Memory & StorageLOT 261122 · 6 min read

Phison Puts an NPU and 64GB of Memory Inside Enterprise SSDs

Phison shows an SSD with an NPU, 64GB of 130 GB/s memory and support for PyTorch and llama.cpp, plus a pSLC-QLC mixed-media drive claiming 4% cache loss per drive failure.

By Tom Whitfield1,197 words

Key takeaways

  1. Phison detailed an SSD with an NPU and 64GB of 130 GB/s memory running PyTorch, vLLM, OpenVINO and llama.cpp at SC26
  2. Phison claims 10 to 100 times less data transferred for applicable workloads; neither design has a name or sample date
  3. The pSLC-QLC Mixed Media design claims 4% cache loss when one of 24 drives fails, versus 25% with four dedicated cache drives
  4. Phison's X3 PCIe 6.0 controller targets 28 GB/s and 6.8 million IOPS, with customer samples expected by year-end
  5. Kioxia, DapuStor, Solidigm and Enmotus have similar or competing tiering approaches; Samsung's SmartSSD left the market
Phison details SSD with a built-in NPU and 64GB of memory that supports AI frameworks on the drive
Phison details SSD with a built-in NPU and 64GB of memory that supports AI frameworks on the drive — geralt / pixabay

Phison has detailed two enterprise SSD designs at Supercomputer 2026 that carry a combined claim of up to 100 times less data transfer for AI workloads. One design embeds an NPU with 64GB of 130 GB/s memory to run PyTorch, vLLM, OpenVINO and llama.cpp directly on the drive. The other merges fast pSLC cache with large QLC capacity behind a single 16-channel controller. Neither design has a product name, a named controller or a sample date yet.

The company notes that "many of the products and features mentioned are still in development" and will arrive once finalized. The claims are therefore roadmap material, not shipping specifications.

How would on-drive AI compute work?

The premise is simple. A server offloads work to the SSD, which runs AI code where the data already sits and sends back results instead of whole datasets. Phison says developers can build on the drive "without specialized programming tools" and that the drives "scan and process data in parallel." The company claims this could mean 10 to 100 times less data transferred for applicable workloads, freeing the host CPU for other tasks.

The compute figure Phison gives is 36,800 DMIPS, a Dhrystone-based integer benchmark from the 1980s. NPUs are normally rated in TOPS, and Phison did not provide that number. The DMIPS figure likely refers to the CPU cores that work alongside the NPU — probably Arm Cortex-based — adequate for the drive but not a measure of AI capability. For reference, Phison's Topaz NPU add-in card, shown at Computex, carried eight NPUs rated at up to 40 TOPS each.

The memory matters more for this workload. Phison does not state the type, though one secondary report suggests LPDDR5X, possibly paired with the Topaz NPU. If the drive uses PCIe 6.0 like Phison's X3 controller, host transfers could reach 28 GB/s, which would put the onboard memory at more than four times that speed. For scale, 130 GB/s is a bit under half the bandwidth of Nvidia's DGX Spark hardware.

Nothing was disclosed about how a host reaches the drive's compute. The open question of host-to-NPU access, not the benchmark number, will decide whether this design is practical.

Is computational storage a proven idea?

No, and the record is mixed. Samsung tried it with SmartSSD, first shown in 2018 with an FPGA on the drive. A 2022 second generation used AMD's Xilinx Versal FPGA, with Samsung claiming more than 50% lower processing time, up to 70% lower energy consumption and up to 97% lower server CPU usage for scan-heavy database queries. The drive has reportedly since disappeared from Samsung's lineup, though the first-generation model was still sold under the Xilinx name as of last year.

ScaleFlux took a narrower path, handling compression inside the SSD controller so the drive looks like any other NVMe SSD. Its CEO claims it is the only company with computational storage in production at hyperscalers. That history suggests the secret to success is not making customers rewrite code — SmartSSD required new code for each workload, while Phison's pitch is flexibility with common AI frameworks.

Phison CTO Sebastien Jean, writing in All About Circuits in 2025, argued generic computational storage had stalled because "each use case is highly unique and it's simply not scalable." He called running Linux on a drive a "pie-in-the-sky approach" and proposed targeted accelerators running set operations over logical block addresses. By his reckoning, a chassis with 90 PCIe 6.0 SSDs could process 2.5 TB/s without touching the CPU. The SC26 design looks closer to the general-purpose model he once dismissed, likely because AI demand has changed the calculus.

What does the Mixed Media drive change?

The second design pairs "high-endurance pSLC and high-capacity QLC flash" in one 16-channel enterprise drive. The pSLC region is 4-bit QLC flash acting in 1-bit SLC mode, trading capacity for speed and endurance. Phison says it builds on standard NVMe interfaces and fits existing infrastructure "without requiring specialized host software."

Jean explained at the conference's Emerging Storage Technology panel that the SLC endurance group is managed separately from the TLC or QLC region, each with its own namespace, so the OS sees multiple drives. High-write data goes to the SLC group; QLC holds more read-heavy data. This matters because enterprises moving to QLC for capacity today use separate cache drives, which consume drive bays.

The tradeoff: one drive failure removes both capacity and a share of cache. Phison argues its approach is more robust, saying "losing one drive out of 24 would result in a 4% cache loss instead of a 25% loss" — the 25% assuming four dedicated cache drives. Host software simply picks the namespace to write to, with no tiering logic on the host or array side.

Who else is building this?

Phison is not first. Kioxia presented its Mixed Mode drive at FMS 2025, with a pSLC-to-QLC ratio set once per drive; Kioxia called 1:4 a "good beginning ratio," while customers typically want pSLC at just 0.5% to 2% of maximum QLC capacity, and its design adds parity protection across drives. DapuStor's dual-mode J5060 QLC SSD, offering 400GB, 800GB or 1.2TB of pSLC on a 30.72TB drive, was in customer validation as of August. Solidigm's Cloud Storage Acceleration Layer takes a different route with a host-side flash translation layer, and the Enmotus FuzeDrive P200 — previewed at CES 2020 with a Phison controller — is the closest ancestor, with 24GB of SLC on a 900GB consumer drive managed by Enmotus software.

Where does this leave the X3?

Both designs invite comparison with Phison's X3 enterprise controller, a PCIe 6.0 part promising up to 28 GB/s sequential throughput and 6.8 million IOPS, unchanged since Computex in June. Phison told press then to expect customer samples by year-end and volume in mid-2026, though the SC26 release gives no dates. Micron and Samsung already offer PCIe 6.0 SSDs, and Marvell's Bravera SC6 controller is expected.

Jean also discussed PCIe 7.0 drives at "50, 56 gigabytes per second" — "faster than a DDR channel" — where 32 drives could reach "1.8 terabytes per second" per chassis, calling it computational storage for HPC. Nvidia's roadmap targets 100 million IOPS for PCIe 7.0, but that counts 512-byte reads rather than 4KB. Jean said Phison is "very optimistic" it can achieve this, while noting the eightfold IOPS jump would mean "eight times as much power," roughly "in the 140 to 180 watt range" — beyond reasonable on current NAND.

Of the two new designs, Mixed Media looks nearer to product: it uses standard NVMe features, and DapuStor's similar drive already reached customer testing. The computational-storage design has one CPU benchmark and no framework-independent performance figures. CES and Computex on the 2027 calendar are where Phison will likely fill in product names, pSLC ratios and sample dates — and where the 100x transfer claim faces its first real test.

Tom's Hardware (Source)

Filed under

Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Staff writer covering companies, deals and strategy in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,200 articles

Read also

« Previous article