Phison Puts an NPU and 64GB of Memory Inside an SSD: PyTorch and llama.cpp Run on the Drive Itself, But There Is Still No Name, No Controller Model, and No Sampling Date
An AI server spends much of its day doing something remarkably inefficient: pulling data off storage, moving it into memory, handing it to a CPU or GPU, computing, then shipping the result back. The larger the dataset, the longer that trip takes, and the more often an expensive accelerator sits idle waiting to be fed. Phison's proposal at SC26 was blunt — stop moving the data, and let the drive do the computing.
On November 18–19, at Supercomputing 2026 (SC26) in Chicago, the Taiwanese SSD controller maker showed three storage technologies aimed at AI and HPC workloads. The one that drew the most attention was a computational storage design: an enterprise SSD with a neural processing unit (NPU) and 64GB of memory built in, capable of running PyTorch, vLLM, OpenVINO and llama.cpp directly on the drive and returning only the final result to the host.
Two caveats belong up front. First, Phison's own newsroom does not carry a separate, crawlable press release for this design, so the facts here are assembled from the Commercial Times on-site report, Tom's Hardware's independent verification, ElectronicsForU's write-up with official quotes, and follow-ups from TechSpot and TechPowerUp. This article is compiled from multiple media reports. Second, the design has no product name, no named controller, and no sampling date; Phison itself says many of the products and features shown at SC26 are still in development.
TL;DR: the short version
- What it is: an enterprise SSD design with an integrated NPU and 64GB of memory, able to run PyTorch, vLLM, OpenVINO and llama.cpp on the drive.
- Why it matters: AI inference is often limited not by compute but by the trip data makes from storage to the compute die. Pushing compute next to the data removes most of that trip.
- The official numbers: NPU rated at 36,800 DMIPS, 64GB of memory, 130GB/s of memory bandwidth; data transfers cut by 10x to 100x on suitable workloads.
- Compared with the host link: 130GB/s is more than four times the 28GB/s that Phison's next-generation PCIe 6.0 controller can deliver off the drive. The drive reads its own data far faster than it can hand it over.
- Can you buy it: no. No name, no controller model, no sampling date, and Phison has not even said what memory media it uses.
- The easiest thing to misread: 36,800 DMIPS is a decades-old CPU integer benchmark, not AI throughput. The entire announcement contains no TOPS figure at all.
Why put compute inside the drive
Start with the problem. When an AI server handles a request, the usual path is: read the needed data off SSD — possibly terabytes of logs, vectors, images or training samples — move it into host memory, then compute on a CPU or GPU. On that path, actual computation may account for only a fraction of the time; the rest is movement.
That movement costs three things: PCIe bandwidth, host CPU and memory cycles, and idle time on accelerators that cost far more per hour than the storage underneath them. Phison's idea is to perform the parts of the job that can be done where the data already sits — parallel scanning, filtering, preliminary inference — inside the SSD, and send back only what matters.
The official claim is that, for suitable workloads, data transfers drop by 10 to 100 times while host CPU load is freed up. The word "suitable" carries the weight here. Tasks that can be parallel-scanned and largely discarded on the drive benefit most; tasks that require repeatedly pulling an entire dataset into VRAM will see far less gain.
What the design actually contains
According to the official figures recorded by the Commercial Times and ElectronicsForU, the architecture has three key parts:
- A neural processing unit (NPU), rated by Phison at 36,800 DMIPS.
- On-drive memory: 64GB at 130GB/s.
- A software compatibility layer: support for PyTorch, vLLM, OpenVINO and llama.cpp, with Phison stating that developers do not need specialized programming tools to build on the architecture.
That third point is the real difference from earlier attempts at in-storage computing. The concept is not new, but historically it required customers to write their own low-level software and adapt their own IP, which kept it out of most deployments. Phison is pitching "your existing frameworks work as-is."
How fast 130GB/s really is
This number deserves its own section, because it is the most persuasive part of the whole proposal.
Phison's X3 (PS5303) enterprise SSD controller, announced in the same breath, uses a PCIe Gen6 x4 link with NVMe 2.3 and targets 28GB/s sequential reads and 6.8 million random IOPS — already at the top of what commercial SSDs deliver, and roughly double a PCIe Gen5 drive.
The 64GB of memory inside the drive runs at 130GB/s, more than four times that host link. In plain terms: data moves around inside the drive much faster than it can leave the drive. That gap is the precondition for the whole idea. If on-drive bandwidth were roughly equal to the outbound link, computing on the drive would offer no structural advantage over computing on the host.
