Phison Previews Computational SSD With NPU
The concept pairs a 36,800 DMIPS NPU and 64GB of on-drive memory with NAND storage to process AI workloads near the data.
Updated on Nov. 21, 2026 in Computers

Phison previewed a computational SSD architecture that places an NPU and 64GB of memory alongside NAND flash, rated at 36,800 DMIPS with 130GB/s memory bandwidth. The company presented the concept alongside its next-generation PCIe 6.0 SSD controller, the X3 (PS5303).
Why it matters
The design aims to reduce data movement in AI and high-performance computing workloads by running frameworks such as PyTorch, vLLM, OpenVINO and llama.cpp close to stored data. Company claims suggest transfers could fall 10 to 100 times for applicable tasks, though no product has been announced.
The on-drive NPU is rated at 36,800 DMIPS with 64GB of memory at 130GB/s, and the X3 controller targets 28GB/s sequential throughput and 6.8 million IOPS over PCIe Gen6 x4. Phison has not disclosed a launch timeline, and the 10 to 100 times transfer reduction remains a company claim.
The players
Phison
Phison is a Taiwan-based NAND flash controller company whose products are widely used in consumer and enterprise solid-state drives.
The details
Phison's concept integrates an NPU and local memory alongside NAND storage, using a Mixed Media design that combines pSLC and QLC flash behind one controller across 16 channels. The X3 controller, also identified as PS5303, is a PCIe 6.0 SSD controller presented alongside the architecture.
Timeline
Nov 21, 2026: Phison's computational SSD preview was published.
The Tech Race
The architecture extends the computational storage concept by embedding an NPU and memory directly in the SSD, placing Phison among firms seeking to shift AI processing toward storage hardware. It arrives as PCIe 6.0 controllers begin to appear in next-generation drives.
The concept could eventually reduce the time AI tools take to process large local datasets by limiting data transfers to the host system. No product, price or availability has been announced, so buyers cannot yet act on it.
The takeaway
Computational storage aims to move AI processing to where data already sits, which could ease bottlenecks in data-heavy workloads if company claims hold up in shipping products. Readers should watch for a product name and release schedule before drawing conclusions about real-world gains.
Further reading
Explore more Computers coverage for context on storage and processing hardware.
Source note: This article includes information reported by Guru3D.



