Urgent.News

What's breaking now, across thousands of outlets.

AI

Phison details SSD with a built-in NPU and 64GB of memory that supports AI frameworks on the drive

Phison showed a computational-storage SSD with an NPU and 64GB of memory, plus Mixed Media pSLC and QLC drives, at SC26.

Phison details SSD with a built-in NPU and 64GB of memory that supports AI frameworks on the drive

Phison showcased two new enterprise SSD designs at Supercomputer 2026 (SC26) to tackle specific challenges: one featuring an integrated Neural Processing Unit (NPU) and 64GB to 130GB of memory capable of running AI frameworks directly on the drive, like PyTorch, vLLM, OpenVINO, and llama.cpp, while the other combines a fast pSLC cache with large QLC capacity within a single controller.

The company mentioned that many of the features are still in development and will be released once finalized. The key idea behind these SSDs is to offload more processing tasks to the drive itself, running AI code on data stored on the drive and sending back results, rather than moving entire datasets around. This approach could potentially reduce the amount of data transferred to the host CPU, freeing it up for other tasks.

The integrated NPU is expected to deliver 36,800 DMIPS of compute performance, which, while an outdated CPU benchmark, could correspond to the performance of the CPU cores working alongside the NPU. More relevant to the expected workload, however, are the memory and its bandwidth. The SSD is designed to support multiple general-purpose inference servers and runtimes without requiring specialized programming tools, making it more accessible for developers.

The drive supports PCIe 6.0 interfaces, which could provide up to 28 GB/s to the host, making the memory bandwidth potentially several times faster than current high-end SSDs. Nonetheless, Phison emphasizes that the drive can be utilized with common AI frameworks without necessitating new code development.

Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at tomshardware.com →

More in AI

More from Friday 20 November →