Power	Under 100W total 20V/5.4A USB-C PD external adapter
Dimensions	150mm x 150mm x 26mm (0.58L)
Operating Temperature	0 to 40C
Compatibility	NVIDIA DGX Spark DGX Station Workstations and servers with RDMA-capable NICs

Performance

Our test configuration paired the GP Spark with a GIGABYTE DGX Spark over a direct 100GbE connection, with four 1TB KIOXIA XG8 client NVMe SSDs populating the bays. It’s important to keep in mind that the drives you pick will play a significant role in the measured performance. We leveraged client Gen5 SSDs; some models, especially enterprise SSDs, may offer higher sustained write performance. We ran FIO sweeps across 4K and 64K random and 1M sequential workloads, read and write, stepping iodepth and numjobs to map the full envelope. Results reflect the final retest after applying Luisuantech’s MTU guidance, which improved transfer behavior over our initial runs.

4K Random Performance

Small-block reads are where the offload engine shows its worth. 4K random reads scaled with queue depth to a peak of 2.43 million IOPS at 9,475 MiB/s, within sight of the vendor’s 2.7M claim and effectively saturating the 100GbE link with 4K transfers. For a passively powered four-bay box feeding a desk-side AI system, that is a remarkable figure.

Writes follow the same shape at roughly half the height, peaking at 1.19 million IOPS. The gap between read and write ceilings is consistent across every workload we ran. The performance is directly related to the underlying drives, so results here will vary depending on configuration.

Read latency bottoms out at 65.3 microseconds on average at low queue depth. That is higher than the vendor’s sub-20-microsecond claim, but results will vary depending on drive selection and network configuration. The network round trip is also doing work in that number; latency stays flat and predictable until the link saturates.

Write latency is the one place the spec sheet claim lands: 20.4 microseconds average at minimal depth, right at the vendor’s under-20-microsecond figure and low enough that the fabric is effectively invisible to the application.

64K Random Performance

At 64K, the story becomes purely about bandwidth. Random reads hold 9,503 MiB/s at peak, statistically identical to the 4K and 1M ceilings. Whatever block size the workload brings, the GP Spark delivers the same answer: the full line rate of its 100GbE port.

64K random writes plateau at 4,742 MiB/s, the same ceiling we measured at every other block size.

1M Sequential Performance

Large-block sequential reads, the profile of model loading and dataset streaming, reach 9,496 MiB/s and hold there from modest queue depths onward. This is the workload the GP Spark exists for, and it runs at the wire.

Sequential writes hold 4,742 MiB/s, roughly half of read throughput, and that ceiling is identical at every block size we tested. We flagged the asymmetry to Luisuantech during testing and worked through a round of tuning with the company, including MTU changes; the figures here represent the best the platform delivers in its current single-port configuration, and Luisuantech confirmed they are consistent with its specifications for the write path. For the read-dominated workloads this box targets, model loading, dataset streaming, and RAG retrieval, it is a footnote; for heavy ingest, size expectations accordingly.

Conclusion

The GP Spark does one thing and does it cleanly: it gives one or more DGX Spark the storage the platform really needs for heavy lifting. Cable it up, load the kernel module, and nearly 10GB/s of GDS-accessible flash appears without a driver install, a storage OS, or an enterprise invoice. Filling it with client M.2 drives is the point; capacity gets relatively cheap when the box accepts whatever 2280 or 22110 SSDs you have, and the offload engine handles the protocol work the drives never see.

Our take is that this is a neat, well-executed add-on rather than a breakthrough. Reads stop at the single link’s line rate; writes stop at roughly half of that. Pricing is the open question: the GP Spark is not yet listed at retail in the US or China, our test unit is a prototype, and Luisuantech has not published pricing. The value argument rests on the box coming in meaningfully below enterprise NVMe-oF alternatives, which its client-drive design should allow. For Spark owners who hit the internal storage wall, and our original review suggests that many of them will, this is an easy path to solve that issue without carving out storage from a large enterprise storage estate.

Product page: Luisuantech GP Spark

The post Luisuantech GP Spark Review: Nearly 10GB/s of Plug-and-Play Storage for the DGX Spark appeared first on StorageReview.com.