Ascend 950
#16 in AI Inference HardwareHuawei · v950 · 2× · last seen Sep 18, 2026
The Huawei Ascend 950 (commercially introduced as Ascend 950PR) is Huawei's current AI accelerator chip for Inference workloads (Prefill and recommendation services), positioned as an alternative to Nvidia GPUs such as the H20. The chip is based on the Da Vinci 3.0 architecture, uses Huawei's proprietary HBM replacement "HiBL 1.0," and is manufactured by SMIC. It commercially debuted in March 2026 on the Atlas 350 accelerator card and delivers up to 1.56 PFLOPS FP4 computing performance at 600 W TDP, according to Huawei. Due to US export controls, the chip is practically only available in China and is distributed through Huawei Cloud as well as major Chinese customers including ByteDance, Alibaba, and Tencent.