For two decades, the SSD-versus-HDD conversation ended the same way: flash wins on performance, disk wins on price per terabyte, and the size of that price gap settled the argument in favor of bulk storage. As storage technology has matured and AI has taken over, that framing is clearly out of date. The largest data center operators are no longer capacity-constrained by what they can afford to buy; they are constrained by what they can power, cool, and physically fit. When we reviewed the 245TB capacity Micron 6600 ION SSD earlier this year, the drive’s benchmark results told one story, but the more consequential one was the math around watts and rack units. In our measurements, a single Micron 6600 ION SSD did the work of eight nearline hard drives, and the flash configuration writing at full tilt drew less power than the HDD configuration at rest. Extend that swap across an exabyte, and the footprint collapses from 22 racks of best-case HDD density to 6 racks of flash, handing 16 rack positions back to the compute plan. A drive that puts nearly a quarter petabyte into a single slot changes how exabytes are planned, and in today’s facilities, both the watts and the racks are spoken for years in advance.

Micron has promoted the 245TB capacity 6600 ION aggressively since it began shipping in May, and we have the test data to check the claims. The FIO and power work here, alongside the GPU Direct Storage and DLIO checkpointing results in our full review of the drive, supports a specific version of the pitch: for read-heavy bulk storage at scale, the drive delivers density and efficiency that nearline HDDs cannot approach, with trade-offs that are predictable when the workload matches the design.

Key Takeaways

• One drive replaced eight: A single 245TB Micron 6600 ION stood in for eight Seagate Exos M 30TB drives in RAID5 in the same Dell R5715, with the HDD backplane and RAID controller pulled from the chassis.
• Flash writing draws less than disk idling: 170.2W under sequential writes against 173.5W for the HDD configuration at rest. Every measured state freed between 49.7W and 63.8W, blending to 55W per unit.
• Watts convert directly to compute: 44.0kW freed per rack of flash servers. At 2,182 drives, less than three racks, the swap frees the 120kW that operates a complete GB200 NVL72.
• Efficiency, not just draw: 72.7 MB/s per watt against 10.4 for the HDD array, or 3.8 watt-hours to read a terabyte against 26.7, roughly 7x on sequential reads.
• An exabyte in 6 racks instead of 22: 4.2 times the capacity per rack with both sides at best-case density, returning 16 rack positions and roughly 320 square feet of white space per exabyte.

Data Center Economics Are Being Rewritten

The International Energy Agency projects data center electricity consumption will more than double to roughly 945 TWh by 2030, with AI the largest driver, and in the facilities absorbing that growth, power, space, and cooling have replaced budget as the binding constraints. In a power-capped building, every watt and every rack position allocated to storage is capacity that cannot be used for revenue-generating GPUs.

The 245TB capacity Micron 6600 ION SSD was built for that arithmetic. Nearline HDDs still win on acquisition cost per terabyte, and for cold archives and rarely accessed data, that advantage remains decisive. But operators planning at exabyte scale are pricing something different: the cost to own and operate that capacity over its service life within a fixed power and space envelope. Measured that way, the drive’s density converts directly into recovered watts, rack positions, and cooling headroom, which is to say, into the compute the facility can now hold. Travis Vigil, Dell Technologies’ senior vice president of ISG product management, made the vendor version of the TCO case at launch, calling the 6600 ION “a meaningful reduction in total cost of ownership for customers building out AI and large-scale data center environments.” IDC’s Jeff Janukowicz, research vice president for solid state drives and enabling technologies, described the same shift at launch: “Rapid AI dataset growth is shifting storage economics from individual drives to rack-level efficiency. Operators need more usable capacity per rack while staying within strict power and cooling constraints.” The following sections provide our own measurements to support that argument.

Micron 6600 ION 245TB Overview

The 6600 ION 245TB is currently the highest-capacity commercially available SSD, shipping since May 5, 2026, in E3.L 9.5mm and U.2 15mm form factors. It is built on Micron’s ninth-generation G9 QLC NAND with a six-plane architecture that pushes NAND I/O to 3.6 GB/s, the fastest QLC currently shipping in a data center SSD. The controller rides a PCIe Gen5 x4 interface, and the drive carries the compliance list a procurement team expects: OCP 2.6, NVMe 2.0d, TAA eligibility, and FIPS 140-3 Level 2 certifiability with CNSA 2.0 and SPDM 1.2 support.

The spec sheet makes the design intent straightforward. Sequential reads are rated at 13,700 MB/s against 3,000 MB/s writes, and random reads at 1.78 million IOPS against 42,000 random write IOPS. The top-capacity model uses a 16K indirection unit rather than 4K, which is why endurance is 1.0 SDWPD for 128KB sequential writes but 0.3 RDWPD for 16K random writes. This is a read-optimized high-capacity SSD.

Specification	Micron 6600 ION 245TB
Platform Overview
Capacity	245TB
Form Factors	E3.L (9.5mm) U.2 (15mm)
Interface	PCIe Gen5 x4, NVMe 2.0d
NAND	Micron G9 QLC, six-plane, 3.6 GB/s NAND I/O
Performance
Sequential Read	13,700 MB/s
Sequential Write	3,000 MB/s
Random Read	1,780,000 IOPS
Random Write (4K/16K)	42,000 IOPS
Latency (QD1, Read/Write)	100µs / 20µs
Power and Endurance
Max Power	≤30W
Idle Power	≤5W
Endurance	1.0 SDWPD (128KB sequential) 0.3 RDWPD (16K random)
MTTF / UBER	2.5 million hours @ 50°C <1 sector per 1017 bits read
Features
Compliance	OCP 2.6 NVMe 2.0d NVMe-MI 1.2d TAA
Security	FIPS 140-3 L2 certifiable CNSA 2.0 SPDM 1.2 Micron SEE, SED options

What Our Testing Showed

We tested the performance and power of the Micron 6600 ION 245TB in a Dell R5715, alongside a configuration of eight 30TB HDDs in RAID5. The platform was specifically chosen as it included 3.5″ HDD support, as well as a PERC12 RAID to put the drives on their best foot forward in terms of performance.

Because the R5715’s platform overhead is the same in both configurations, the per-workload delta between the Micron 6600 ION and eight HDDs in RAID5 represents the storage-attributable power difference. We used our Quarch Mains Power Analysis Module to measure the server’s power draw. Idle power measurements include all server components at rest; the gap reflects the difference in drive power, standby electronics, and fan response between the two configurations. Active workload deltas follow the same logic: the server’s non-storage draw changes only marginally between workloads, so the active savings is the storage subsystem, as well as some CPU uplift in areas where the SSD is driving higher I/O through the system.

Configurations Tested

• HDD RAID5: Eight Seagate Exos M 30TB drives installed via the R5715’s standard HDD backplane, connected to the onboard RAID controller and configured as a RAID5 group presenting 210TB of usable capacity. The backplane, controller, and all eight drives were fully populated and operational during measurement.
• NVMe SSD: The HDD backplane was removed from the chassis entirely. A dedicated E3.L riser was installed in its place, and the Micron 6600 ION 245TB was seated into that riser and connected via PCIe Gen5 x4 NVMe, presenting the full 245TB as a single namespace. No spinning drives, RAID controller activity, or backplane electronics were present during SSD measurement.

On paper, the 245TB capacity Micron 6600 ION SSD is rated at an idle power draw of less than 5W and a maximum power draw of under 30W. The Seagate Exos M 30TB HDDs, by comparison, include an idle power rating of 6.9W and a max operating rating of 9.5W.

FIO workloads

• Sequential 128K: Read and write tested independently. We focused on single-threaded sequential transfers for both SSDs and HDDs, reflecting the scan and ingest patterns that dominate high-capacity data lake tiers.

Each workload ran for 3 minutes, with power and performance data averaged over that duration.

Workload Power Consumption Comparison

At rest, the gap is immediate and structural:

• Idle draw: 115.9W for the 6600 ION configuration against 173.5W for eight Exos M HDDs on their backplane with the RAID controller.
• The gap: 57.6 watts. The HDD configuration draws 50% more power at rest, or, put another way, the flash configuration cuts idle power draw by a third.
• Normalized to capacity: 0.47 W/TB at idle against the HDD array’s 0.72 W/TB on raw capacity parity, a 35% reduction in power per terabyte.

That idle figure matters more than it first appears. Storage isn’t always saturated, particularly in large-scale object storage, archival, and AI data lake environments where data is ingested in bursts and read intermittently.

Under load, the separation holds across both directions of traffic:

• Sequential read: 175.0W for the 6600 ION configuration against 224.7W for the HDD array.
• Sequential write: 170.2W against 234.0W.
• No overlap: the flash configuration under its heaviest measured load drew 49.7W less than the HDD array under its lightest.
• The headline number: sequential write draw of 170.2W sits below the HDD array sitting completely idle at 173.5W.

A quarter petabyte of flash writing at full speed uses less system power than eight idle hard drives.

Power is draw at an instant; energy is draw multiplied by time. That distinction matters here, because the flash configuration does not merely draw less, it finishes sooner. Our review of the 245TB Micron 6600 ION measured 12,729.8 MB/s in 128K sequential reads, while our review of the Seagate Exos M 30TB measured 292MB/s from a single drive. Granting the HDD array perfect linear scaling across all eight spindles, a best case no RAID5 group reaches in practice, the disk side is bounded near 2,336 MB/s. Set those against the power we measured during the sequential read runs and the efficiency gap opens well past the wattage gap:

• Throughput per watt: 72.7 MB/s per watt for the 6600 ION configuration against 10.4 MB/s per watt for the HDD array.
• Energy to read one terabyte: roughly 3.8 watt-hours on flash against roughly 26.7 watt-hours on disk.
• Either way, about 7x, and that is a floor, since it credits the hard drives with scaling they do not deliver.

The throughput figures come from our respective product reviews rather than from the power runs above, so the ratio is derived rather than measured end to end, and it describes sequential reads, the pattern this drive was built for.

Micron reports a much larger advantage from its own testing, citing up to 84x better energy efficiency for AI preprocessing. Micron isolates drive-level power, which strips out the roughly 116W of shared platform draw sitting under both of our configurations, compares against sixteen hard drives rather than eight, and uses AI pipeline access patterns that punish spinning media harder than sequential reads do. Our 7x is the system-level floor; Micron’s ceiling can be much higher depending on the workflow applied.

The Watts-and-Racks Math

The relevant framing for storage power consumption is not the electricity bill; rather, data center power budgets are finite, grid access is increasingly constrained, and cooling capacity is a hard ceiling on what can be deployed. Every watt a storage system consumes is a watt that cannot go to a GPU. In environments where Blackwell deployments are throttled by available power rather than by procurement limits or software readiness, storage efficiency is a compute-capacity decision.

Our system-level measurements provide the basis. Each comparison unit, one 6600 ION (245TB) replacing eight Seagate Exos M drives in a RAID5 group (210TB usable), shows a measurable power delta across all tested states, ranging from 49.7W (sequential read, 1T) to 63.8W (sequential write, 1T). To get a single representative number, we weighted the three measured states by a duty cycle typical of an AI data lake tier: 40% idle, 45% sequential reads, and 15% sequential writes dominated by ingest and checkpoint traffic, consistent with published characterizations of training-tier storage from Meta and others. That blend lands at 55W freed per unit. The weighting barely matters, which is the point: because every measured state frees roughly 50W to 64W, any realistic mix of idle, read, and write time lands within a few watts of the same answer.

Applying that 55W average to real deployments: a 2U flash storage server holds 40 E3.L drives, and a standard 42U rack holds 20 such servers with 2U left for networking. Mapping the freed power against the current GPU lineup: the H200 SXM carries a 700W TDP, the B200 SXM (Blackwell) 1,000W, the B300 SXM (Blackwell Ultra) 1,400W, and the GB200 NVL72, which packs 72 Blackwell GPUs and 36 Grace CPUs into a single liquid-cooled rack, draws approximately 120kW at full load per NVIDIA’s system documentation.

The threshold tipping point highlights how compelling the power savings can be at scale. Replacing 17,456 spinning drives with 2,182 Micron 6600 ION SSDs, less than three racks of dense flash servers, frees 120kW at the 55W average: enough to operate one complete GB200 NVL72 within the same facility power allocation. That system delivers 1.44 exaFLOPS of FP4 compute across 72 Blackwell GPUs with 13.5TB of unified HBM3e memory. The storage freed that headroom by consuming less power while doing the same job.

The 55W average spans the full range of measured states. In the best case (sequential write 1T, 63.8W), the NVL72 threshold is 1,881 SSDs, about 2.4 racks. In the worst case (sequential read 1T, 49.7W), it rises to 2,414 SSDs, about 3.0 racks. The 55W blend lands the threshold at 2,182 SSDs (2.73 racks), and no realistic workload mix moves it far.

Every figure above is an IT-level delta, which keeps the comparison honest: GPUs have their own cooling overhead, so converting freed watts into GPUs at the IT level is an apples-to-apples comparison. At the facility level, the savings only grow. With a PUE of 1.2 to 1.5, the 44.0kW of IT load removed per rack represents 53 to 66kW of total facility load, since cooling overhead tracks the IT watts it no longer has to reject. The HDD drive counts also use raw capacity parity; serving the same usable capacity through RAID5 would require roughly 17% more hard drives, raising the HDD-side power in every row.

Jeremy Werner, Micron’s senior vice president and general manager for the core data center business unit, framed the trajectory at launch: “AI workloads are driving massive growth in shared data, continuing the shift of data center storage share from HDDs toward SSDs.”

One Exabyte in 6 Racks: The Floorspace Dividend

Power is only half of the constraint set. The other half is physical: how much capacity fits in a rack, and how many racks an exabyte occupies. The fair way to frame this is best case against best case. For disk, that is Seagate’s Exos 4U106, the densest HDD enclosure shipping, filled with 44TB drives, the largest announced. 10 enclosures fill a 42U rack with 1,060 drives and 46.6PB of raw capacity. For flash, the best case is the E3.L form factor the 6600 ION was built around: 2U servers holding 40 drives each, 20 per rack, 800 drives, and 196.6PB. Rack against rack, each side at its densest, hard drives deliver 46.6PB while the 245TB SSDs deliver 196.6PB, 4.2 times the capacity in the same footprint.

If anything, the comparison is as favorable as possible to disk. A fully loaded 4U106 is a deep, top-loading enclosure that weighs over 200 pounds populated; many facilities cap HDD racks well below 10 enclosures on floor-loading grounds alone. We also gave the hard drives a capacity point that is still ramping while the 245TB SSD is shipping today.

Stretch the math to a deployed exabyte, and the difference becomes a floor-plan decision. On 44TB disk at best-case density, one exabyte requires roughly 22,730 drives across 22 racks. On 245TB flash, it lands in about 4,070 drives across 6 racks, roughly 73% less floor space. That is 16 rack positions handed back before counting the switching, cabling, and cooling distribution that those racks drag along.

Data center planners typically model net white space at roughly 20 square feet per rack once aisles, power distribution, and cooling clearances are allocated, a planning proxy drawn from the Department of Energy’s best-practices guidance for data center design. By that measure, the 16 rack positions returned per exabyte represent about 320 square feet of white space that never has to be built, leased, or provisioned, and at 10EB the arithmetic clears 3,200 square feet. For facility planners, that is the most valuable kind of capacity: deferral. An expansion that slips a year or is canceled outright returns its budget and its construction timeline to the compute plan.

The savings extend past the concrete. Every rack that is never deployed removes a top-of-rack switch, its power distribution units, and the structured cabling that ties it into the fabric, along with the management overhead it carries for its service life. Network architects see fewer ports to light and fewer devices to patch, operations teams see fewer field units to monitor and replace, and finance sees rack-linked line items, from colocation space to cooling distribution, that simply never appear. Micron’s own space- and power-economics analysis for the 6600 ION reaches the same conclusion from modeled rack math: capacity that grows within existing racks preserves both space and power headroom, while HDD-based growth adds racks, servers, and supporting infrastructure in lockstep.

The reason those 16 racks matter is that rack positions have become the scarcest commodity in the industry. CBRE’s Global Data Center Trends report for 2026 puts Northern Virginia vacancy at 0.3%, Atlanta at 1.0%, and 80% of all capacity under construction in the top four US markets already preleased before it opens. Absorption hit a record 2,236MW globally in the past year, while rents climbed double digits in several major markets. An operator cannot simply buy more floorspace; it doesn’t exist to buy, and new capacity is spoken for years before the concrete is poured. Space recovered within an existing footprint is the only inventory available immediately and at no marginal cost.

Floorspace and power converge on the same conclusion from different directions. A rack that no longer holds nearline disk is not merely emptier; it is available. In a build where every rack position is provisioned for power and cooling before the first server arrives, handing 16 positions back to the compute plan is worth more than the real estate itself. The watts follow the racks, and in current AI facilities, both are spoken for years in advance.

