Phison presents SSD design for running AI workloads on stored data
The design includes an NPU and 64GB of memory, but Phison has not announced a product name or sampling date.

Phison has presented an SSD architecture intended to process AI workloads within the drive instead of moving entire datasets to a server's processors. Shown at Supercomputing 2026 (SC26) in Chicago, the design combines an NPU with 64GB of memory.
The company says the drive would process stored data in parallel and return relevant results to the host. For suitable workloads, Phison claims the approach could reduce data transfers by 10 to 100 times while easing demand on the host CPU.
Phison lists memory bandwidth of 130 GB/s, compared with the 28 GB/s host-transfer capability of its upcoming PCIe 6.0 controller. The design is intended to run PyTorch and llama.cpp, and the company says support for vLLM and OpenVINO would let developers work without specialized programming tools.
Details about AI performance remain limited. Phison disclosed a figure of 36,800 DMIPS but provided no TOPS rating. Tom's Hardware said the DMIPS figure likely refers to companion CPU cores rather than the NPU's AI performance.
Phison has not identified the memory type or provided a sampling date. The company also said many products and features shown at the event remain under development.




