Huawei’s Ascend 950 AI processor

Second in a line AI processors.

September 19, 2025

Jon Peddie

Huawei’s Ascend 950DT NPU targets training and decode-stage inference, and anchors the Atlas 950 platform. DaVinci architecture combines matrix, vector, and scalar engines with 144GB HiZQ 2.0 HBM (4 TB/s) and 2 TB/s interconnect, supporting FP8, MXFP8, and XMFP4. The hierarchy scales from 8,192-chip supernodes to 64 SuperPoDs; Huawei cites configurations over 524,000 devices. HiBL 1.0 (128GB, 1.6 TB/s) pairs with 950PR for prefill. CANN and MindSpore provide kernels, graph optimization, and scheduling. Roadmaps add 960/970 parts. Market context includes Nvidia Rubin, AWS Inferentia/Trainium, Google TPU, and Microsoft Maia platforms. (Source: Huawei) Huawei introduced the Ascend 950DT NPU at Connect

...

Enjoy full access with a TechWatch subscription!

TechWatch is the front line of JPR information gathering service, comprising current stories of interest to the graphics industry spanning the core areas of graphics hardware and software, workstations, gaming, and design.

A subscription to TechWatch includes 4 hours of consulting time to be used over the course of the subscription.

Already a subscriber? Login below

Huawei’s Ascend 950 AI processor

Enjoy full access with a TechWatch subscription!

This content is restricted

Login