Final Thoughts

The 245TB Micron 6600 ION is the highest-capacity SSD shipping today, engineered around a single job: put a quarter petabyte in one slot and hold it there at under 30 watts. This is a read-optimized design, so sequential writes cap at 3,000 MB/s and the 16K indirection unit sets random write endurance accordingly. That is a specific lane, something SSDs have gotten better at identifying over the years. When performance is the goal, Micron has a part for that: the Gen6 9650 is rated at 14,000 MB/s sequential writes and up to 900,000 random write IOPS, more than twenty times the ION on random writes. For the read-heavy bulk capacity AI data lakes are built on, that combination changes the arithmetic of the facilities it goes into.

Our measurements bear that out at a scale we did not expect going in. One of these drives displaced eight nearline hard drives in the same server, and the flash configuration writing at full speed drew less system power than the disk configuration sitting idle. Extended across a rack, the swap frees 44.0kW. Extended to 2,182 drives, it frees the 120kW that operates a GB200 NVL72. The floorspace math moves in the exact same way: an exabyte lands in 6 racks instead of 22, returning 16 rack positions and roughly 320 square feet of white space per exabyte to the compute plan.

None of this makes hard drives obsolete. Acquisition cost per terabyte still favors disk, and for cold archives and rarely accessed data, that advantage remains decisive. What has changed is the set of workloads where that comparison is the right one. In a facility where the power budget and the floor plan are both spoken for years in advance, storage stops being a line item measured in dollars per terabyte and becomes infrastructure competing directly with GPUs for the same finite resources. Measured that way, a drive that gives back watts and rack units is not a storage purchase; it is a compute purchase made in the storage tier.

Micron 6600 ION Product Page

This report is sponsored by Micron. All views and opinions expressed in this report are based on our unbiased view of the product(s) under consideration.

The post Micron 6600 ION 245TB: Swap the Hard Drives, Power an NVL72 for Free appeared first on StorageReview.com.

The KIOXIA CM9-R is the first enterprise SSD family built on eighth-generation BiCS FLASH, and the 15.36TB E3.S model in our lab is the volume sweet spot of the range: 3,400K random read IOPS and 14,800 MB/s of sequential read from a drive with a 25W typical active rating. Gen5 enterprise drives routinely push into the high 20s to hit their headline figures, and KIOXIA’s pitch for the CM9 generation is that BiCS8’s CBA (CMOS directly Bonded to Array) design delivers top-of-class throughput without the power tax.

The CM9-R is the 1 DWPD read-intensive arm of the family; a 3 DWPD CM9-V mixed-use line runs alongside it. Capacities span 1.92TB to 30.72TB in E3.S, with a 61.44TB flagship reserved for the 2.5-inch U.2 version. Below 7.68TB, the series carries BiCS generation 5 flash; the 7.68TB, 15.36TB, and 30.72TB models are where generation 8 lives, and our 15.36TB review unit (KCM9XRJE15T3) sits at the exact capacity point where the series posts its best write specs: 11,000 MB/s sequential write and 540K random write IOPS, both series highs shared with no other capacity.

The drive is PCIe 5.0 x4 with dual-port x2 support for high-availability topologies, NVMe 2.0 and NVMe-MI 1.2c compliant, and supports the OCP Datacenter NVMe SSD v2.5 specification (not all requirements). Power-loss protection and end-to-end data protection are standard, with SIE, SED, and FIPS 140-3 SED security variants available in the model matrix.

KIOXIA CM9-R Specifications

Specification	KIOXIA CM9-R 15.36TB (E3.S)
Platform Overview
Model	KCM9XRJE15T3 (SIE) KCM9DRJE15T3 (SED) KCM9FRJE15T3 (FIPS SED)
Capacity	15,360 GB
Form Factor	E3.S, 7.5 mm
Interface	PCIe 5.0 (single x4, dual x2), NVMe 2.0, NVMe-MI 1.2c
NAND	KIOXIA BiCS FLASH generation 8 3D TLC (CBA)
Performance (single port x4, up to)
Sequential Read (128 KiB)	14,800 MB/s
Sequential Write (128 KiB)	11,000 MB/s
Random Read (4 KiB)	3,400K IOPS
Random Write (4 KiB)	540K IOPS
Read / Write Latency (4 KiB QD1, typ.)	65 µs / 10 µs
Power and Endurance
Power (Active / Ready)	25 W typ. / 5 W typ.
Endurance	1 DWPD
MTTF / Warranty	2,500,000 hours / 5 years
Features
Protection	Power Loss Protection End-to-End Data Protection Dual-port for HA
Compliance	OCP Datacenter NVMe SSD v2.5 (partial)

 

Testing Background and Comparables

We use a Dell PowerEdge R760 running Ubuntu 22.04.2 LTS as our test platform for all workloads in this review. Equipped with a Serial Cables Gen5 JBOF, it offers wide compatibility with U.2, E1.S, E3.S, and M.2 SSDs. Our system configuration is outlined below:

• 2 x Intel Xeon Gold 6430 (32-Core, 2.1GHz)
• 16 x 64GB DDR5-4400
• 480GB Dell BOSS SSD
• Serial Cables Gen5 JBOF
• NVIDIA L4

Drives Compared

• KIOXIA CD9P-R 7.68TB (1 DWPD)
• Sandisk DC SN861 7.68TB (1 DWPD)
• Solidigm PS1010 7.68TB (1 DWPD)
• Solidigm PS1030 12.8TB (3 DWPD)
• Micron 7600 MAX 6.4TB (3 DWPD)
• Micron 9550 MAX 12.8TB (3 DWPD)
• Micron 9550 Pro 7.68TB (1 DWPD)

The pairing with KIOXIA’s own CD9P-R matters most: the CD9P-R showed what BiCS8 does at data center positioning, and the CM9-R is the enterprise flagship version of the same flash.

FIO Performance

128K Sequential Write (IODepth 16 / NumJobs 1)

The CM9-R opened its FIO run with a strong steady-state 128K sequential write result: 8,668.1 MB/s at 230.4 µs, third in the group behind only the Micron 9550 MAX (10,957.9 MB/s) and 9550 Pro (10,354.6 MB/s), and comfortably clear of the rest of the field, which clustered between 6,370 and 7,127 MB/s. Against its own CD9P-R sibling at 6,912.4 MB/s, the enterprise drive had a 25% advantage, a meaningful gap given that both use the same BiCS8 flash.

128K Sequential Read (IODepth 64 / NumJobs 1)

The read side is the one result in our dataset that’s far from the drive’s rating. At IODepth 64 on a single worker, the CM9-R delivered 9,974.6 MB/s at 801.7 µs, last in the group and well short of both the 14,800 MB/s spec and the 14,235.9 MB/s its CD9P-R sibling posted in the identical configuration. This is a single-job test, and the sweeps below show the CM9-R reading at full pace when work is spread across jobs. However, buyers with single-stream large-block read patterns should note this behavior and test their own pipeline.

64K Random Write

The 64K random write sweep placed the CM9-R second in the group, peaking at 9,635.3 MB/s (IODepth 8 / NumJobs 2, at just 103.4 µs), behind only the Micron 9550 MAX at 10,878.1 MB/s. The low-queue-depth entry was the best in the field outright: 3,334.8 MB/s at 18.4 µs at IODepth 1 / NumJobs 1, ahead of the Solidigm PS1030 (2,901.7 MB/s at 21.1 µs) and everything else. The latency ceiling remained controlled, topping out at 1,942.5 µs across the sweep, while several competitors exceeded 2,300 µs and the PS1010 spiked to nearly 6,000 µs.

64K Random Read

At 64K reads, the CM9-R posted the best single-stream opening in the group, 1,359.0 MB/s at 45.6 µs at IODepth 1 / NumJobs 1, edging out even the CD9P-R (1,334.0 MB/s), with the rest of the field starting at roughly half that. Just as striking is how little concurrency it needed: the drive reached its 13,402.4 MB/s peak at IODepth 4 / NumJobs 8 with only 148.7 µs of latency, while most of the field needed IODepth 32 configurations to reach theirs. The deep-queue crown went to the Solidigm PS1030 at 14,162.9 MB/s, but nothing in the group serves 64K reads at low concurrency as well as the two KIOXIA drives.

16K Sequential Write

One programming note: our 16K coverage is now sequential rather than random, a workload we first broke out in our Micron 9550 MAX review and are carrying across the group, since mid-size sequential streams dominate real deployments, particularly AI pipelines that move ordered data in runs rather than scattering random hits.

The change suits the CM9-R. Its 10,812.2 MB/s peak was second only to the Micron 9550 MAX at 10,970.8 MB/s, but where the 9550 MAX needed IODepth 32 / NumJobs 4 and 181.9 µs to reach its number, the CM9-R hit its peak at IODepth 1 / NumJobs 16 with 22.8 µs of latency, the same low-queue-depth signature that defines this drive everywhere else. Single-worker latency of 10.4 µs was again the best in the group, and the drive cleared 10,400 MB/s at three separate points spanning IODepth 1 to 32.

16K Sequential Read

On reads, the CM9-R posted 13,393.2 MB/s at IODepth 32 / NumJobs 8, second to its CD9P-R sibling at 13,819.7 MB/s, and it delivered effectively the same figure, 13,390.1 MB/s, at IODepth 16 / NumJobs 8 with half the latency, 149.0 versus 298.3 µs. One honest wrinkle: unlike the random tests, single-stream 16K sequential reads favored the Sandisk SN861 (12.5 µs) and the Micron (13.9 to 21.3 µs) over the KIOXIA pair, which were near 32 µs. Once parallelism was introduced, the CM9-R scaled well, achieving 12,400 MB/s at IODepth 4 / NumJobs 16.

4K Random Write

The CM9-R reached 1,502.9K IOPS at IODepth 8 / NumJobs 16 with just 84.6 µs of latency, fourth in the group behind the Micron 7600 MAX (1,781.2K), Solidigm PS1030 (1,595.8K), and 9550 MAX (1,544.2K), and nearly 2.8x its 540K rated figure, which KIOXIA specs at a fixed queue depth. At queue depth 1, it led the entire field with 115.8K IOPS and 8.2 µs, the fastest small-block write response we measured among drives in this group.

4K Random Read

The deep-queue 4K read sweep is the CM9-R’s weakest chart: 1,973.4K IOPS at its IODepth 16 / NumJobs 16 peak, seventh of eight, ahead of only the Micron 7600 MAX, and well below the Sandisk SN861’s group-leading 2,555.6K. Note the rating context: KIOXIA’s 3,400K IOPS spec is measured at queue depths beyond our sweep’s IODepth 32 / NumJobs 16 ceiling, so we treat the gap as a difference in test configuration rather than a shortfall, but drives like the SN861 and PS1030 found substantially more within the same sweep. The other half of the chart belongs to the CM9-R: 33.6K IOPS at 29.3 µs at queue depth 1, the lowest 4K read latency in the group, fractionally ahead of the CD9P-R, which made the same trait its signature in June.

GDSIO Performance

GPU Direct Storage reads are where the CM9-R’s thread scaling shows the same character as its CD9P-R sibling demonstrated in June. The two KIOXIA drives led the 16K block-size segment throughout, and the CM9-R’s peak 16K result of 136.4K IOPS was the best in the entire group, ahead of the CD9P-R’s 134.2K and far clear of the Solidigm pair at the bottom (the PS1030 managed 101.2K). In the 1M segment, the CM9-R peaked at 6.06 GiB/s, second only to the CD9P-R’s 6.16 GiB/s, with the Solidigm PS1010 and Micron 9550 MAX at 6.05 GiB/s behind it. Single-thread 16K read latency of 44.5 µs sat mid-group, effectively identical to the CD9P-R; the Sandisk SN861 held the single-thread edge at 26.1 µs.

The write sweep is where the CM9-R quietly sets itself apart. Its 5.51 GiB/s peak (1M block, 16 threads) was second to the Micron 9550 MAX’s 5.69 GiB/s on paper, but the shape of the two lines is the real story: the 9550 MAX dipped to roughly 2.2 GiB/s at 1M/64 in its documented volatility window, while the CM9-R held the most stable line in the group across the entire 1M segment. It also avoided the 128K high-thread collapse exhibited by both Solidigm drives. Single-thread 16K write latency of 21.4 µs tied the CD9P-R for the group’s best. For sustained GPU-adjacent write streams, this was the most predictable drive on the bench.

DLIO Checkpoint Performance

Our DLIO checkpoint test writes recurring training checkpoints across three passes, and it is a test of consistency as much as speed; lower times are better, and the gap between pass one and pass three shows how a drive behaves once its FTL is working under sustained pressure. One note on this data: checkpoint counts scale with drive capacity, so larger drives, such as the 15.36TB CM9-R, log more checkpoints per run, and per-checkpoint results are not aligned point-for-point across drives. That is why the table below presents pass averages, which normalize each drive’s run into a directly comparable figure; lower is better. On the first pass, the entire comparison group falls within a tight 459- 465-second range, with the CM9-R at 462.8 seconds. The separation comes later. By pass three, the CM9-R settles at 580.9 seconds, mid-pack toward the rear of the group: quicker than the Solidigm PS1030’s 599.2 seconds, but behind the Sandisk SN861 (553.3), Micron 9550 MAX (555.3), and its own CD9P-R sibling (570.6).

Drive	Pass 1 Average (seconds)	Pass 2 Average (seconds)	Pass 3 Average (seconds)
Sandisk DC SN861 7.68TB	461.3	558.6	553.3
Micron 9550 MAX 12.8TB	462.8	558.9	555.3
Micron 9550 Pro 7.68TB	461.4	577.9	559.7
Solidigm PS1010 7.68TB	458.8	561.1	564.6
Micron 7600 MAX 6.4TB	464.2	581.5	567.3
KIOXIA CD9P-R 7.68TB	464.7	575.6	570.6
KIOXIA CM9-R 15.36TB	462.8	571.9	580.9
Solidigm PS1030 12.8TB	462.3	578.0	599.2

 

The pass-three position deserves interpretation rather than alarm. A 1 DWPD read-intensive drive is not bought for sustained checkpoint writes, and the FIO results carry the drive’s actual argument: the group’s best low-queue-depth latency on both reads and writes, delivered consistently. Notably, the CM9-R degraded gradually and predictably across its passes rather than swinging, showing the same steadiness as its GDS write line, which is what checkpoint scheduling cares about once a drive is in the rotation. Buyers who need the checkpoint burst should look at the Micron 9550 family or wait for the CM9-V mixed-use sibling.

Conclusion

The KIOXIA CM9-R 15.36TB is the first enterprise-class SSD with BiCS8, tested at the capacity point where the series delivers its best write performance, and our data gives it clear framing: the lowest low-queue-depth latency in the comparison group across every workload we ran. At queue depth 1, it led the field outright in 4K reads (33.6K IOPS at 29.3 µs), 4K writes (115.8K IOPS at 8.2 µs), 16K sequential writes (10.4 µs single-worker), and 64K reads (1,359.0 MB/s at 45.6 µs), and it reached its 64K read peak at a fraction of the concurrency its rivals required. In GPU Direct Storage, it produced the group’s best small-block read IOPS at 136.4K and the most stable sustained write line on the bench, peaking at 5.51 GiB/s, without the high-thread volatility seen in several competitors.

The deep-queue 4K read peak of 1,973.4K IOPS sat seventh of eight in our sweep (KIOXIA’s 3,400K rating is measured at queue depths beyond our test ceiling, but rivals found more inside the same sweep), and the single-worker 128K sequential read of 9,974.6 MB/s was the one result off from its rating.

For latency-sensitive read tiers, OLTP frontends, virtualized estates on dual-port HA infrastructure, and read-heavy AI serving where response time at real-world queue depths matters more than synthetic deep-queue peaks, the CM9-R is the strongest argument yet that BiCS8’s CBA design changes the class. Sustained heavy-write pipelines should look to the CM9-V or the mixed-use drives in this group. Everyone else gets the drive that answers fastest when it matters, in a capacity stack that runs to 30.72TB in E3.S and 61.44TB in 2.5-inch.
The CM9-R’s queue-depth-1 sweep also earned it a leaderboard spot: it now holds Best for Latency-Critical Workloads on our Best Enterprise SSDs page.

KIOXIA CM9-R E3.S Product Page

The post KIOXIA CM9-R 15.36TB Review: BiCS8 Flash Hits Full Speed at Low Queue Depths appeared first on StorageReview.com.

The Solidigm D7-PS1030 is the mid-endurance arm of the company’s first PCIe 5.0 data center family, pairing the platform it shares with the D7-PS1010 with a 3 DWPD rating and a random write ceiling of up to 800K IOPS, double the 400K of its standard-endurance sibling. The family ranges from 1.6TB to 12.8TB across E3.S and U.2, using 176-layer TLC NAND, with rated performance up to 14,500 MB/s sequential read and 10,000 MB/s sequential write. Solidigm targets write-centric and mixed workloads: OLTP, metadata logging, HPC, and AI/ML pipelines where write pressure never lets up. Our review unit is the 12.8TB E3.S model.