TechSpot put the comparison most directly: the drive can read its own data far faster than it can transfer it to the server.
Two other technologies shared the stage
Computational storage was one of three announcements. The other two matter because they say more about how close any of this is to shipping.
The first is the PCIe Gen6 controller X3 (PS5303). PCIe Gen6 doubles per-lane throughput from 32 GT/s to 64 GT/s, at the cost of much harder signal integrity work. Phison's answer spans from the controller down to the physical layer: alongside the X3 controller there is the PS7161 redriver, which the company says is the first PCIe 6.0 redriver to reach the PCI-SIG Integrators List. This line has model numbers, published specs and a certification record — it is the part that is genuinely close to orderable.
The second is Mixed Media. A single controller with 16 channels puts high-endurance pSLC and high-capacity QLC NAND on the same drive, effectively merging the "fast cache drive" and the "big capacity drive." The reliability illustration Phison offers is easy to grasp: in a 24-drive system, losing one drive costs about 4% of cache capacity, versus roughly 25% under a dedicated-cache-drive architecture. It uses a standard NVMe interface and needs no extra host software.
Read together, the three sit at different maturity levels. X3 and Mixed Media have clear paths to product. Computational storage is an architecture proposal.
Can you order one: no
This is the most important section of this article.
Tom's Hardware, reporting on November 20, listed the gaps one by one: neither new design has a product name, a named controller, or a sampling date; Phison has not even said what type of memory media the drive uses. Phison's own framing is that many of the products and features shown at SC26 are still in development and will ship once finalized.
In other words, what is confirmed is that Phison built the architecture and can demonstrate it. What is not confirmed is when it can be bought, what it will cost, or how fast it runs. Any account describing this as a launched product goes beyond the available evidence.
How this differs from Samsung's SmartSSD
Samsung has walked this road before. It introduced the SmartSSD in 2018 with an FPGA on board, followed by a second generation in 2022. But TechRadar reported last year that the line had all but disappeared from Samsung's portfolio — partly because the generative AI boom created far more demand for raw storage capacity than for on-drive processing.
The other obstacle cited repeatedly for Samsung's approach was that customers had to develop their own software and IP, which made deployment expensive.
Phison's difference sits exactly there. It promises support for PyTorch, vLLM, OpenVINO and llama.cpp — tools developers already have. TechSpot's assessment is that if the NPU turns out to be as capable as claimed, that difference could carry in-storage computing further this time. The "if" matters: no third party has yet tested that NPU.
Where the numbers stop: 36,800 DMIPS is not AI throughput
This is where headlines are most likely to mislead.
The only compute figure Phison gave for the whole announcement is 36,800 DMIPS. DMIPS is Dhrystone MIPS, an integer benchmark designed in the 1980s to measure general-purpose CPU integer performance. It is not comparable to the TOPS or FLOPS used to rate AI accelerators today.
Tom's Hardware states the point plainly: the figure most likely describes the companion CPU cores on the drive rather than its AI processing capability, and there is no TOPS rating anywhere in the announcement.
So the correct reading is: Phison has shown that the drive can host PyTorch and llama.cpp. It has not given any comparable measure of how fast they run. Whether the design can carry real inference workloads will require third-party measurement — specifically token throughput, time to first token, and how large a model can be sustained at 130GB/s.
Timeline
- November 16–19, 2026: SC26 takes place in Chicago; Phison demonstrates the three storage technologies on site.
- November 19, 10:36 Taipei time: DIGITIMES publishes the first on-site report.
- November 19, 10:49 Taipei time: the Commercial Times publishes the fullest Chinese account, including quotes from the CEO.
- November 19, 08:28 UTC: ElectronicsForU publishes an English report quoting the president of Phison US.
- November 20: Tom's Hardware publishes its independent verification, itemizing what is missing.
- November 21–22: TechSpot, TechPowerUp, Guru3D, IT之家 and cnBeta follow up.
- Product timeline: not announced. Neither design has a name, a controller model, or a sampling date.
Who it is a good fit for
- Operators of data-heavy inference pipelines: log analysis, pre-filtering for vector search, and first-pass image screening — jobs that start by picking a small number of useful records out of an enormous pile — match the "10x to 100x less data movement" description most closely.
- Data centers already at their power and cooling ceiling: moving some compute into existing SSD bays adds no accelerators and no liquid cooling.
- Storage clusters trying to reclaim host CPU: Phison explicitly lists freed host CPU capacity as one of the benefits.
