Phison unveils enterprise SSD concepts with onboard NPU and pSLC-QLC caching
At Supercomputer 2026, Phison showed two enterprise SSD designs still under development. One pairs an NPU with 64GB of memory running at 130 GB/s, letting frameworks like PyTorch, vLLM, OpenVINO, and llama.cpp execute directly on the drive, while the other combines fast pSLC cache with large QLC capacity under a single controller.
GoKawiil's interpretation of the reporting above, not reported fact.
Running AI frameworks directly on storage could let servers process data where it sits rather than shuttling entire datasets to the CPU, potentially easing bottlenecks in AI-heavy workloads. Folding cache into every drive bay, rather than dedicating separate cache drives, could also simplify system design and recover usable capacity, though these remain early-stage concepts pending finalization. The NPU's quoted performance figure is an outdated benchmark type, so its actual AI capability is still unclear.
- Phison's computational-storage SSD design embeds an NPU and 64GB of 130 GB/s memory to run AI frameworks on-drive.
- A separate design merges pSLC cache and QLC capacity under one controller to avoid dedicating drives solely to cache.
- Both designs remain in development, and the NPU's quoted 36,800 DMIPS figure is an outdated benchmark that doesn't clearly indicate AI performance.
Source: tomshardware.com — Shane Downing, 2026-11-20
Published there as: “Phison details SSD with a built-in NPU and 64GB of memory that supports AI frameworks on the drive”
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.