This is not our first time with this drive; however, it is the first time a single drive has been benchmarked. Eight of these same 12.8TB units carried the flash tier in our KV cache offload work on the Dell PowerEdge XE7740, where the array sustained 1.9GB/s of continuous KV writes around the clock, a duty cycle that works out to roughly 3.2 drive writes per day per drive under mirroring, or about 1.6 DWPD striped as RAID0. That is exactly the bracket the PS1030’s 3 DWPD rating is built to straddle: KV offload is a workload where endurance, not capacity or peak speed, is the defining constraint of the tier. The read-intensive class that dominates Gen5 headlines would burn through its 1 DWPD budget quickly with that duty.

At 12.8TB, Solidigm rates the drive at 2.75 million 4K random read IOPS and a full 800K random write IOPS, with the family’s 3.1M random read peak living lower in the capacity stack. Active power is 23W typical and 5W idle, in line with the Gen5 field, with five configurable power states from 5W to 25W for operators working against a fixed rack budget. Solidigm also makes two additional claims worth noting: up to 70% better energy efficiency than comparable drives, and up to 90% IOPS consistency across the life of the drive. Endurance can also be spent faster over shorter horizons; the same media supports 4.98 DWPD over a three-year period, and the 12.8TB model is rated for 70PB written either way. Reliability specs are in line with the class: 2.5-million-hour MTBF, a five-year warranty, and a UBER Solidigm test to 1E-18. Security options include TCG Opal 2.02 SED trim, FIPS 140-3 Level 2-certifiable hardware, OCP-standard secure boot and firmware signing, plus device attestation and key revocation.

Solidigm D7-PS1030 Specifications

Specification	Solidigm D7-PS1030 (12.8TB E3.S, family range noted)
Platform Overview
Capacities	1.6TB 3.2TB 6.4TB 12.8TB (as tested)
Form Factors	E3.S 7.5mm (as tested) U.2 15mm
Interface / Protocol	PCIe 5.0 x4, NVMe
NAND	Solidigm 176-layer TLC 3D NAND
Performance (Up To, Vendor Rated)
Sequential Read (128K)	14,500 MB/s (family)
Sequential Write (128K)	10,000 MB/s (family)
Random Read (4K)	2,750K IOPS (12.8TB) Up to 3,100K IOPS (family peak)
Random Write (4K)	800K IOPS
Power and Endurance
Power (Active / Idle)	23W typ. / 5W typ. Five configurable power states, 5W to 25W
Endurance	3.0 DWPD (5-year basis) 4.98 DWPD (3-year basis) 70 PBW at 12.8TB
Reliability and Security
MTBF / UBER	2,500,000 hours Tested to 1E-18
Security	TCG Opal 2.02 (SED variant) FIPS 140-3 Level 2 certifiable OCP-standard Secure Boot and firmware signing Device attestation, key revocation
Warranty	5 Years

 

Solidigm D7-PS1030 Performance

The context for the charts that follow is that the PS1030 lands in our comparison field as the mixed-use counterweight to the read-intensive drives that have driven recent Gen5 coverage. Against the KIOXIA CD9P-R we reviewed in June, the PS1030 gives up rated sequential read (14,500 vs 14,800 MB/s) but nearly doubles rated random write (800K vs 450K IOPS) and triples the write budget (3 vs 1 DWPD). Its closest philosophical rival in the group is the Micron 7600 MAX, the other 3 DWPD drive in the field, with the Micron 9550 MAX bringing the performance-tier mixed-use fight at the same 12.8TB capacity as our unit.

Drive Testing Platform

We use a Dell PowerEdge R760 running Ubuntu 22.04.2 LTS as our test platform for all workloads in this review. Equipped with a Serial Cables Gen5 JBOF, it offers wide compatibility with U.2, E1.S, E3.S, and M.2 SSDs. Our system configuration is outlined below:

• 2 x Intel Xeon Gold 6430 (32-Core, 2.1GHz)
• 16 x 64GB DDR5-4400
• 480GB Dell BOSS SSD
• Serial Cables Gen5 JBOF
• NVIDIA L4

Drives Compared

• KIOXIA CD9P-R 7.68TB (1 DWPD)
• KIOXIA CM9-R 15.36TB (1 DWPD)
• SanDisk DC SN861 7.68TB (1 DWPD)
• Solidigm PS1010 7.68TB (1 DWPD)
• Micron 7600 MAX 6.4TB (3 DWPD)
• Micron 9550 MAX 12.8TB (3 DWPD)
• Micron 9550 Pro 7.68TB (1 DWPD)

DLIO Checkpointing Benchmark

To evaluate SSD real-world performance in AI training environments, we utilized the Data and Learning Input/Output (DLIO) benchmark tool. Developed by Argonne National Laboratory, DLIO is specifically designed to test I/O patterns in deep learning workloads. It provides insights into how storage systems handle challenges such as checkpointing, data ingestion, and model training.

The table below shows each drive’s average checkpoint completion time across three passes; lower is better. One note on this data: the number of checkpoints in a run scales with each drive’s capacity, so larger drives log more checkpoints, and per-checkpoint results are not aligned point for point across drives of different sizes. That is why we publish pass averages, which normalize each drive’s run into a directly comparable figure. When training machine learning models, checkpoints are essential for periodically saving the model’s state, preventing loss of progress during interruptions or power failures. This storage demand requires robust performance, especially under sustained or intensive workloads. We used the DLIO benchmark version 2.0 from the August 13, 2024, release.

To ensure our benchmarking reflected real-world scenarios, we based our testing on the LLAMA 3.1 405B model architecture. We implemented checkpointing using torch.save() to capture model parameters, optimizer states, and layer states. Our setup simulated an eight-GPU system, implementing a hybrid parallelism strategy with 4-way tensor parallelism and 2-way pipeline parallel processing distributed across the eight GPUs. This configuration yielded a checkpoint size of 1,636GB, reflecting the requirements of training modern large language models.

Drive	Pass 1 Average (seconds)	Pass 2 Average (seconds)	Pass 3 Average (seconds)
SanDisk DC SN861 7.68TB	461.3	558.6	553.3
Micron 9550 MAX 12.8TB	462.8	558.9	555.3
Micron 9550 Pro 7.68TB	461.4	577.9	559.7
Solidigm PS1010 7.68TB	458.8	561.1	564.6
Micron 7600 MAX 6.4TB	464.2	581.5	567.3
KIOXIA CD9P-R 7.68TB	464.7	575.6	570.6
KIOXIA CM9-R 15.36TB	462.8	571.9	580.9
Solidigm PS1030 12.8TB	462.3	578.0	599.2

 

Looking at the pass averages, the Solidigm PS1030 opened right in the pack at 462.3 seconds in Pass 1, where the entire comparison group landed within a seven-second band of roughly 459 to 465 seconds. The separation came later. In Pass 2 the PS1030 stepped up to 578.0 seconds, in the upper portion of a field that ranged from the SanDisk SN861’s 558.6 to the Micron 7600 MAX’s 581.5 seconds, and by Pass 3 it had drifted to 599.2 seconds, the highest average in the comparison group, with the rest of the field settling between the SN861’s 553.3 and the KIOXIA CM9-R’s 580.9 seconds.

The PS1030’s own checkpoint log shows the shape of that drift. The PS1030 opened at 465.3 seconds and held near 461 seconds through checkpoint 5 before stepping up. Once it transitioned, it ranged from roughly 553 to 593 seconds, closed checkpoint 12 at 614.7 seconds, and its latest checkpoints reached as high as 629.0 seconds. The result is consistent with the drive’s one established weakness rather than a new one: DLIO checkpointing is exactly the kind of large sequential write burst that the 128K single-worker FIO test flagged. The gap is real but bounded, about 5% against the KIOXIA CD9P-R on Pass 3 averages, and the PS1030 scaled predictably across passes rather than swinging.

FIO Performance Benchmark

To measure the storage performance of each SSD across common industry metrics, we leverage FIO. Each drive undergoes the same testing process, which includes a preconditioning step of two full drive fills with a sequential write workload, followed by steady-state performance measurement. As each workload type being measured changes, we run another preconditioning fill of that new transfer size.

In this section, we focus on the following FIO benchmarks:

• 128K Sequential
• 64K Random
• 16K Sequential
• 4K Random

128K Sequential Write (IODepth 16 / NumJobs 1)

The steady-state 128K sequential write test is the one area where the PS1030 shows some weakness compared to the other SSDs. The drive posted 6,370.3 MB/s at 313.7 µs, the lowest bandwidth and highest latency in the group, trailing even the read-intensive KIOXIA CD9P-R at 6,912.4 MB/s. The Micron 9550 MAX led at 10,957.9 MB/s, with the 9550 Pro at 10,354.6 MB/s, and the KIOXIA CM9-R took a clear third at 8,668.1 MB/s. The PS1030’s sibling, the PS1010 (7,126.5 MB/s), and the SanDisk DC SN861 (7,116.5 MB/s) occupied the middle, with the Micron 7600 MAX at 6,960.6 MB/s.

Worth keeping in frame: this is a single-job workload, and the PS1030’s write architecture is built to spread work, not to win one stream. The random write sections below show the same drive moving far more data once parallelism is introduced, which matches the access pattern that its target workloads generate.

128K Sequential Read (IODepth 64 / NumJobs 1)

The read side flipped the script. The PS1030 delivered 14,156.4 MB/s at 564.8 µs, effectively tied with its PS1010 sibling (14,163.3 MB/s) and within 0.6% of the group-leading CD9P-R (14,235.9 MB/s). The Micron 9550 Pro (14,050.1 MB/s) and 9550 MAX (14,047.5 MB/s) completed the pack of five drives, saturating the Gen5 interface within a 200 MB/s band. The SN861 followed at 12,631.2 MB/s, the 7600 MAX at 11,240.5 MB/s, and the CM9-R, so strong in the write test, came in last here at 9,974.6 MB/s in this single-job configuration. For a drive sold on its write budget, giving up nothing on big-block reads is the quiet win in this chart.

64K Random Write

The 64K random write sweep is where the PS1030’s character shows. The drive peaked at 7,224.0 MB/s, fourth in the group behind the Micron 9550 MAX (10,878.1 MB/s), KIOXIA CM9-R (9,635.3 MB/s), and Micron 9550 Pro (9,069.4 MB/s), and ahead of the Micron 7600 MAX (6,960.5 MB/s) and the rest of the field. What sets the PS1030 apart is where that peak occurred: at IODepth 2 / NumJobs 2, with just 34.3 µs of latency, while most of the field needed deep queues to reach their best numbers. The drive saturates almost immediately and then stays flat for the rest of the sweep, which is exactly the profile you want for a steady-state write tier running at moderate concurrency around the clock. The gap to its sibling is also the endurance tier earning its keep on performance terms: the PS1010 peaked at 5,873.9 MB/s, 23% below the PS1030.

On latency, the PS1030 opened at 21.1 µs at IODepth 1 / NumJobs 1, second only to the CM9-R’s 18.4 µs, and its worst point in the entire sweep was 2,831 µs, less than half the PS1010’s 5,987 µs spike. The 9550 MAX remained the most controlled at high concurrency, topping out at 1,714 µs.

64K Random Read

The PS1030 took the group’s best 64K random read peak at 14,162.9 MB/s (IODepth 32 / NumJobs 8), a hair ahead of the Micron 9550 Pro (14,049.9 MB/s), 9550 MAX (14,049.7 MB/s), and PS1010 (14,013.6 MB/s), with the CM9-R at 13,402.4 MB/s and the CD9P-R at 12,036.0 MB/s further back. The low-queue-depth story belongs to the KIOXIA drives, as it did in our CD9P-R review: the CM9-R opened at 1,359.0 MB/s and the CD9P-R at 1,334.0 MB/s at IODepth 1 / NumJobs 1, roughly 45 µs latency territory, while the PS1030 started at 768.4 MB/s and 81.0 µs, mid-pack. The PS1030 ranks at the top of this chart for scaling, not for single-stream response.

16K Sequential Write

A note on the test itself: beginning with this comparison group, our 16K coverage is sequential rather than random, a workload we first broke out in our Micron 9550 MAX review. Mid-size sequential streams better represent what these drives do in production, especially in AI pipelines, which stream ordered training data in, stage intermediate results out, and feed inference tiers in runs rather than scattering random hits across the drive.

The results reshuffle the field compared to the random sweeps. The Micron 9550 MAX led at 10,970.8 MB/s (IODepth 32 / NumJobs 4), with the KIOXIA CM9-R nearly matching it at 10,812.2 MB/s and the 9550 Pro third at 9,772.8 MB/s. The PS1030 peaked at 5,977.7 MB/s (IODepth 8 / NumJobs 4), seventh in the group, ahead of the SanDisk DC SN861 at 5,772.5 MB/s and just under its PS1010 sibling at 6,271.7 MB/s. The shape of the sweep, though, is the same signature the random tests showed: the PS1030 was already at 5,856.7 MB/s by IODepth 2 / NumJobs 4 with 21.1 µs of latency, then held a flat band between roughly 4,700 and 6,000 MB/s across the rest of the matrix. Its single-worker latency of 10.9 µs sat in the leading cluster with the CM9-R (10.4 µs) and CD9P-R (11.0 µs), well under the Microns at 15.1 to 17.8 µs. The ceiling is modest for a drive with a write-focused brief, but it arrives at minimal queue depth and microsecond-class latency, the operating point where a sustained cache or staging tier lives, and the same profile that carried this drive through weeks of continuous sequential write streams in our KV cache deployment.

16K Sequential Read

On the read side, the KIOXIA CD9P-R led at 13,819.7 MB/s, with the CM9-R at 13,393.2 MB/s and the Micron 9550 Pro at 13,273.5 MB/s close behind. The PS1030 peaked at 11,142.8 MB/s (IODepth 16 / NumJobs 8) at 179.1 µs, seventh in the group, ahead of only the SN861 at 10,995.0 MB/s and just behind its sibling at 11,656.3 MB/s, about a fifth off the group lead. The single-stream picture also inverts the random-read story: on ordered 16K reads, the SN861 opened at 12.5 µs and the Microns between 13.9 and 21.3 µs, while the Solidigm platform started near 59 µs, the same low-concurrency read trait both PS-series drives showed at 64K. Mid-size reads remain the corner of the matrix where this platform gives the most away; the streaming-read side of an AI pipeline is better served by the read-intensive class, which is not the argument this drive was built to win.

4K Random Write

The headline chart for a 3 DWPD drive, and the PS1030 delivered. Its peak of 1,595.8K IOPS at IODepth 16 / NumJobs 8 was second in the group, behind only the fellow 3 DWPD Micron 7600 MAX at 1,781.2K, and ahead of the 9550 MAX (1,544.2K), PS1010 (1,504.7K), CM9-R (1,502.9K), 9550 Pro (1,467.6K), SN861 (1,438.3K), and CD9P-R (1,273.1K). That measured peak is also double the drive’s 800K rated figure, which Solidigm specs at a fixed queue depth; steady-state sweeps find more.

Latency behavior seals the argument. The PS1030 opened at 8.8 µs at IODepth 1, in the leading cluster with the CM9-R (8.2 µs) and PS1010 (8.3 µs), reached its peak throughput at just 79.8 µs, and never exceeded 359.6 µs anywhere in the sweep. The PS1010, by contrast, needed IODepth 32 / NumJobs 16 to reach a peak of 339.7 µs, on the way to a 735.6 µs worst case. For the OLTP logs and KV-cache-style traffic this drive is aimed at, small writes in the single-digit microseconds, with a sub-400 µs ceiling, are the profile that matters.

4K Random Read

The PS1030 backed up its write showing with the group’s second-best 4K random read peak: 2,255.8K IOPS at IODepth 16 / NumJobs 16 and 112.8 µs, behind the SanDisk SN861’s 2,555.6K IOPS and ahead of the Micron 9550 MAX (2,217.6K IOPS) and CD9P-R (2,165.0K IOPS). At IODepth 1 / NumJobs 1, the KIOXIA low-latency signature led again: the CM9-R at 29.3 µs and CD9P-R at 30.4 µs, but the PS1030’s 56.8 µs was the best of the rest, edging its sibling and both 9550s (roughly 65 µs) and the SN861 (67.8 µs). A mixed-use drive that lands second in both 4K read and 4K write peaks in a field this read-heavy is covering both halves of its job description.

GPU Direct Storage

One of the tests we conducted on this testbench was the Magnum IO GPU Direct Storage (GDS) test. GDS is a feature developed by NVIDIA that allows GPUs to bypass the CPU when accessing data stored on NVMe drives or other high-speed storage devices. Instead of routing data through the CPU and system memory, GDS enables direct communication between the GPU and the storage device, significantly reducing latency and improving data throughput.

How GPU Direct Storage Works

Traditionally, when a GPU processes data stored on an NVMe drive, the data must first travel through the CPU and system memory before reaching the GPU. This process introduces bottlenecks, as the CPU acts as an intermediary, adding latency and consuming valuable system resources. GPU Direct Storage eliminates this inefficiency by enabling the GPU to access data directly from the storage device via the PCIe bus. This direct path reduces data-movement overhead, enabling faster, more efficient data transfers.

