Phison has outlined enterprise storage designs that put processing closer to stored data, including a computational-storage architecture with an integrated neural processing unit. Its November 17 announcement at SC26 describes technology under development, rather than a newly available consumer SSD.
The company lists 64 GB of onboard memory with 130 GB/s bandwidth and support for PyTorch, vLLM, OpenVINO and llama.cpp. That memory-bandwidth figure is not a claim that the SSD transfers files to a computer at 130 GB/s.
Phison says processing data inside drives could reduce transfers by 10 to 100 times for suitable workloads. That is a manufacturer projection about data movement, not independently measured application speed or a promise that every AI task runs 100 times faster.
A separate Mixed Media design combines faster pSLC storage and higher-capacity QLC flash behind one controller. Phison presents it as a way to put caching and capacity in each drive instead of reserving bays for dedicated cache drives.
The release warns that products and features remain in development and gives no release date for these two designs. Tom’s Hardware’s November 20 analysis also identifies unanswered implementation questions. Neither source provides an independent benchmark of the proposed computational-storage drive.
Editorial Opinion
Moving less data could be useful, but the missing measurements matter. Phison labels its compute figure as 36,800 DMIPS; Tom’s Hardware notes that this older integer benchmark does not establish AI-accelerator performance. That unresolved metric should not be converted into an AI speed ranking. Useful next evidence would include shipping hardware, documented software interfaces, reproducible workloads and comparisons that count power use as well as execution time.
References
Phison Electronics. (2026, November 17). SC26: Phison unveils three next-generation storage innovations for the new compute era. Announcement.
Downing, S. (2026, November 20). Phison details SSD with a built-in NPU and 64GB of 130 GB/s memory to run PyTorch and llama.cpp on-drive — claims up to 100x less data transfer alongside new mixed media pSLC and QLC design. Tom’s Hardware. Analysis.
Have a factual correction or website problem? Contact PartialMedia with the article link, disputed passage and supporting source.
Discussion
Comments load automatically through Disqus. This connects to Disqus and its partners, which may use cookies and process your IP address and browser information. Disqus privacy policy.