Phison details SSD with a built-in NPU and 64GB of memory that supports AI frameworks on the drive
Reported by Tom's Hardware and Mezha2 sources ↓
At SC26, Phison detailed an enterprise SSD with an onboard NPU and 64GB of memory, designed to run AI frameworks on the drive. It also described a separate design combining pSLC cache and QLC capacity.
Key points
- The computational-storage SSD has 64GB of memory with stated bandwidth of 130 GB/s
- Phison says it supports PyTorch, vLLM, OpenVINO and llama.cpp on the drive
- A separate 16-channel design combines pSLC and QLC flash behind one controller
Sources
Details
Phison presented the two enterprise SSD designs at Supercomputer 2026. The computational-storage model is intended to run AI code on data stored on the drive and return results to the host, rather than transferring entire datasets. Phison says applicable workloads could involve 10 to 100 times less data transferred, but the company did not explain how a host would access the drive’s compute.
The computational-storage design has an NPU and 64GB of memory, with Phison specifying 130 GB/s of bandwidth. The company lists PyTorch, vLLM, OpenVINO and llama.cpp as supported frameworks or runtimes. It cited 36,800 DMIPS of compute performance, but Tom’s Hardware notes that this is a CPU-oriented benchmark and does not establish the NPU’s AI performance.
The second design combines high-endurance pSLC and high-capacity QLC flash in one 16-channel drive. Phison says the regions are managed separately and appear as separate namespaces to the operating system, allowing data to be directed to either region. The company says the design uses standard NVMe interfaces and does not require specialized host software.
It did not provide availability dates.
Written from coverage as of
Images
- Tom's Hardware
- Tom's Hardware