AI workloads, especially those involving deep learning, are highly data-intensive. Training large neural networks requires processing terabytes of data, and any delay in data transfer can lead to underutilized GPUs and longer training times. GPU Direct Storage addresses this challenge by ensuring that data is delivered to the GPU as quickly as possible, minimizing idle time and maximizing computational efficiency.

In addition, GDS is particularly beneficial for workloads that involve streaming large datasets, such as video processing, natural language processing, or real-time inference. By reducing the reliance on the CPU, GDS accelerates data movement and frees up CPU resources for other tasks, further enhancing overall system performance.

GDSIO Sequential Read Throughput

In the 16K block-size segment, the PS1030 opened at approximately 0.2 GiB/s on a single thread, the lowest entry point in the group, consistent with its 71.6 µs single-thread GDS read latency (its sibling posted 71.1 µs, the same platform trait we flagged in the CD9P-R review). It scaled steadily into the pack through the mid-range, finishing the segment around 1.6 GiB/s at 16K/128 while the KIOXIA pair held the segment lead near 2.0 GiB/s. The KIOXIA drives’ thread-scaling advantage carried through the 128K segment as well, with the CD9P-R and CM9-R pulling away through 128K/16 while the PS1030 tracked the main group, reaching approximately 4.7 GiB/s at 128K/64 and 5.0 GiB/s at 128K/128.

In the 1M segment, the field converged. The PS1030 climbed to its peak of 5.95 GiB/s (1M/32), within 3.5% of the group-best 6.16 GiB/s from the CD9P-R, with the CM9-R at 6.06, the PS1010 and 9550 MAX at 6.05, and the 9550 Pro at 5.97 GiB/s. The 7600 MAX came in last at 5.59 GiB/s. At peak 16K IOPS, the KIOXIA pair led (136.4K for the CM9-R, 134.2K for the CD9P-R), while the PS1030’s 101.2K was the group’s lowest; small-block GPU-direct reads are simply not this platform’s strength.

GDSIO Sequential Write Throughput

The write sweep is the chart PS1030 buyers should study, because it contains both the drive’s best manners and its one real anomaly. In the 16K segment, all eight drives tracked together in the 0.5 to 1.5 GiB/s band, and the PS1030’s single-thread write latency of 22.3 µs sat in the leading cluster with the two KIOXIA drives (21.4 µs each) and its sibling (21.9 µs). In the 128K segment, the PS1030 scaled cleanly to approximately 3.95 GiB/s at 128K/32, then fell off a cliff: roughly 2.35 GiB/s at 128K/64 and 1.55 GiB/s at 128K/128, less than a third of the segment leaders’ performance. This is the same high-thread-count write collapse we documented on the PS1010 in the CD9P-R review, reproduced almost point-for-point on the PS1030 (the PS1010 fell to 2.5, then 1.65 GiB/s at the same marks). Whatever sits behind it lives in the platform, not the endurance tier, and it remains the most significant blemish in the dataset.

The 1M segment softened but did not erase the pattern. The PS1030 peaked at 4.22 GiB/s (1M/8), ahead of only its sibling’s 4.17 GiB/s, and drifted down to roughly 3.35 GiB/s at 1M/128 while the CM9-R held the most stable line in the group toward its 5.51 GiB/s peak, and the 9550 MAX hit the highest peak at 5.69 GiB/s with its own documented volatility, dipping near 2.2 GiB/s at 1M/64. The 9550 Pro (5.54), 7600 MAX (5.44), CD9P-R (4.86), and SN861 (4.59 GiB/s) filled out the order. The saving grace for the PS1030’s target buyer: KV cache and log-style tiers write at moderate thread counts per drive, where the PS1030 behaves impeccably, not at the 64-plus GPU-direct write threads where the collapse lives. Our own KV cache deployment ran eight of these drives for weeks without the array ever approaching the bottleneck. But anyone planning heavy multi-threaded GDS write streaming onto this platform should benchmark their exact pattern first.

Conclusion

The Solidigm D7-PS1030 does what a mid-endurance Gen5 drive is supposed to do, and the data shows a drive with a distinct skill set. Its signature is parallel small-block work: second in the group in both 4K random write (1,595.8K IOPS, behind only the fellow 3 DWPD Micron 7600 MAX) and 4K random read (2,255.8K IOPS), the group’s best 64K random read at 14,162.9 MB/s, and full Gen5 line rate on 128K sequential reads at 14,156.4 MB/s. Just as interesting is how it gets there: peaks at IODepth 2 or NumJobs 8 with latency in the tens of microseconds, where competitors need deep queues and pay for them in latency. The drive saturates early, holds flat, and keeps its worst-case write latency at half that of its PS1010 sibling.

There are trade-offs, however. Single-worker 128K sequential writes landed last in the group at 6,370.3 MB/s, and DLIO checkpointing, which stresses exactly that pattern, ran to the group’s highest Pass 3 average at 599.2 seconds. The 16K sequential tests also trailed the field on both sides, with the write peak of 5,977.7 MB/s and read peak of 11,142.8 MB/s each landing seventh of eight. The GDSIO write sweep reproduced the PS1010’s high-thread 128K drop nearly point-for-point, confirming it as a platform behavior rather than a one-off, and small-block GPU-direct reads favor the KIOXIA drives by a wide margin. None of this undermines the drive’s brief, but it does highlight it: the PS1030 is not the drive for single-stream ingest.

What it is built for, it has already proven in longer form than any bench run. Eight of these drives spent weeks as the KV cache offload tier in our XE7740 inference work, absorbing 1.9 GB/s of sustained, around-the-clock writes at a duty cycle that brushed past their 3 DWPD rating, and the tier never became the bottleneck. The bench data explains why that worked: early-saturating write behavior, single-digit-microsecond 4K write latency at low queue depths, and a controlled latency ceiling are exactly the traits a continuously written cache or log tier rewards. Endurance-bound, latency-sensitive, moderately concurrent write workloads, KV cache offload, OLTP logging, and metadata tiers are where the PS1030 belongs, and where its 3 DWPD budget and 4.98 DWPD three-year option let it run duty cycles that would disqualify the read-intensive class.

Solidigm D7-PS1030 Product Page

The post Solidigm D7-PS1030 Review: 3 DWPD Gen5 That Earned Its Keep in the KV Cache Tier appeared first on StorageReview.com.

Sandisk has introduced two SSD families designed specifically for network-attached storage, covering both SATA-based systems and higher-performance NVMe NAS deployments. The Sandisk NAS 600 SATA SSD offers capacities up to 4TB and endurance up to 2,500 TBW, while the Sandisk NAS 800 NVMe SSD, a PCIe 5.0 drive, scales to 7.68TB with sequential read speeds up to 14,900MB/s and endurance up to 14 PBW.

The two drives address very different parts of the NAS market. The NAS 600 uses the familiar 2.5-inch SATA form factor and can replace hard drives or serve as a cache or storage tier in existing SATA-based systems. The NAS 800 is an M.2 2280 NVMe drive designed for all-flash and hybrid NAS configurations that require substantially higher throughput and I/O performance.

Both drives feature five-year limited warranties and are designed for extended operation in NAS environments.

Specification	SANDISK NAS 600 SATA SSD	SANDISK NAS 800 NVMe SSD
Form Factor	2.5″ SATA/7mm	M.2 2280
Interface	SATA III 6Gb/s	PCIe 5.0 (NVMe)
Capacity	500GB, 1TB, 2TB, 4TB	960GB, 1.92TB, 3.84TB, 7.68TB
Endurance	Up to 2,500 TBW [4TB model]	Up to 14 PBW [7.68TB model]
Sequential Read/Write (MB/s)	Up to 560/520MB/s [4TB model]	Up to 14,900/13,200MB/s [1.92TB model]
Random Read/Write (IOPS)	Up to 86K/77K IOPS [4TB model]	Up to 2.3M/2M IOPS [3.84TB model]
MTTF	2.25M hours	1.75M hours
Warranty	5-year limited	5-year limited

Sandisk NAS 600 SATA SSD for HDD-Based NAS Upgrades

The Sandisk NAS 600 SATA SSD provides a direct solid-state option for NAS systems with conventional 2.5-inch SATA drive bays. Its 2.5-inch/7mm design means it can be installed in compatible systems without requiring an NVMe-capable NAS or M.2 storage slots. Sandisk is offering the NAS 600 in 500GB, 1TB, 2TB, and 4TB capacities. Performance tops out at sequential read and write speeds of 560MB/s and 520MB/s, respectively, on the 4TB model. Random performance reaches up to 86K IOPS reads and 77K IOPS writes.

The 4TB NAS 600 is rated for up to 2,500 TBW and has an MTTF of 2.25 million hours. The higher endurance is designed for the frequent reads and writes common in NAS environments, including backups, file sharing, and media workloads.

The drive can serve several roles depending on the NAS configuration. In addition to replacing hard drives with solid-state storage, it can be used as an SSD cache or as part of a storage tier. This gives existing SATA-based NAS owners another option for adding flash storage without replacing the NAS itself.

Sandisk NAS 800 NVMe SSD for Higher-Performance NAS Workloads

The Sandisk NAS 800 NVMe SSD moves into a very different performance class, using a PCIe 5.0 M.2 2280 design and Sandisk TLC 3D NAND. Capacities include 960GB, 1.92TB, 3.84TB, and 7.68TB, providing all-flash and hybrid NAS configurations with ample capacity per M.2 slot.

Sequential read performance reaches up to 14,900MB/s, while sequential writes reach 13,200MB/s on the 1.92TB model. Random read and write performance tops out at 2.3 million and 2 million IOPS, respectively, on the 3.84TB model. This gives the NAS 800 a broader range of potential uses than caching alone. Sandisk lists primary storage and storage tiering alongside caching, with workloads including collaborative content creation, databases, virtualization, and AI applications.

The NAS 800 also has a substantially higher endurance rating than the SATA model, reaching up to 14 PBW with the 7.68TB version. Its MTTF rating is 1.75 million hours, and the drive features the same five-year limited warranty as the NAS 600.

Sandisk rates the NAS 800 for continuous 24/7 NAS operation across a 0°C to 70°C operating range (85°C is its non-operating ceiling) and says the drive has been tested for compatibility across a range of NAS systems.

With the NAS 600 and NAS 800, Sandisk now covers both SATA-based NAS upgrades and higher-performance all-flash and hybrid systems. The NAS 600 gives existing SATA NAS users a flash option for primary storage, caching, or tiering, while the NAS 800 adds substantially higher throughput and endurance for more demanding NVMe-based storage workloads.

Availability

The Sandisk NAS 600 SATA and NAS 800 NVMe SSDs are slated for release sometime in September. Sandisk’s own store already lists pricing ahead of availability, showing the 500GB NAS 600 at $179.99 and the 7.68TB NAS 800 at $2,199.99, both marked down from list prices of $224.99 and $2,749.99. The NAS push also caps a busy month for Sandisk, which just taped out its first HBF memory die on the datacenter side.

The post Sandisk NAS 800 Brings PCIe 5.0 and 14,900MB/s to NAS, With the SATA NAS 600 Covering Legacy Bays appeared first on StorageReview.com.

Sandisk has taped out the first High Bandwidth Flash memory die. The company put it on a slide at its 2026 Investor Day on August 13, under the header HBF Roadmap, next to what the slide labels an actual die picture. The second half of that slide covers the first HBF inference product samples: coming soon, 2027.

A tapeout is a substantial milestone with any new silicon. It means the design is finished and committed to a mask set, which is the point where an architecture stops being a slide and starts being silicon. It is also several steps short of a product. The wafers have to come back from the fab, the die has to hit its target specs, yield has to climb to something economic, and the die then has to survive stacking into 8-high and 16-high configurations with a working logic die and controller underneath. After that comes thermal and endurance qualification, the accelerator software work to actually address the memory, and customer qualification cycles that run in quarters. Sandisk is telling investors that the first of those steps is done.

What Sandisk Says HBF Delivers

HBF stacks NAND rather than DRAM and puts it in an HBM-style package next to the accelerator. Per Sandisk’s own HBF fact sheet, the first generation targets 512GB per stack, built from sixteen 256Gb die, at 1.6TB/s of read bandwidth. Sandisk claims that lands at up to 8 to 16 times the capacity of HBM at a similar cost, in a package that closely matches HBM4’s footprint, stack height, and power profile. Because it is NAND, it is non-volatile and spends no power on refresh.

The roadmap on the fact sheet runs further. A second generation targets more than 2TB/s and up to 1TB per stack at 0.8 times the first generation’s power, and a third pushes past 3.2TB/s and up to 1.5TB per stack at 0.64 times the power. The Investor Day deck frames the target workloads plainly: mixture-of-experts LLMs, long context lengths, and large KV caches, with the architecture developed using input from major cloud and AI customers.

The deck also lays out three deployments. HBF can augment HBM, filling some of the stack positions around an xPU while HBM keeps the rest. It can replace HBM stacks outright in a similar footprint. Or it can sit disaggregated, holding decode weights and KV cache while a smaller HBM tier acts as cache. That flexibility matters more than it sounds, because it lets HBF into a socket without requiring an accelerator vendor to abandon HBM.

The Internal Numbers Sandisk Put On Screen

Sandisk showed an inference token output comparison covering one HBF GPU, four HBF GPUs, and eight HBM GPUs. From it, the company draws two claims, both labeled as based on internal testing. The first is an 8x capex efficiency figure, defined as the minimum configuration required to run the model: one HBF GPU against eight HBM GPUs. The second is a 2x GPU efficiency figure: four HBF GPUs delivering the same token output as eight HBM GPUs. The fact sheet adds a related simulation result: HBF landing within 2.2% of unlimited-capacity HBM when reading pretrained weights for Llama 3.1 405B.

These are vendor numbers on unreleased silicon, and the chart in question carries no axis values for tokens per second. They describe the concept though: if a model fits in memory that is eight to sixteen times larger for the same bandwidth and roughly the same power, you need fewer accelerators to hold it, and the ones you have spend less time waiting.

Where the Timeline Stands

For planning purposes, two Sandisk statements bracket the schedule. In August 2025, the company said the first HBF samples were targeted for the second half of calendar 2026, with AI inference devices using HBF expected in early 2027. As of the August 2026 Investor Day, the first HBF inference product samples are listed as coming soon in 2027. Schedules for new memory classes firm up as designs become silicon, and first-generation platforms across this cycle have moved to the right as qualification realities set in. The tapeout is the evidence that HBF is progressing; however.

The Ecosystem Piece Is Further Along Than the Silicon

The standards work is moving faster than the product. On August 3, Sandisk and SK hynix released the first HBF technical specification through the Open Compute Project, six months after the consortium formed, with Google and Tenstorrent among the contributors. We covered that specification and what it defines when it landed.  Tensorent’s Jim Keller was named as a technical advisory board member at the event.

SK hynix, which co-authored the spec and showcased its own tiered memory pitch at FMS 2026, is the other half in determining whether this becomes a standard or a single-vendor product. For now, Sandisk has a taped-out die, a spec in the open, and a 2027 date on samples. The next real checkpoint is silicon that measures up, and that will come from a fab report, which hopefully Sandisk talks more about in the near future.

The post Sandisk Tapes Out Its First HBF Memory Die, Targets 2027 for Inference Product Samples appeared first on StorageReview.com.

Enterprise SSD prices increased another 5 percent in July 2026, according to the latest VDURA Flash Volatility Index. While the monthly increase was lower than the extreme fluctuations seen over the past year, flash pricing remains approximately 6.5 times higher than it was in the third quarter of 2025.

The index, which VDURA launched in January and updated August 11, places the price of a 30TB TLC enterprise SSD at $22,600, compared with $3,460 in Q3 2025. A 30TB QLC SSD now costs $18,080, up from $2,768 during the same period.

The TLC-to-HDD Multiple Sits at 18.6x

Hard drive pricing has also increased. VDURA lists a 30 TB HDD at $1,216, compared with $495 in Q3 2025. Despite the increase, the price difference between 30TB TLC SSDs and 30TB HDDs remains substantial. The current 18.6x multiple is lower than the 23.2x peak recorded in Q1 2026, but significantly higher than the 7.0x multiple measured a year earlier.

The pricing gap is particularly relevant to AI infrastructure. Training reads, checkpoint write bursts, and inference lookups require flash performance, while training datasets, checkpoint histories, and model archives can represent much larger capacity requirements. Placing all of that data on premium flash can significantly increase infrastructure costs.

“A 5% monthly increase on top of a 6.5x year-over-year jump is not relief. It is confirmation that elevated flash pricing is structural, and every AI cloud and AI factory business plan needs to account for it,” said Erik Salo, VDURA’s senior vice president of marketing and business operations.

25PB AI Factory Reference Deployment

VDURA used its Storage Economics Optimizer Tool to model a 25 PB storage deployment that delivers 1,000 GB/s of sustained read performance. The reference system is sized for a cluster of approximately 2,000 GPUs, based on guidance from AMD and NVIDIA.