- Hardware teams evaluating storage architectures: this is a concrete reference point for the computational storage route and belongs in any comparison.
Not a good fit: who should skip it
- Production systems that need predictable performance: with no name, no sampling date and no third-party testing, capacity planning is impossible today.
- Large-model inference bounded by VRAM bandwidth: 130GB/s is enormous for an SSD, but HBM delivers multiple terabytes per second per card. This design addresses data sitting in storage, not model weights sitting in VRAM.
- Small, single-machine deployments: the value scales with data volume; a personal workstation or a small team will not see it.
- Anyone hoping to replace GPUs: Phison makes no such claim, and 36,800 DMIPS does not support that reading.
How to use this: where to start
First, measure how much of your time actually goes into moving data. Take a real inference or analytics job and time three phases separately: reading from storage, moving between host memory and VRAM, and actual computation. If the first phase dominates, this class of design is relevant. If the bottleneck sits in the second or third phase, on-drive compute will not help.
Second, define "suitable workload" concretely for your case. Phison's 10x to 100x is a range, and the tenfold spread between its ends shows how strongly the benefit depends on task shape. A practical proxy: simulate "filter first, then transfer" on the host and measure how much data volume can be eliminated. That is roughly the theoretical ceiling for on-drive compute.
Third, wait for product information. There is no model number, no sampling date and no price, so no procurement decision can be made yet. The sensible move is to put it on a watch list and track two signals: an official TOPS figure or measured token throughput, and the appearance of a named controller with a sampling plan.
Common mistakes and pitfalls
- Do not read 36,800 DMIPS as AI throughput. It is a CPU integer benchmark, not TOPS, and the announcement contains no TOPS figure.
- Do not read "supports PyTorch" as "can run large models." Running a framework and running a specific model at usable speed are different claims; Phison gave no model sizes and no tokens per second.
- Do not compare 130GB/s with HBM. That figure is on-drive memory bandwidth; its counterpart is the 28GB/s PCIe link, not the multiple terabytes per second of HBM.
- Do not read "10x to 100x less data transfer" as "10x to 100x faster." What shrinks is the volume moved, not the compute time. If compute is the bottleneck, the gain shrinks accordingly.
- Do not treat this as a launch. No product name, no controller model, no sampling date — Phison says the work is still in development.
- Do not ignore Samsung's precedent. SmartSSD was technically workable and the market did not take it up. Whether this attempt fares differently depends on the NPU's real capability and on developer adoption.
Cost and time
Phison has published no pricing and no sampling or production schedule, so any specific number would be an estimate rather than a fact.
Two reference points are useful. The X3 (PS5303) controller was announced with a model number and full specs, which suggests Phison is sequencing PCIe Gen6 silicon ahead of computational storage. And Samsung's SmartSSD ran from a 2018 first generation to a 2022 second generation before the line contracted — roughly four to five years — which indicates this class of product does not move quickly.
A reasonable expectation: a named controller and a sampling plan some time in 2027, with entry into enterprise procurement lists requiring another round of customer validation after that. All of it remains subject to what Phison announces.
Sources
- Commercial Times (November 19, 2026): https://www.ctee.com.tw/news/20261119700981-430201
- Tom's Hardware (November 20, 2026): https://www.tomshardware.com/pc-components/ssds/phison-details-ssd-with-a-built-in-npu-and-64gb-of-memory-that
- ElectronicsForU (November 19, 2026): https://www.electronicsforu.com/news/three-storage-technologies-target-ai-bottlenecks
- TechSpot (November 21, 2026): https://www.techspot.com/news/114653-phison-shows-ssd-design-npu-64gb-memory-runs.html
- TechPowerUp (November 22, 2026): https://www.techpowerup.com/355089/phison-previews-an-ssd-that-runs-ai-frameworks-on-its-own-npu
- DIGITIMES (November 19, 2026): https://www.digitimes.com/news/a20261118PD260/phison-hpc-pcie-performance-data.html
Sources
- https://www.ctee.com.tw/news/20261119700981-430201
- https://www.tomshardware.com/pc-components/ssds/phison-details-ssd-with-a-built-in-npu-and-64gb-of-memory-that
- https://www.electronicsforu.com/news/three-storage-technologies-target-ai-bottlenecks
- https://www.techspot.com/news/114653-phison-shows-ssd-design-npu-64gb-memory-runs.html
- https://www.techpowerup.com/355089/phison-previews-an-ssd-that-runs-ai-frameworks-on-its-own-npu
- https://www.digitimes.com/news/a20261118PD260/phison-hpc-pcie-performance-data.html