Skip to content
Tech News
← Back to articles

Phison unveils SSD prototype with onboard NPU and 64GB memory for in-drive AI

read original more articles
GoKawiil Brief

Phison showcased an SSD design at Supercomputing 2026 that pairs an NPU with 64GB of memory, letting the drive run frameworks like PyTorch and llama.cpp directly on stored data instead of sending it to a host CPU. The company says the design could reduce data transfers by 10 to 100 times for certain workloads, with onboard memory bandwidth of 130 GB/s versus 28 GB/s for its upcoming PCIe 6.0 controller. No product name, sampling date, memory type, or TOPS performance figure has been announced, and Phison describes much of what it showed as still in development.

Why It Matters

GoKawiil's interpretation of the reporting above, not reported fact.

Moving computation closer to stored data could ease bottlenecks in AI servers where CPUs and GPUs spend significant time waiting on data transfers from storage, potentially improving efficiency for data-heavy AI workloads. However, the lack of a clear AI performance metric and product timeline suggests this remains a concept demonstration rather than a near-term commercial offering. Similar past efforts, like Samsung's SmartSSD with onboard FPGAs, struggled to gain lasting traction, which raises questions about whether in-storage compute designs can achieve broad adoption.

Key Takeaways

Source: techspot.com — Zo Ahmed, 2026-11-21

Published there as: “Phison shows an SSD design with an NPU and 64GB of memory that runs AI on the data it stores”

Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.
Keep reading
llama.cpp 15 briefs → npu 13 briefs → supercomputing 2026 5 briefs →