Quarter	30TB TLC SSD	30TB QLC SSD	TLC vs HDD	30TB HDD
Q3 2025	$3,460	$2,768	7.0x	$495
Q4 2025	$7,765	$6,212	13.4x	$580
Q1 2026	$17,500	$14,000	23.2x	$755
Q2 2026	$18,900	$15,120	16.3x	$1,158
Q3 2026	$22,600	$18,080	18.6x	$1,216

At Q3 2026 pricing, an all-flash design using 30TB TLC SSDs carries an estimated three-year cost of $51.60 million. A design combining storage-class memory with QLC flash totals $48.42 million over the same period.

VDURA’s mixed-fleet architecture uses 5.78 PB of flash and 22.68 PB of HDD capacity within a single namespace. The configuration delivers 1,040 GB/s of sustained performance at an estimated three-year cost of $12.86 million.

The resulting difference is approximately $38.74 million compared with the TLC all-flash design. The comparison illustrates how separating performance and capacity requirements can reduce the amount of data stored on NVMe media without limiting the performance available to hot workloads.

For AI cloud operators, storage cost directly affects the economics of GPU utilization and token delivery. Training data, checkpoint operations, and inference lookups can remain on flash, while less performance-sensitive data is placed on hard drives. This approach limits exposure to elevated flash pricing while retaining NVMe performance where it matters most.

VDURA said the current pricing environment challenges the assumptions behind all-flash AI storage architectures. The company characterized the combination of a 5 percent monthly increase and a 6.5x year-over-year increase as evidence that flash pricing has become a structural consideration for AI cloud and AI factory deployments rather than a temporary market fluctuation.

The VDURA Flash Volatility Index and Storage Economics Optimizer Tool provide pricing data and modeling capabilities for comparing storage media mixes, capacity requirements, GPU counts, and performance targets. VDURA said it plans to update the index as market conditions change.

The post Enterprise SSD Prices Run at 6.5x Last Year: VDURA Pegs a 30TB TLC Drive at $22,600 appeared first on StorageReview.com.

SK hynix and its subsidiary Solidigm are following a two-part manufacturing plan to keep mature flash production separate from next-generation memory. Recent updates show they are expanding older production capacity overseas while also investing heavily at home. These moves will support future demand for AI storage, but current production schedules mean that enterprise SSD shortages are likely to continue for now.

Dalian Fab 2 Restart and Offshore Capacity Expansion

Solidigm has restarted investment in its second factory in Dalian, China, after about four years, according to the Seoul Economic Daily. The building for Fab 2 is finished, and equipment could start arriving as early as November 2026. The reports put full production in the first half of 2027. Neither Solidigm nor SK hynix has announced the restart through its own channels, so the timing rests on Korean press reporting for now.

If completed, the new line could add about 50,000 wafer starts per month to the current Dalian Fab 1 output of around 100,000 wafers per month. This would increase the company’s flash manufacturing in China by about 50 percent. The Dalian site will continue to focus on mature-node NAND, helping Solidigm boost production of mainstream enterprise flash products while adhering to rules on advanced semiconductor equipment.

Multi-Trillion-Won Domestic Buildout for Advanced Nodes

Alongside its overseas expansion, SK hynix announced in early August a ₩54 trillion (about $38 billion) investment in its South Korean factories, according to company reports. Most of the money will go to two main sites: ₩35.2 trillion for the Yongin Y2 facility, which will make advanced DRAM and high-bandwidth memory, and ₩19.1 trillion for the Cheongju M17 fab, which will focus on next-generation NAND flash and enterprise SSDs.

The M17 investment will create a dedicated site for making high-layer, high-density flash needed for large-scale AI tasks. However, this new capacity will take time to come online. The first cleanroom at M17 should be finished by the end of 2028, with commercial production starting in 2029 or later.

Market Position and Enterprise SSD Pricing Implications

TrendForce data for the first quarter of 2026 shows that SK hynix Group, including Solidigm, stayed in second place for global NAND flash revenue. The group posted $7.53 billion in revenue for a 17.6 percent share, behind Samsung’s 31.6 percent. To keep or grow this share, they need to increase production of both mature and advanced memory types.

The demand side explains the urgency. Counterpoint Research’s second-quarter tracker puts enterprise SSDs at 48 percent of global NAND shipments, nearly double the 26 percent share a year earlier, and the firm expects eSSDs to absorb more than half of all NAND bits by the end of 2026. Solidigm stood out in that data, growing bit shipments roughly 40 percent quarter over quarter per Counterpoint, and that mature-node, high-capacity product line is exactly what the reported Dalian expansion would feed.

The main issue is still how much product is available and at what price. By sending mature-node production to Dalian and focusing on advanced memory at home, SK hynix will not solve the current shortage of enterprise SSDs. Since Dalian Fab 2 will not start producing until at least the first half of 2027, and Cheongju M17 will not be ready until the end of the decade, enterprise storage supplies will stay limited for now.

The post SK hynix and Solidigm Split the NAND Map: 54 Trillion Won at Home, a Reported Dalian Fab 2 Restart in China 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.

DapuStor used FMS 2026 to show what it calls an industry-first 512TB SSD: a new top capacity for the R6060 PCIe 5.0 QLC line that doubles the company’s previous 245TB design and puts a full petabyte of flash in just two drives. The drive was demonstrated on the show floor through a performance test video, alongside a liquid-cooled E1.S model, mixed-mode QLC solutions, and dual-port E3.S drives spanning TLC and QLC.

The 512TB figure resets the top of the announced-capacity table. The largest SSD we have tested is Micron’s 245.76TB 6600 ION, and DapuStor’s own R6060 came through the lab in May in its 122.88TB configuration, where its read-heavy Gen5 QLC design held up well at scale. DapuStor has not published performance specifications for the 512TB model, so for now, the claim rests on the show-floor demonstration.

512TB R6060: A Petabyte in Two Drives

The new R6060 arrives in next-generation EDSFF form factors, including E3.L and E2, the emerging EDSFF variant aimed at maximum-capacity drives. DapuStor positions the drive for large AI datasets and capacity-intensive infrastructure, with the familiar density argument: fewer drives per rack for a given capacity target, which reduces slot count, power draw, and overall TCO. Those are the vendor’s claims, and they will ultimately depend on pricing, endurance, and performance figures that the company has not yet disclosed.

Liquid-Cooled E1.S for Dense AI Servers

Alongside the capacity play, DapuStor introduced an 8TB R6 PCIe 5.0 TLC SSD in E1.S with cold-plate liquid-cooling support. The compact form factor targets dense, GPU-heavy AI servers and liquid-cooled data center environments, where flash increasingly shares the cooling loop with the accelerators it feeds.

Mixed-Mode QLC and Dual-Port E3.S

The J5060 QLC Series supports mixed SLC and QLC deployment, letting a system carve a high-performance flash tier and a high-capacity QLC tier within the same architecture. DapuStor also showed R6 E3.S dual-port enterprise SSDs in both TLC and QLC configurations, with TLC models scaling to 30.72TB and QLC to 61.44TB; dual-port support adds path redundancy for high-availability enterprise and AI storage systems.

DapuStor ran the show from Booth #115, and on the final day, its VP of R&D, Xiang Chen, presented a technical session on optimized solutions for large-capacity QLC SSDs in the AI era. The company was one of several vendors leaning hard into QLC capacity at FMS 2026, where the show’s throughline was flash positioning itself as the capacity tier for AI infrastructure. No availability dates or pricing were announced for the 512TB R6060; we will follow up as the drive moves toward production, and our full test data on the current R6060 is in the 122TB review.

The post DapuStor Shows a 512TB QLC SSD at FMS 2026: 1PB of Flash in Two Drives appeared first on StorageReview.com.

Smart IOPS and H3 Platform have announced an AI compute storage platform targeting up to one billion random-read IOPS. The proposed system combines Smart IOPS’ Unobtanium T50 solid-state devices with an air-cooled H3 Platform appliance designed around NVIDIA accelerated computing and networking technologies.

The platform targets data-intensive AI inference and analytics workloads that require GPUs to access large datasets through highly parallel, fine-grained I/O. It is designed to support NVIDIA’s Storage-Next initiative and SCADA, or Scaled Accelerated Data Access, programming model. These technologies enable GPU-initiated access to NVMe storage when application working sets exceed local GPU memory.

By reducing the distance between GPU compute and storage, the architecture is intended to reduce CPU control-path overhead and help keep accelerators supplied with data. The approach is aimed at workloads such as graph neural networks, vector search, recommendation systems, retrieval-augmented generation, embedding stores, and other sparse analytics applications that generate large volumes of small, random requests.

“AI infrastructure is entering a phase in which storage must behave less like a passive repository and more like an active extension of the compute fabric,” said Ashutosh Das, CEO of Smart IOPS.

Smart IOPS Unobtanium T50

The Unobtanium T50 is a PCIe Gen6 x4, NVMe 2.0 device in an E3.S form factor. Smart IOPS specifies up to 50 million random-read IOPS and 10 million random-write IOPS at a 512-byte block size. Sequential performance is rated at up to 28 GB/s for reads and 24 GB/s for writes. Sequential throughput in that range matches PCIe Gen6 enterprise SSDs already entering mass production; the 512-byte random-read figure is the specification that stands apart.

Four T50 devices can provide a stated aggregate target of up to 200 million random-read IOPS at the device level. Smart IOPS positions that capability as a match for the I/O requirements of next-generation PCIe Gen6 x16 GPUs, including NVIDIA Rubin. Actual performance will depend on workload characteristics, software, PCIe topology, system configuration, and other system-level factors.

Smart IOPS ACS Unobtanium T50
	Preliminary Specification
Performance
Random read	Up to 50 million IOPS
Random write	Up to 10 million IOPS
Block size	512 bytes
Sequential read	Up to 28 GB/s
Sequential write	Up to 24 GB/s
Hardware
Capacity	9 TB, 18 TB and 36 TB
Media	TLC NAND operated in pseudo-SLC mode; support for specialized low-latency SLC NANDs
Form factor	E3.S 2T
Interface	PCIe Gen6 x4; NVMe 2.0
Availability
Planned availability	Evaluation samples: Q1 2027; production: Q2 2027

H3 Platform High IOPS Appliance
	Preliminary Specification
Compute & Networking
GPU	4 × NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs
Networking	4 × NVIDIA ConnectX-8
Storage
SSD bays	20 × E3.S 2T slots for high-IOPS SSDs
Per-slot power	Up to 60 W per E3.S slot
System
Cooling	Air-cooled
Availability
Planned availability	Evaluation systems: Q1 2027; production: Q2 2027
Preliminary specifications; subject to change without notice.

The T50 uses the Smart IOPS TruRandom controller architecture with widely available TLC NAND operated in pseudo-SLC mode. The company says this approach allows the devices to use multiple qualified NAND suppliers while retaining access to the manufacturing scale and supply-chain capacity of the broader flash industry. The controller can also support specialized low-latency NAND where the workload or deployment economics justify it.

H3 Platform Appliance

H3 Platform is developing an air-cooled appliance with four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, four NVIDIA ConnectX-8 network adapters, and 20 E3.S 2T storage slots. Each slot is designed to support up to 60 watts, providing the power and thermal capacity required by high-performance storage devices.

With 20 slots populated by T50 devices, the system is designed to reach a target of up to one billion random-read IOPS. The configuration is intended to balance GPU compute, networking, and storage connectivity in a single platform for evaluating and deploying GPU-initiated, small-block storage workloads.

“The next wave of AI systems must be designed as an integrated data path,” said Brian Pan, founder and CEO of H3 Platform, “not as disconnected compute, network and storage components.” The design places high-IOPS E3.S storage alongside GPUs and high-speed network adapters while retaining an air-cooled form factor.

High IOPS is relevant to these workloads because bandwidth alone does not describe how effectively a storage system handles irregular, concurrent accesses. Small-block requests generated by many GPU threads can expose queuing and synchronization overhead in CPU-orchestrated storage architectures. A higher-IOPS, lower-latency device can service more concurrent requests and potentially reduce the time GPUs spend waiting for data.

Smart IOPS and H3 Platform expect evaluation samples and systems to become available in the first quarter of 2027. Production availability is planned for the second quarter of 2027. The specifications, configurations, and schedules remain preliminary and may change. Customer qualification, software ecosystem readiness, and any NVIDIA validation or approval are separate processes and are not implied by the announcement.

The post Smart IOPS Unobtanium T50: 50 Million IOPS Per Gen6 SSD, With a One Billion IOPS Appliance Target appeared first on StorageReview.com.

Netlist and Samsung have signed a five-year set of agreements covering patent cross-licensing, DRAM and NAND supply, technology cooperation, and the settlement of all pending litigation between the companies. “We’re excited to renew this partnership and look forward to working closely with Samsung,” said Netlist CEO C.K. Hong, a nod to the joint development agreement the two companies signed in 2015 before years of litigation. “These strategic agreements reflect the companies’ shared commitment to innovation in the AI-memory space and validates the value of Netlist’s IP.”

The deal gives Samsung a worldwide license to specified Netlist patents, including technology related to server DIMMs and High Bandwidth Memory. The license is royalty-bearing for Samsung, while Netlist receives a royalty-free license to specified Samsung patents. Samsung will pay Netlist an upfront patent license fee of $239 million, or approximately $200 million after Korean withholding taxes, followed by quarterly license payments of up to $32.9 million gross through the second quarter of 2031. The recurring payments are based on a revenue formula and remain subject to contractual adjustments and refund rights.

Netlist said the arrangement could generate up to $750 million in net license fees over its five-year term. The agreement replaces years of patent disputes across U.S. district court, International Trade Commission, and Patent Trial and Appeal Board proceedings. Those disputes included a $303 million jury award in favor of Netlist against Samsung in 2023, although prior litigation outcomes had remained subject to appeal and patent-validity proceedings.

A separate supply agreement allows Netlist to purchase up to $300 million annually in DRAM and NAND products from Samsung Semiconductor, with aggregate purchasing capacity of up to $1.5 billion over five years. The agreement gives Netlist a defined source of memory components that can be used in its own specialized products or supplied to customers through its existing sales channels. The supply limit represents a maximum purchasing capacity, not a minimum purchase commitment.

Samsung Semiconductor also agreed to buy 10 million shares of Netlist common stock for $1 million in a private transaction. The shares will be subject to a five-year transfer restriction, with 20% released on each of the first four anniversaries and the balance released in the fifth year.

The companies also signed a five-year ITC cooperation agreement under which Samsung will provide specified information, documents, or declarations for Netlist’s future ITC actions involving third parties. Netlist’s dispute resolution with Samsung therefore does not end its broader IP enforcement activity.

While technical development deliverables were not disclosed, the natural candidates are the areas Netlist has been developing publicly: hybrid CXL memory and low-power MRDIMM designs. Those technologies address two emerging infrastructure concerns: expanding memory capacity at a lower cost than DRAM-only configurations, and reducing power consumption in next-generation high-density memory modules.

Primary terms are disclosed in Netlist’s Form 8-K filing.

The post Netlist and Samsung Sign Five-Year Patent, Supply, and Technology Agreements appeared first on StorageReview.com.

CoreWeave said on August 5 that it has signed a multi-year agreement with Solidigm for priority access to enterprise SSD capacity. Storage has become the second component after GPUs that an AI cloud has to contract for years in advance, and CoreWeave has now done for flash what cloud providers spent the first half of 2026 doing for DRAM.

What the Agreement Says, and What It Does Not

The disclosed terms amount to three things: the agreement is multi-year, it covers priority access to enterprise SSD capacity, and it is meant to keep storage scaling with customer demand on CoreWeave’s integrated AI cloud platform.

What both executives chose to talk about is telling. Sachin Jain, chief operating officer at CoreWeave, framed it entirely around supply certainty: the agreement “gives us priority access to the storage our roadmap depends on, allowing customers to focus on what they are building instead of the infrastructure underneath it.” Paul Palonsky, executive vice president and head of global sales at Solidigm, said that the agreement “reflects the confidence Solidigm has in CoreWeave’s growth and its position as a full-stack AI cloud provider.” CoreWeave’s own framing in the release is that storage has become a critical constraint in capacity planning as enterprise AI adoption accelerates and industry-wide storage supply tightens.

Why Flash Allocation Became a Scheduling Problem

The memory market explains the deal better than the press release does. TrendForce projects NAND flash contract prices rising 10 to 15 percent quarter over quarter in the third quarter of 2026, with conventional DRAM up 13 to 18 percent, and describes the DRAM market as extremely tight. Enterprise SSD supply is the one bright spot in that picture, improving as vendors shift capacity away from consumer segments. However, TrendForce still expects a clear NAND shortage across 2026, with meaningful capacity expansion unlikely before late 2027 or 2028.

