Newshot

Phison Showcases Smart SSD with Built-in NPU and 64GB Memory That Can Run AI Models Directly

AI summary

At the SC26 conference, Phison showcased two enterprise SSDs designed for AI scenarios, one of which features a built-in NPU and 64GB of memory, enabling it to run AI frameworks and inference tools such as PyTorch, vLLM, OpenVINO, and llama.cpp directly on the drive with 130GB/s of bandwidth. The other product uses a Mixed Media design that combines high-endurance pSLC cache with high-capacity QLC flash under a single 16-channel controller, with each storage area assigned its own namespace.

Translation(model translation, not the source's own text)

Phison Demonstrates Smart SSD with Built-in NPU and 64GB Memory That Can Run AI Models Directly

IT Home, November 22 news: According to a report by tech media Tom's Hardware yesterday, Phison showcased two enterprise-grade SSDs designed specifically for AI scenarios at the SC26 conference. One of them even comes with an NPU and 64GB of memory, and can run AI models directly.

IT Home learned from the original report that one of the products, an SSD, has a built-in NPU (neural network engine) and 64GB of memory, with bandwidth reaching 130GB/s, and supports running AI frameworks and inference tools such as PyTorch, vLLM, OpenVINO, and llama.cpp directly on the drive.

The design approach of this product is to offload some computing tasks from the server to the SSD, letting the device process stored data directly and return only the results to the host, without the need to transmit the full dataset.

Phison says this solution is expected to reduce data transmission volume by 10 to 100 times under specific workloads, thereby lowering the load on the host CPU.

The other product adopts a new Mixed Media hybrid media design, integrating high-endurance pSLC cache with high-capacity QLC flash in the same SSD, managed by a 16-channel controller.

This solution switches part of the QLC flash to SLC operating mode, improving write performance and endurance at the cost of some storage capacity. The two storage areas are each assigned independent namespaces, and the host can select the corresponding storage area based on data type without installing additional specialized software.

Phison says this approach can reduce reliance on separate cache drives. In a system containing 24 SSDs, if one of the drives fails, the cache loss ratio is only 4%, whereas a traditional solution using 4 separate cache drives could lose 25% of cache capacity.

Related Articles

  • Phison Launches Pascari Space: Custom SSDs Built for Space Storage
  • Lite-On Storage Technology Launches Its First Industrial-Grade QLC SSD PQ2, Offering 3-Year 1DWPD Warranty
  • ORICO Launches FlashMem FM9800 SSD with pSLC AI Cache Area
  • Nordic Launches Wi-Fi SoC "nRF71" Series Supporting Edge AI Inference, Reducing Standby Power Consumption by Up to 1/20
  • Huawei Kirin 9050 Series Chips Officially Announced with Next-Generation Da Vinci Architecture NPU, Industry's First 30B MoE Large Model on Device
  • Phison's Revenue for the First Three Quarters of Fiscal 2026 Grew 290% Year-on-Year, Hitting a New High for the Same Period in History, with Net Profit Attributable to Shareholders of NT$71.938 Billion

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