Buried in the same forecast is the mechanism behind this announcement. TrendForce notes that price gains are moderating in part because long-term supply agreements now govern a portion of procurement, and its earlier take on the year was that cloud providers are willing to accept higher prices and sign LTAs to secure stable supply. An LTA is what a buyer signs when the binding constraint is allocation rather than price. CoreWeave is not trying to get cheaper SSDs; it’s just trying to be first in line.

The pattern is not confined to flash. Western Digital told investors last week that its hard drive long-term agreements now extend into calendar 2029 through 2031, which we covered in our look at WD shipping 40TB UltraSMR drives. On the supply side, SK hynix is committing KRW 100 trillion to its Cheongju fabs precisely because AI has pushed NAND demand past supply. Every layer of the storage stack is being contracted forward at once.

What CoreWeave Is Actually Buying

CoreWeave is large enough that a supply agreement is a major commitment on Solidigm’s side. First quarter 2026 revenue was $2.078 billion, up 111.6 percent year over year, against a revenue backlog of $99.4 billion as of March 31. The company spent $7.695 billion on property, equipment, and capitalized software in that quarter alone, guided full-year 2026 capital expenditures to a range of $31 billion to $35 billion, and crossed 1GW of active power with more than 3.5GW contracted. Storage is a small slice of that capex next to GPUs and buildings, but the absolute number is still large, and it has to arrive on the same schedule as everything else.

An AI cloud consumes enterprise SSD capacity in three places. Checkpointing during training writes enormous files at high frequency. Object storage tiers feed data to GPU nodes fast enough to keep them busy, which is the problem CoreWeave was addressing when it removed egress fees from data migration in its zero egress migration push. The third is newer and growing fastest: KV cache offload during inference.

We have measured that last one directly on Solidigm hardware. In our testing of KV cache offload to flash, a Dell PowerEdge XE7740 with four NVIDIA RTX PRO 6000 Blackwell GPUs and eight Solidigm D7-PS1030 12.8TB drives cut worst-case time to first token on resumed sessions from 13.9 seconds to 3.2 seconds. It held roughly 30,000 tokens per second at sustained load, 2.2 times the VRAM-only baseline. Moving KV cache off VRAM onto flash converts a memory capacity ceiling into a storage capacity purchase. Every AI cloud that adopts it needs more SSDs than it did the year before, and that ramp isn’t slowing.

Density is the other half of the argument. Solidigm’s current flagship is the D5-P5336, a 122.88TB drive built on 192-layer QLC NAND, which we reviewed last year at 7GB/s sequential reads and 900,000 IOPS on 4K random reads, drawing 24W active with a 0.6 DWPD endurance rating. In a Dell PowerEdge R7725xd, that capacity point puts nearly 3PB in a single chassis. For an operator filling gigawatts of contracted power, the difference between 61.44TB and 122.88TB per bay is measured in racks, floor space, and switch ports, not just in dollars per terabyte.

Solidigm’s Side of the Timing

The same day CoreWeave published this release, the Korea Economic Daily reported that Solidigm was pursuing a pre-IPO capital raise ahead of a possible Nasdaq listing, with figures in the coverage ranging from roughly 5 trillion to 10 trillion won. SK hynix responded the next day in a regulatory filing, saying that its overseas subsidiary “is reviewing various measures to strengthen its competitiveness” but that “no matters have been determined as of the date hereof.” The company committed to disclosing details either when they are confirmed or within one month of that August 6 filing.

So the raise is unconfirmed, and the reported figures do not agree with each other. What is on the record is that SK hynix reorganized its wholly owned US NAND unit under a new California entity in January, with Solidigm described as the anchor asset of a $10 billion American AI investment vehicle, and that the company now has a self-imposed September deadline to say more. A marquee multi-year supply agreement with the highest-profile AI cloud on the market is a useful item to have on the sheet during a month like that. Neither company drew the connection, and we are not suggesting the announcement was timed to anything, but the sequence is worth considering when reading a release this light on specifics.

Solidigm has also said where its capacity ceiling is headed. Roger Corell, the company’s senior director of AI and leadership marketing, has stated publicly that Solidigm plans to ship 245TB-class drives before the end of 2026. That would match the density Micron reached with the 6600 ION, which we put through the lab in June at a quarter petabyte per drive bay. Solidigm does not pre-announce products, so there is no SKU or date beyond that statement.

What to Watch

Three things will show whether this release is a milestone or a placeholder. The first is whether either company ever attaches a number to it. Multi-year with no capacity, no term, and no value is a signal to customers and investors rather than a contract disclosure. If a figure surfaces later in a CoreWeave filing, it will be worth comparing against the language used here.

The second is whether the 245TB-class Solidigm drive ships on the stated timeline and whether CoreWeave is an early taker. Priority access matters most at the top of the capacity stack, where allocation is tightest, and the rack math changes the most.

The third is whether other AI clouds follow with named SSD supply agreements of their own. Power contracts are already standard disclosure for this class of operator. If flash allocation joins them, this announcement will signal the first of a pattern rather than a one-off. There is a nearer date than that, too: SK hynix has told regulators it will say more about Solidigm’s capital plans by early September, and whatever it discloses will set the context for how much weight this agreement was meant to carry.

The post CoreWeave Locks In Solidigm SSD Supply as Flash Allocation Becomes an AI Cloud Problem appeared first on StorageReview.com.

Sandisk and SK hynix have released the first HBF (High Bandwidth Flash) technical specification through the Open Compute Project, six months after the two companies formed the HBF workstream in February. Google and Tenstorrent joined the consortium during the standardization process, and the spec is now openly available to any company designing AI inference systems or accelerators.

HBF is a NAND-based answer to a DRAM problem. AI inference wants high-bandwidth memory close to compute, but HBM capacity is limited and expensive, while model sizes and context windows keep growing. HBF slots between the two extremes: flash-based stacks that trade some of HBM’s speed for far greater capacity and persistence, giving system designers a tier for model weights and long-context data that no longer fits in DRAM but cannot live on an SSD.

The specification puts real numbers behind the concept. Per SK hynix, HBF stacks come in 8-high and 16-high NAND configurations at up to 512GB per stack, with three performance grades spanning 0.4TB/s to 3.0TB/s, connected over the UCIe standard. The document defines the xPU-HBF host interface, electrical guidelines, baseline performance expectations, reliability and packaging guidance for an HBF die stack, and a software user guide for read and write operations. The companies designed the spec so HBF can coexist with HBM in the same system rather than replace it.

Sandisk and SK hynix are explicit about the strategy: publish early and openly through OCP to position HBF as the de facto standard for the AI storage market before rivals define an alternative. “AI inference is creating a new set of memory requirements, and HBF technology is designed to meet that moment,” said Alper Ilkbahar, chief technology officer at Sandisk, calling the spec a milestone “for the next generation of AI systems built to improve token economics at scale.” The capacity math is the same one driving KV cache offload to flash: keeping context in a cheaper, denser tier raises how many tokens and users each accelerator can serve.

BiCS10 QLC Claims the Density Crown

Sandisk’s other FMS announcement came with Kioxia: BiCS10 QLC, the pair’s tenth-generation quad-level-cell 3D NAND. The companies say the new die delivers a 60% bit density increase over their 8th-generation QLC and, at more than 37Gb/mm², the industry’s highest bit density for QLC NAND. It carries the same 332 active layers as the BiCS10 TLC die that began sampling in July, built on the same CBA (CMOS directly Bonded to Array) process that manufactures logic and memory array on separate wafers before bonding them.

On the interface side, BiCS10 QLC runs Toggle DDR6.0 at 4.8Gb/s with a Separate Command Address protocol, and adds Power-Isolated Low-Tapped Termination to improve data-out transfer efficiency. “Our 10th-generation QLC 3D flash memory delivers simultaneous gains in density, bandwidth, and energy efficiency,” Ilkbahar said, while Kioxia CTO Hideshi Miyajima pointed at the target market: efficiently storing rapidly expanding data volumes for AI systems. Neither company gave a sampling date for the QLC die.

Keynote Wednesday, Memory Wall Panel Thursday

Sandisk’s FMS keynote, “NAND: The Versatile & Scalable Foundation of the AI Era,” runs Wednesday, August 5 at 11:40 a.m. PT, presented by chief revenue officer Jim Elliott, chief product officer Khurram Ismail, and Ilkbahar. On Thursday at 9:45 a.m. PT, Sandisk, SK hynix, and Google share a panel titled “Breaking the Memory Wall with High Bandwidth Flash,” moderated by Tom Coughlin, digging into architectural integration, standardization timelines, and the economics of the new tier. Sandisk is showing HBF, next-generation NAND, enterprise SSDs, and its updated AI Data Cycle framework at booth #607.

The spec release comes in a week thick with memory-hierarchy news; Samsung used its own FMS keynote to preview zHBM and zNAND-O concepts aimed at the same AI inference wall. The difference is that HBF now has a published, open specification behind it, and with Google and Tenstorrent already in the consortium, the standards race for flash-as-memory has a front-runner. All figures above are vendor-stated; we hope to see HBF hardware in the lab before anyone ships production systems built on it.

The post High Bandwidth Flash Gets Its First Open Spec: 512GB Stacks and Up to 3.0TB/s appeared first on StorageReview.com.

KIOXIA has introduced the GP1 Series PCIe 6.0 NVMe SSD, the first product in its GP Series of Super High IOPS drives optimized for GPU direct access. KIOXIA describes it as building on the GP Series technology it introduced in March, now realized in shipping silicon. Built around second-generation KIOXIA XL-FLASH memory, the drive is meant to serve as a flash-based memory extension tier for AI systems that need lower latency and faster access to large datasets.

The GP1 Series extends High Bandwidth Memory (HBM) with a high-speed flash storage layer. Instead of relying only on adding more HBM, AI platforms can use flash memory to expand capacity and improve GPU utilization. KIOXIA says this approach reduces the cost of scaling AI memory compared to expanding HBM alone.

KIOXIA GP1 Specifications

Specification	Details
Interface	PCIe 6.0, NVMe 2.2
Flash Memory	KIOXIA XL-FLASH generation 2 low-latency flash memory
Random Read Performance	Up to 10 million random read IOPS (512-byte block size)
Form Factors	E3.S, E1.S 9.5mm, E1.S 15mm
Cooling Support	Cold-plate liquid cooling supported on E3.S and E1.S 9.5mm models; all form factors also support traditional air-cooled environments
Endurance	Up to 50 DWPD

 

The GP1 delivers up to 10 million random-read IOPS at a 512-byte block size, with KIOXIA noting the usual caveat that real-world numbers vary by host, drivers, and workload. The company says the architecture is designed to scale from today’s 10 million random read IOPS to future generations targeting up to 100 million IOPS.

The drives use KIOXIA’s second-generation XL-FLASH memory, designed to provide lower latency and finer-grained 512-byte data access than conventional TLC-based SSDs. KIOXIA says this also reduces power consumed per I/O operation, an important factor for dense AI deployments. That is the dividing line between this drive and the CM10 KIOXIA announced days earlier: the CM10 is a capacity-oriented enterprise drive on 332-layer BiCS FLASH generation 10 TLC, while the GP1 trades capacity for latency and IOPS on XL-FLASH.

“The AI memory wall is a critical challenge for our industry to solve on the way to growing the scale and capability of AI deployment,” said Neville Ichhaporia, senior vice president and general manager of the SSD business unit at KIOXIA America. “KIOXIA’s XL-FLASH is uniquely positioned to deliver the extremely high performance and low-latency data access needed to support GPUs as a memory extension tier. It also offers a significantly lower cost per gigabyte than expanding in-node HBM or DRAM capacity.”

Several enterprise form factors will be available, including E3.S and E1.S in 9.5mm and 15mm heights. The E3.S and E1.S 9.5mm versions support cold-plate liquid cooling. All models can operate in traditional air-cooled server environments.

KIOXIA GP1 Availability

KIOXIA plans to provide evaluation samples of the GP1 Series to select customers by the end of 2026, noting that the samples are for functional check purposes only and that final specifications may differ in mass production. The drives will be shown at FMS: the Future of Memory and Storage, taking place August 4 to 6 in Santa Clara, California.

The post KIOXIA GP1 Series Hits 10 Million Random Read IOPS on XL-FLASH Gen 2 appeared first on StorageReview.com.

ScaleFlux has introduced an AI-optimized SSD platform designed for NVIDIA CMX and other inference architectures that use SSDs as a shared KV-cache tier beyond GPU HBM and host DRAM. The platform combines high-endurance SSD hardware, Flexible Data Placement (FDP) support, and workload telemetry intended to improve data placement, reduce write amplification, and extend effective endurance in high-churn inference environments.

The platform addresses three storage challenges with offloaded KV cache: characterizing real-world workload behavior, separating data by lifecycle, and sustaining heavy write activity without using excess flash capacity to absorb writes.

Long-context inference, shared-prefix reuse, agentic application flows, and retained idle sessions increase the volume of reusable runtime state stored outside GPU memory. Unlike conventional enterprise workloads, KV-cache blocks may be written frequently, retained for varying periods, reactivated after inactivity, and invalidated asynchronously across sessions, workers, and tenants. These patterns increase garbage collection activity and write amplification when data with incompatible lifecycles share the same flash blocks.

ScaleFlux’s high-endurance architecture is designed to deliver 7 to more than 10 effective drive writes per day at five years for KV cache workloads, with the company noting that effective endurance depends on workload characteristics, FDP utilization, and device configuration. Higher effective endurance reduces the raw flash capacity operators must deploy purely to absorb write traffic, leaving more installed capacity available to hold active KV cache and other AI runtime state. ScaleFlux calls that overhead the “endurance tax,” and reducing it is the platform’s core economic argument.

The SSD platform supports over 200 FDP write streams per drive. This lets inference software group data by lifecycle, session, tenant, shared-prefix classification, ownership, or reuse behavior before placing it on flash media. The goal is to write data with similar invalidation patterns together, reducing internal data movement during garbage collection and limiting interference across data classes.

In preliminary controlled testing, ScaleFlux measured more than a twofold reduction in write amplification using lifecycle-aware FDP placement compared with a baseline placement configuration. The company notes that actual results depend on workload characteristics, lifecycle classification, software integration, and device configuration.

“AI inference infrastructure needs SSDs that provide more than additional capacity,” said Hao Zhong, CEO and co-founder of ScaleFlux. “Infrastructure teams need to understand how KV workloads affect the drive, separate data according to lifecycle, and sustain high write rates without deploying excess capacity simply to dilute writes. ScaleFlux brings workload intelligence, scalable FDP placement, and 7-10+ effective DWPD together in one AI-optimized SSD platform.”

At the software and telemetry layer, ScaleFlux Context-Insight SSD shows how KV-cache policies affect SSD operation. The platform captures latency, queue depth, throughput, request-size distribution, data age, write-to-first-read intervals, read reuse, NAND write volume, garbage collection movement, and write amplification.

Context-Insight can operate in SSD-only mode for initial workload analysis without changes to upper software layers. With deeper integration, it correlates SSD telemetry with application metadata, including session IDs, worker or tenant identifiers, shared-prefix IDs, KV-block ownership, lifecycle state, and key-to-block mappings. This lets operators associate latency, endurance consumption, and write amplification with specific workload classes instead of treating the SSD as an opaque shared resource.

ScaleFlux positions the platform as a complement to NVIDIA’s recently announced CMX Context Memory Storage Platform, which provides a shared, pod-level context tier for high-speed KV cache access and reuse. The pitch is aimed at AI factory operators: CMX handles the context tier, while ScaleFlux addresses the endurance, data placement, and write amplification challenges specific to the underlying SSDs.

“As AI inference systems extend KV cache beyond GPU Memory and DRAM, understanding the behavior and requirements of the SSD tier becomes increasingly important,” said Jason Hardy, vice president of storage technology at NVIDIA. “Our engagement with ScaleFlux is helping characterize how KV cache offload affects storage requirements for latency, endurance, and write amplification, contributing to the broader storage ecosystem around NVIDIA CMX.”

The company is developing a trace-driven simulator that models KV-cache movement across GPU HBM, host memory, and SSD tiers. The simulator generates replayable SSD traces to evaluate placement, eviction, and lifecycle-grouping policies under controlled conditions.

ScaleFlux plans to showcase the platform at FMS, covering Context-Insight workload analysis, KV metadata correlation, lifecycle-aware FDP placement, write amplification reduction, and high-endurance operation for write-intensive KV cache workloads. The company has a substantial presence at the show: ScaleFlux said in July that its experts would lead seven presentations there, including a keynote co-delivered with NVIDIA on memory solutions for scaling the AI data pipeline.

The platform announcement caps a busy stretch of silicon news. Two days earlier, ScaleFlux unveiled two PCIe Gen6 parts it will introduce at the same show: the FC6116 NVMe SSD controller and the MC600 CXL 3.2 Type 3 memory controller. ScaleFlux rates the FC6116 at up to 28 GB/s sequential read and 25 GB/s sequential write, up to 7 million 4K random read IOPS and more than 1 million sustained 4K random write IOPS, under 9W active controller power, with support for TLC, QLC, and SLC NAND up to 256TB across E1.S/L, E3.S/L, and U.2/3. The MC600 draws under 9W typical in a Gen6 x8 configuration and handles quad-channel DDR5 or dual-channel DDR4 with up to 2TB of DDR5, a dual-generation capability ScaleFlux positions as a way to carry existing DDR4 into a CXL deployment. Both begin sampling with key customers in Q4 2026.

The post ScaleFlux KV Cache SSD Platform Claims 7-10+ DWPD and 200+ FDP Streams appeared first on StorageReview.com.

KIOXIA has introduced the CM10 Series, the company’s first PCIe 6.0 enterprise SSD and its first enterprise drive built on BiCS FLASH generation 10 TLC memory. The drives cover E3.S, E1.S, and 2.5-inch form factors, with capacities ranging from 1.60TB to 61.44TB and endurance options for both read-intensive and mixed-use workloads.

The PCIe 6.0 portion of the CM10 family includes E3.S drives with a 7.5mm thickness and E1.S models in 9.5mm and 15mm formats. These EDSFF versions use 332-layer BiCS FLASH generation 10 TLC memory, while the 15mm 2.5-inch model uses a PCIe 5.0 interface and BiCS FLASH generation 8 TLC memory.

All three form factors support NVMe 2.1 and the Open Compute Project Datacenter NVMe SSD Specification 2.7, although not all requirements in the OCP specification are supported. NVMe Flexible Data Placement is also included, allowing compatible hosts to organize data placement to reduce internal write amplification and improve flash utilization.

KIOXIA CM10 Targets AI Inference and KV Cache Workloads

The CM10 Series is aimed at enterprise and AI workloads, including AI inference and key-value cache. KIOXIA says the drives are designed to support the NVIDIA CMX context memory storage solution, which extends effective GPU memory with a shared, pod-level context tier optimized for ephemeral KV cache.

KIOXIA frames the demand in specific terms: as AI models scale to trillions of parameters and agentic AI context windows expand to millions of tokens, the need for high-performance KV cache storage grows accordingly. Moving part of that cache from GPU memory to high-speed flash gives infrastructure designers another tier for retaining context data without holding all of it in accelerator memory.

Compared with the previous CM9 Series, the CM10 is rated for up to approximately 92% higher sequential read performance and up to approximately 85% higher random read performance. As always, actual real-world performance will depend on the host platform, software, drivers, operating system, and workload conditions.

“As AI infrastructure evolves, storage must deliver not only higher performance but also greater efficiency and deployment flexibility,” said Neville Ichhaporia, senior vice president and general manager of the SSD business unit at KIOXIA America. “The KIOXIA CM10 Series reflects our continued commitment to advancing flash storage that helps customers build more capable, scalable, and efficient AI infrastructure.”

KIOXIA CM10 Series Specifications

Specification	EDSFF Models	2.5-inch Model
Interface and Form Factor
Interface	PCIe 6.0; NVMe 2.1 specification compliant	PCIe 5.0; NVMe 2.1 specification compliant
OCP support	OCP Datacenter NVMe SSD Specification v2.7 support; not all requirements are supported	OCP Datacenter NVMe SSD Specification v2.7 support; not all requirements are supported
NVMe Flexible Data Placement	Supported	Supported
Form factor	E3.S, 7.5 mm thickness E1.S, 9.5 mm or 15 mm thickness	2.5-inch, 15 mm thickness
Capacity and endurance	Capacities from 1.60 TB to 61.44 TB across the CM10 Series, varying by form factor; read-intensive (1 DWPD) and mixed-use (3 DWPD) endurance options
Architecture and Data Protection
Architecture	KIOXIA controller and firmware with BiCS FLASH generation 10 TLC memory	KIOXIA controller and firmware with BiCS FLASH generation 8 TLC memory
High availability	Dual-port design on E3.S models only	Dual-port design
Data protection	Power loss protection and end-to-end data protection	Power loss protection and end-to-end data protection
Security options	SIE, SED, planned SED FIPS 140-3, post-quantum cryptography, CNSA 2.0, and SPDM 1.4 attestation support
Power and Cooling
Power states	Power states above 25W supported	Not specified
Cooling	Traditional air cooling supported on all EDSFF models Direct cold-plate liquid cooling supported on E3.S and 9.5 mm E1.S models	Traditional air cooling supported
Power management	NVMe TP4199 Self-Reported Drive Power Support	NVMe TP4199 Self-Reported Drive Power Support

Direct Liquid Cooling for E3.S and E1.S

Direct cold-plate liquid cooling is supported by the E3.S and 9.5mm E1.S models, while every CM10 form factor can operate in a traditional air-cooled environment. The capability is not unique to this family: KIOXIA introduced its first liquid-cooled SSD a day before the CM10 with the E1.S NX1 Series, a PCIe 5.0 data center drive built on generation 8 memory.

The PCIe 6.0 EDSFF drives support power states above 25W, giving the controller and flash a larger power envelope for higher-performance workloads. NVMe TP4199 Self-Reported Drive Power support allows the SSD to communicate its power use to compatible systems.

Dual-port connectivity is included on the E3.S and 2.5-inch versions for high-availability deployments. Power loss protection and end-to-end data protection are available across the listed form factors. Moreover, security options include Secure Instant Erase, Self-Encrypting Drive models, a planned FIPS 140-3 SED option, post-quantum cryptography, CNSA 2.0 support, and SPDM 1.4 attestation. Optional SED models support TCG Opal SSC, with some feature exceptions, and may not be available in every country due to local regulations.

KIOXIA CM10 Series Availability

KIOXIA CM10 Series drives are currently being sampled for select customers for functional testing, and the sample specifications may differ from the eventual production versions. KIOXIA will display the drives at FMS: the Future of Memory and Storage in Santa Clara, California, from August 4 to 6.

The post KIOXIA CM10 Is Its First PCIe 6.0 Enterprise SSD, Aimed at NVIDIA CMX KV Cache appeared first on StorageReview.com.

KIOXIA has introduced the NX1 Series, a new family of PCIe 5.0 NVMe data center SSDs built around the E1.S form factor. Capacities range from 1.92 TB to 15.36 TB, with read-intensive endurance rated at 1 DWPD, while cooling options include both air and direct liquid cooling. The drives succeed the XD Series and are designed for GPU-enabled AI servers and hyperscale infrastructure where dense storage configurations require more attention to heat removal.

KIOXIA NX1 Series Specifications

Specification	Details
Form factor and cooling	Available in E1.S 9.5 mm (supports direct liquid cooling and air cooling) and E1.S 15 mm form factor (supports air cooling)
Capacity and endurance	Capacities ranging from 1.92 TB to 15.36 TB with read-intensive endurance (1 DWPD)
Standards and features	Compliant with PCIe 5.0, NVMe 2.0, and Open Compute Project (OCP) Datacenter NVMe SSD 2.6 specifications, as well as NVMe Flexible Data Placement (FDP) support
Flash memory	Built with KIOXIA BiCS FLASH generation 8 TLC memory technology utilizing CMOS directly Bonded to Array (CBA) architecture
Security	Optional TCG Opal-compliant Self-Encrypting Drive (SED) security model

 

Direct liquid cooling is available on the 9.5 mm E1.S model, which can also use air cooling, while the thicker 15 mm version supports air cooling only. The NX1 is KIOXIA’s first SSD with direct liquid cooling support and can be used in cold-plate-compatible configurations, allowing heat to be removed from the drive in dense, accelerator-rich server designs.

KIOXIA built the NX1 around a new in-house controller architecture and BiCS FLASH generation 8 TLC memory using its CMOS directly Bonded to Array (CBA) technology. Platform support includes PCIe 5.0, NVMe 2.0, the Open Compute Project Datacenter NVMe SSD 2.6 specification, and NVMe Flexible Data Placement.

Compared with the previous-generation KIOXIA XD8 Series, the NX1 delivers up to 38% higher sequential write performance and up to 20% higher random write performance. As always, those figures can vary depending on the host system, drivers, operating system, software, and read/write conditions.

Security is available through an optional TCG Opal-compliant Self-Encrypting Drive model. This version does not support every TCG Opal SSC feature and will not be offered in every country due to local regulations.

The KIOXIA NX1 Series is currently being sampled by select hyperscale customers. KIOXIA will also exhibit the drives at FMS: the Future of Memory and Storage in Santa Clara, California, from August 4 through August 6.

The post KIOXIA’s First Liquid-Cooled SSD Arrives in the E1.S NX1 Series appeared first on StorageReview.com.

Samsung’s new 990 SSD joins the consumer lineup as a mainstream PCIe Gen4 drive that prioritizes efficiency and value over raw speed. Samsung calls it its most power-efficient SSD to date, claiming up to 38% better power efficiency than the 990 PRO. The 2TB model is rated at 7,250MB/s sequential read and 6,450MB/s write, with random I/O up to 850K IOPS read and 1,200K IOPS write. The 1TB model steps down slightly to 7,150MB/s sequential read, 700K IOPS random read, and 1,100K IOPS random write. The drive ships in 1TB and 2TB capacities at MSRPs of $269.99 and $529.99, respectively.

The name deserves a modest explanation. Samsung’s 990 family already includes the 990 PRO, 990 EVO, and 990 EVO Plus. The plain “990” now sits alongside them as an entry SSD rather than starting a new generation. It pairs a Samsung in-house controller with V-NAND (more on that later) in a DRAM-less design that relies on a Host Memory Buffer. It carries a three-year warranty, compared to the PRO’s five-year warranty, and is rated for 400TB and 800TB of writes at 1TB and 2TB capacities. Sequential read matches the 990 EVO Plus at 7,250MB/s but trails the 990 PRO’s 7,450MB/s. The 850K IOPS random read rating is well under the PRO’s 1,400K. Samsung’s pitch centers on two numbers: sequential writes are over 50% faster than the 990 EVO, and a 38% efficiency gain; 1,686MB/s per watt on 2TB reads versus 1,221MB/s per watt for the 990 PRO in Samsung’s internal testing. Buyers who prioritize outright speed already have Samsung’s Gen5 9100 PRO family, which we reviewed at launch and again in its 8TB capacity, so the 990 is not meant to compete directly with those drives.

Samsung has been down this road before. The 980, its first DRAM-less consumer NVMe drive, came through our lab in 2021 and left a poor impression. The smaller capacities, in particular, landed at the bottom of our charts, and the drive was relentlessly mocked around the lab. But the market has shifted dramatically since then. The AI buildout is soaking up NAND and DRAM supply, component pricing is climbing, and consumers are feeling the pinch across the board. In that environment, a Gen4 drive that offers solid sequential speed, low power draw, and decent capacity starts to look like a sensible choice, provided the street price cooperates.

One spec Samsung refused to provide is the NAND itself. The reviewer’s guide lists only “Samsung V-NAND.” When we asked directly, the company said it cannot disclose component details beyond official specifications, pointing us back to rated performance and its “latest Samsung V-NAND technology.” So we are left to read between the lines, which point to TLC rather than QLC. The endurance spec is the giveaway: 400TB and 800TB over a three-year warranty works out to roughly 133TB and 267TB of writes per year, essentially the same annual allowance as the 990 PRO’s 600TB and 1,200TB across five years. The likeliest answer is a lower-bin Samsung TLC V-NAND in this drive or a lower-cost, shorter-warranty option to make the drive more affordable. It’s odd that they’re being intentionally coy on what’s typically a foundational specification.

Our review unit is the 2TB model (MZ-V9V2T0), a pre-production sample running firmware 0B2QLXL7, which we put through fio, GDS, and AI model-load testing detailed below.

Samsung 990 SSD Specifications

Specification	Samsung 990 1TB	Samsung 990 2TB
Platform Overview
Interface	PCIe 4.0 x4, NVMe 2.0 (backward compatible with PCIe 3.0)
Form Factor	M.2 2280 Max 80.15 x 22.15 x 2.38 (mm)
Controller	Samsung in-house controller
NAND	Samsung V-NAND
Cache Memory	HMB (Host Memory Buffer), DRAM-less
Model Code	MZ-V9V1T0	MZ-V9V2T0
Performance
Sequential Read	Up to 7,150MB/s	Up to 7,250MB/s
Sequential Write	Up to 6,450MB/s	Up to 6,450MB/s
Random Read	Up to 700K IOPS	Up to 850K IOPS
Random Write	Up to 1,100K IOPS	Up to 1,200K IOPS
Power and Endurance
Active Power (Avg. Read)	4.0W	4.3W
Active Power (Avg. Write)	3.7W	3.8W
Idle Power (Typical)	55mW PS3 (APST on) 3mW PS4 (L1.2)
Endurance (TBW)	400TB	800TB
MTBF	1.5 million hours
Warranty	3 years limited
Features
Supporting Features	TRIM (OS support required) Garbage Collection S.M.A.R.T.
Data Security	AES 256-bit Full Disk Encryption TCG/Opal V2.0 Encrypted Drive (IEEE1667)
Software	Samsung Magician 9.0
MSRP	$269.99	$529.99

Samsung 990 SSD Design and Build

The Samsung 990 uses the familiar M.2 2280 form factor, measuring up to 80.15 x 22.15 x 2.38mm. It has a single-sided design and launches without a dedicated heatsink option. This makes it a good physical fit for notebooks, compact PCs, and desktop motherboards with their own M.2 cooling. The overall construction is simple, with Samsung keeping the controller, NAND, and supporting circuitry on one side of the PCB.

The front label displays the Samsung 990 branding, 2TB capacity, model number, firmware, and electrical specifications. There is no integrated heat spreader, so cooling relies on system airflow and an M.2 heatsink you provide.

With the label removed, you can see Samsung’s in-house controller sitting close to the M.2 connector, with the power components packed around it. The NAND sits at the other end of the board, leaving quite a bit of unused space in the middle. Since this is a DRAM-less drive, there is no separate DRAM chip on the PCB.

The back of the drive is mostly taken up by the regulatory label, with no active components underneath. Again, because the 990 uses a single-sided layout, it should be easier to fit in thin laptops and compact systems where space around the M.2 slot can be tight.

On the software side, the 990 is managed through Samsung Magician 9.0, which covers the essentials: firmware updates, drive health and S.M.A.R.T. monitoring, diagnostic scans, benchmarking, and secure erase, along with setup for the drive’s AES 256-bit encryption features. There’s nothing 990-specific to configure, since the HMB arrangement requires no user tuning. It’s worth installing at first boot to keep the firmware up to date, and Magician is generally a very capable tool that adds value.

Samsung 990 Performance

Peak Synthetic Performance

The FIO test is a flexible and powerful benchmarking tool for measuring the performance of storage devices, including SSDs and HDDs. It evaluates metrics such as bandwidth, IOPS, and latency under different workloads, like sequential and random read/write operations. This test helps to assess the peak performance of storage systems, making it useful for comparing different devices or configurations. We measured the peak burst performance for this test, limiting the workload to a 10GB footprint on both SSDs.

Peak Synthetic Performance: The Samsung 990 delivered 7,177 MB/s sequential read, 6,070 MB/s sequential write, 872K random read IOPS, and 1.08M random write IOPS, placing it near the bottom of this PCIe Gen4/Gen5 comparison group. Compared to the Samsung 990 Pro, the 990 trailed by about 4% in sequential read, but fell 15.7% behind in sequential write, 37.7% behind in random read IOPS, and 23.0% behind in random write IOPS. Against the fastest Gen5 drive, the SanDisk SN8100, the gap widened considerably, with the 990 delivering roughly 52% lower sequential read throughput, 57% lower sequential write throughput, 62% lower random read performance, and 50% lower random write performance.

FIO Test (higher MB/s/IOPS is better)	Sequential 128K Read (1T/64Q)	Sequential 128K Write (1T/64Q)	Random 4K Read (16T/32Q)	Random 4K Write (16T/32Q)
SanDisk SN8100	15,000MB/s (0.56ms avg latency)	14,100MB/s (0.59ms avg latency)	2.312M IOPS (0.22ms avg latency)	2.144M IOPS (0.24ms avg latency)
Kingston FURY Renegade G5	14,600MB/s (0.57ms avg latency)	14,100MB/s (0.59ms avg latency)	2.028M IOPS (0.25ms avg latency)	2.028M IOPS (0.25ms avg latency)
Samsung 9100 Pro	14,600MB/s (0.57ms avg latency)	13,300MB/s (0.63ms avg latency)	2.734M IOPS (0.18ms avg latency)	2.734M IOPS (0.19ms avg latency)
SK hynix Platinum P51	14,500MB/s (0.58ms avg latency)	13,500 MB/s (0.62ms avg latency)	2.369M IOPS (0.22ms avg latency)	2.669M IOPS (0.19ms avg latency)
Crucial T705	14,400MB/s (0.58ms avg latency)	12,300MB/s (0.68ms avg latency)	1.585M IOPS (0.32ms avg latency)	2.703M IOPS (0.19ms avg latency)
TEAMGROUP GE Pro 2TB	13,900MB/s (0.60ms avg latency)	12,800MB/s (0.65ms avg latency)	2.585M IOPS (0.23ms avg latency)	1.818M IOPS (0.28ms avg latency)
Lexar Professional NM1090 PRO	13,800MB/s (0.61ms avg latency)	13,600MB/s (0.62ms avg latency)	2.251M IOPS (0.23ms avg latency)	1.818M IOPS (0.28ms avg latency)
TEAMGROUP GC Pro 2TB	13,600MB/s (0.62ms avg latency)	12,700MB/s (0.66ms avg latency)	2.110M IOPS (0.24ms avg latency)	1.686M IOPS (0.28ms avg latency)
PNY CS2150	10,400MB/s (0.80ms avg latency)	8,801MB/s (0.95ms avg latency)	1.379M IOPS (0.371ms avg latency)	1.623M IOPS (0.32ms avg latency)
Corsair MP700 MICRO 4TB	9,169MB/s (0.91ms avg latency)	7,948MB/s (1.06ms avg latency)	1.277M IOPS (0.40ms avg latency)	1.540M IOPS (0.33ms avg latency)
Crucial P510	8,835MB/s (0.90 ms avg latency)	9,961MB/s (0.80 ms avg latency)	1.163M IOPS (0.44ms avg latency)	1.196M IOPS (0.51ms avg latency)
Micron 3610 2TB	6,839MB/s (1.23ms avg latency)	9,673MB/s (0.87ms avg latency)	1.523M IOPS (0.34ms avg latency)	1.871M IOPS (0.27ms avg latency)
Samsung 990 Pro	7,483MB/s (1.12ms avg latency)	7,197MB/s (1.16ms avg latency)	1.400M IOPS (0.36ms avg latency)	1.403M IOPS (0.36ms avg latency)
Crucial P310 2TB	7,197MB/s (1.16ms avg latency)	6,376MB/s (1.31ms avg latency)	1.163M IOPS (0.44ms avg latency)	1.196M IOPS (0.43ms avg latency)
Samsung 990 2TB	7,177MB/s (1.17ms avg latency)	6,070MB/s (1.38ms avg latency)	872K IOPS (0.59ms avg latency)	1.08M IOPS (0.47ms avg latency)
WD SN850X 2TB	6,632MB/s (0.76ms avg latency)	7,235MB/s (0.92ms avg latency)	1.2M IOPS (0.43ms avg latency)	825K IOPS (0.62ms avg latency)
Micron 2600 2TB	5,702MB/s (1.47ms avg latency)	6,612MB/s (1.27ms avg latency)	1.11M IOPS (0.46ms avg latency)	1.36M IOPS (0.38ms avg latency)

Average LLM Load Time

The Average LLM Load Time test evaluated the load times of three different LLMs: DeepSeek R1 7B, Meta Llama 3.2 11B, and DeepSeek R1 32B. Each model was tested 10 times, and the average load time was calculated. This test measures the drive’s ability to load large language models (LLMs) into memory quickly. LLM load times are critical for AI-related tasks, especially for real-time inference and processing large datasets. Faster loading enables the model to process data more quickly, thereby improving AI responsiveness and reducing wait times.

Average LLM Load Time: AI model loading was the Samsung 990’s weakest test, with the drive finishing at or near the bottom across all three workloads. It recorded 5.06 seconds for DeepSeek R1 7B, 7.61 seconds for Meta Llama 3.2 11B Vision, and 7.86 seconds for DeepSeek R1 32B. Compared to the fastest drive, the SK hynix Platinum P51, the Samsung 990 took approximately 99% longer to load the 7B model, 112% longer to load the 11B Vision model, and 88% longer to load the 32B model. The more interesting result is the Samsung 990 Pro, which lands at the bottom of these charts right alongside its value sibling: the 990 edged out the Pro by about 1% on the 7B load and trailed it by 15% on the 11B Vision model and 8% on the 32B model. Whatever Samsung’s Gen4 drives give up in this workload, they give it up together, so stepping up to the Pro buys little for AI model loading.

Average LLM Load Time (lower is better)	DeepSeek R1 7B	Meta Llama 3.2 11B Vision	DeepSeek R1 32B
SK hynix Platinum P51	2.5481s	3.5809s	4.1790s
SanDisk SN8100	2.5702s	3.5856s	4.2870s
Samsung 9100 Pro 4TB	2.6173s	3.6017s	4.3735s
PNY CS2150	2.8107s	3.6820s	4.8962s
Crucial T705 2TB	2.8758s	3.6312s	5.1080s
Crucial P510 1TB	2.8817s	3.6631s	5.0594s
TEAMGROUP GE Pro 2TB	2.9092s	3.9136s	4.8974s
TEAMGROUP GC Pro 2TB	2.9379s	3.9267s	4.8188s
WD SN850X 2TB	3.0082s	3.6543s	5.4844s
Kingston FURY Renegade G5	3.1843s	4.8009s	4.6523s
Crucial P310 2TB	3.1889s	3.7083s	5.4844s
Lexar Professional NM1090 PRO	3.2135s	4.9504s	7.2108s
Micron 2600 2TB	3.3178s	3.9174s	5.9060s
Corsair MP700 MICRO 4TB	3.4694s	5.2106s	5.3990s
Micron 3610 2TB	3.5348s	5.3853s	5.5731s
Samsung 990 2TB	5.0645s	7.6087s	7.8619s
Samsung 990 Pro 2TB	5.1255s	6.6051s	7.3021s

 

One of the tests conducted on this testbench was the Magnum IO GPU Direct Storage (GDS) test. GDS is a feature developed by NVIDIA that allows GPUs to bypass the CPU when accessing data stored on NVMe drives or other high-speed storage devices. Instead of routing data through the CPU and system memory, GDS enables direct communication between the GPU and the storage device, significantly reducing latency and improving data throughput.

How GPU Direct Storage Works

Traditionally, when a GPU processes data stored on an NVMe drive, the data must first travel through the CPU and system memory before reaching the GPU. This process introduces bottlenecks because the CPU acts as a middleman, adding latency and consuming valuable system resources. GPU Direct Storage eliminates this inefficiency by enabling the GPU to access data directly from the storage device via the PCIe bus. This direct path reduces data-movement overhead, enabling faster, more efficient data transfers.

AI workloads, especially those involving deep learning, are highly data-intensive. Training large neural networks requires processing terabytes of data, and any delay in data transfer can lead to underutilized GPUs and longer training times. GPU Direct Storage addresses this challenge by ensuring that data is delivered to the GPU as quickly as possible, minimizing idle time and maximizing computational efficiency.

In addition, GDS is particularly beneficial for workloads that involve streaming large datasets, such as video processing, natural language processing, or real-time inference. By reducing the reliance on the CPU, GDS accelerates data movement and frees up CPU resources for other tasks, further enhancing overall system performance.

Throughput on the read side climbed steadily as thread count increased. At the 1M block size, the Samsung 990 started at 2.27 GiB/s on a single thread and peaked at 2.89 GiB/s with 64 threads, then settled to 2.79 GiB/s at 128 threads. The 128K block size followed a similar curve, increasing from 1.21 GiB/s at 1 thread to 2.12 GiB/s at 128 threads, roughly a 75% gain. The 16K block size behaved differently. It peaked at a single thread (0.82 GiB/s), dropped sharply once concurrency was introduced, and plateaued around 0.3 GiB/s from 8 threads onward. This pattern is consistent with small block I/O saturating on per-operation overhead rather than raw bandwidth.

Latency scaled as expected, rising with thread count. At 1M, average latency grew from 430 microseconds on one thread to 44.7 milliseconds on 128 threads, roughly a 100x increase, reflecting increased queue depth due to more concurrent GPUDirect Storage threads. The 128K and 16K block sizes showed the same upward trend, reaching 7.4 milliseconds and 6.2 milliseconds, respectively, at 128 threads. Notably, 16K had the lowest single-thread latency of the three (18 microseconds), reflecting its smaller per-operation payload, even though its overall throughput ceiling was the lowest.

Write throughput varied by block size. The 1M block size stood out, jumping from 0.32 GiB/s at a single thread to 3.89 GiB/s at 8 threads, its peak, before tapering slightly to 3.63 GiB/s at 128 threads as queueing overhead increased. The 128K and 16K block sizes remained flat across the thread range, hovering near 0.3 GiB/s regardless of concurrency. This suggests the write path is limited by per-I/O overhead or controller queuing rather than bandwidth at those sizes.

Latency on writes rose more steeply than on reads, particularly at 128K, which climbed from 382 microseconds at one thread to 51.0 milliseconds at 128 threads, the highest figure recorded across either write or read testing. The 1M block size showed an unusual dip, with latency dropping from 3.06 milliseconds on one thread to 1.03 milliseconds on four threads, likely because the single-thread run was not yet saturating the write path, then climbing steadily to 34.4 milliseconds on 128 threads. The 16K block size stayed the most consistent, closing the sweep at 9.1 milliseconds, the lowest ceiling of the three block sizes.

Conclusion

The Samsung 990 isn’t chasing the top of the charts, and the numbers make that clear. It trailed the 990 Pro across every FIO test we ran, from a 4% gap in sequential read to a 38% deficit in random read IOPS, and fell well behind the fastest Gen4 and Gen5 drives in this comparison group. AI model loading tells a similar story, with the 990 and 990 Pro finishing at the bottom of the field together; the 990 actually edged its Pro sibling on the DeepSeek R1 7B load while trailing by 8 to 15% on the larger models. Buyers who need peak throughput for demanding workloads should look elsewhere in Samsung’s lineup, starting with the Gen5 9100 Pro.

That said, judging the 990 against the fastest drives on the market misses the point of the drive. This is a mainstream Gen4 SSD built around efficiency and value, not benchmark supremacy, and compared to its real predecessor, the 990 EVO, it’s a solid upgrade. Sequential writes are more than 50% faster, power efficiency is up 38% by Samsung’s own numbers, and the drive still delivers sequential read speeds in line with the 990 EVO Plus. Samsung chose to stay quiet on the NAND itself, but the endurance ratings suggest a competent TLC implementation rather than a QLC design.

For buyers who don’t need Gen5 speeds or Pro-tier random I/O, and who want a dependable, power-efficient Gen4 drive at a workable price, the 990 is a sensible upgrade path from the EVO line. Notably, stepping up to the 990 Pro buys nothing for AI model loading, so the choice between the two comes down to random I/O and sustained writes rather than anything AI-related. At $269.99 for 1TB and $529.99 for 2TB, the MSRPs reflect the current memory market more than the drive’s entry positioning, so the 990’s value case will ultimately be set by street prices. It’s not the drive to buy if raw performance is the priority, but it does what it’s meant to do.

Product Page – Samsung 990

The post Samsung 990 SSD Review: A Value Gen4 SSD for Expensive Times appeared first on StorageReview.com.

Samsung has started mass production of the PM1763, its first PCIe Gen6 enterprise SSD, pairing 9th-generation V-NAND with a newly developed 4nm controller. The 15.36TB flagship is rated for sequential reads up to 28,400 MB/s and writes up to 21,000 MB/s, with reads landing at 1.96x the 14,500 MB/s ceiling of the Gen5 PM1753 it replaces. Samsung says the PM1763 has completed validation for next-generation AI platforms, and the timing aligns with the first wave of Gen6-capable servers expected in the coming year.

Why PCIe Gen6 Matters

PCIe Gen6 doubles the per-lane signaling rate to 64 GT/s using PAM4, which puts roughly 32 GB/s of bandwidth each direction on the standard x4 SSD link, up from about 16 GB/s on Gen5. Gen5 drives have been bumping against that ceiling for a while, with the fastest models rated in the 14-14.5 GB/s range the interface allows.

The PM1763’s 28.4 GB/s rating uses most of the new headroom, and for AI infrastructure the practical effect is that fewer drives are needed to saturate a GPU server’s storage path, while checkpoint and model-load operations spend less time blocking accelerators. Samsung frames it in model terms, claiming a 40GB LLM can move from drive to memory in roughly 1.4 seconds, versus about 2.7 seconds on the PM1753. That 1.4-second figure is the sequential read rating for a 40GB file, not a measured transfer, so treat it as a best-case illustration of the interface math.

Random performance scales alongside the sequential numbers. Samsung’s product materials list up to 6.8 million random read IOPS and 950,000 random write IOPS for the 16TB-class configuration; the SCADA white paper’s measured 6.92 million IOPS per drive (covered below) suggests the read figure is conservative.

Power efficiency improves up to 1.8x over the prior generation, which matters as much as the raw speed in dense deployments, since Gen6 controllers run hot and every watt saved per drive multiplies across a dense chassis. To that point, Samsung has optimized the PM1763 for liquid-cooled servers with direct-to-chip (D2C) cooling, targeting sustained peak performance under extended load rather than burst figures.

PM1763 Specifications

Specification	Samsung PM1763
Platform Overview
Interface	PCIe Gen6 x4, NVMe 2.1, OCP 2.6
NAND	Samsung 9th-generation V-NAND
Controller	New Samsung 4nm controller
Capacities	4TB, 8TB, 16TB-class (15.36TB formatted) at launch Product page lists 30.72TB and 61.44TB
Form Factors	E1.S E3.S U.2 (PCIe Gen5 only)
Performance (16TB)
Sequential Read	Up to 28,400 MB/s
Sequential Write	Up to 21,000 MB/s
Random Read	Up to 6,800,000 IOPS
Random Write	Up to 950,000 IOPS
Power and Cooling
Power Efficiency	Up to 1.8x improvement over PM1753
Cooling	Optimized for liquid-cooled servers, direct-to-chip (D2C)
Security
Features	Post-quantum cryptography (PQC) TEE Device Interface Security Protocol (TDISP)

GPU-Driven I/O: 281 Million IOPS Across 42 Drives

Samsung has also published a white paper pairing the PM1763 with SCADA (Scaled Accelerated Data Access), the NVIDIA-developed framework that lets GPU threads submit NVMe commands directly to the drives, bypassing the CPU and kernel storage stack entirely. The argument for GPU-initiated I/O is concurrency: a CPU can keep roughly 45 million IOPS in flight across its thread pool, while a GPU dispatching from around 100,000 threads pushed past 95 million IOPS per GPU in Samsung’s testing.

The numbers come from a 512-byte random read workload on an H3 Falcon 6048 Gen6 server with one H100, two H200s, and 42 PM1763 E1.S 15.36TB drives behind three Broadcom PEX90144 Gen6 switches. A single PM1763 processed roughly 6.92 million GPU-issued IOPS, an 86% gain over the 3.72 million the Gen5 PM1753 managed in the same setup. With 14 drives per GPU, the system held near-linear scaling at about 96 million IOPS per GPU group, and the full 42-drive configuration aggregated 281 million IOPS, with per-drive results staying within 5% of the single-drive peak. We did not conduct this testing, nor did we independently audit the results, but the scaling behavior is interesting: latency consistency across drives, not peak IOPS, determined how well the aggregate held up.

What Gen6 Does for Dense Storage Servers

The question is what a shelf of these drives does inside a single box. We recently pushed the Dell PowerEdge R7725xd past 300 GB/s of local throughput with 24 Gen5 SSDs, each drive on dedicated x4 lanes from the CPU complex, and served 160 GB/s over the network with PEAK:AIO’s software keeping the queues saturated. That system rivals multi-node storage clusters from a single 2U chassis. Swap in Gen6 drives at the PM1763’s rated speeds and the same 24-bay topology carries a theoretical ceiling of roughly 681 GB/s of raw read bandwidth, though CPU lane budgets and network egress become the binding constraints well before the drives do.

Capacity density moves in tandem. PEAK:AIO’s 2U AI Data Server already packs 1.5PB using 61.44TB QLC drives while delivering 120 GB/s over RDMA. When Samsung ships the 61.44TB PM1763, that class of system gets Gen6 bandwidth and petabyte-plus density in a similar footprint (next-gen servers are likely to add a rack unit for cooling). For AI shops trying to keep GPU clusters fed without building out a parallel file system across a dozen nodes, the single-server storage argument keeps getting stronger.

Security and Availability

Samsung has also extended the drive’s security stack for multi-tenant AI environments. The PM1763 supports post-quantum cryptography algorithms ahead of anticipated quantum attacks on classical encryption, as well as TDISP (TEE Device Interface Security Protocol), which secures the data path between confidential VMs and the device in virtualized deployments.

“Built on industry-leading performance, PM1763 has successfully completed validation for next-generation AI platforms and is well positioned to support evolving AI infrastructure requirements,” said Jangseok Choi, Vice President and Head of Memory Product Planning at Samsung Electronics. “As AI models continue to grow in size and complexity, PM1763 will serve as a key solution that enables customers to efficiently scale memory capacity and optimize AI operations.”

The PM1763 is in mass production now in 4TB, 8TB, and 16TB capacities. Samsung has not announced ship dates for the larger capacity points.

Samsung PM1763 Product Page

The post Samsung PM1763 PCIe Gen6 SSD Enters Mass Production With 28.4 GB/s Reads appeared first on StorageReview.com.