Huawei Advances Domestic Semiconductor Ecosystem with Ascend AI Chips

Huawei’s Ascend AI chip ecosystem is expanding as China seeks to establish an independent semiconductor industry. The company’s in-house processors and their associated supplier network form the backbone of a national effort for full domestic production of chips, including high-end AI chips, optical networks, packaging materials, photoresists, and gas delivery systems.

More than 60 semiconductor companies are backed by Huawei’s investment arm Hubble, while local partners like Empyrean are working on design toolchains to create an alternative to Nvidia’s software ecosystem. This was showcased at the China Hi-Tech Fair in Shenzhen with the unveiling of the CloudMatrix 384 system, which integrates 384 Ascend 910C processors and is positioned as a direct competitor to Nvidia’s GB200 platform.

The Ascend 910C, built using stacked HBM2E memory and a DaVinci NPU architecture tailored for AI workloads, delivers up to 780 TFLOPS of dense BF16 compute while consuming 350 watts. Despite trailing Nvidia’s products in peak throughput and power efficiency, Huawei aims to offset this with scale. The CloudMatrix 384 system combines twelve racks of Ascend modules and four optical interconnect racks, delivering around 300 PFLOPS in total.

Chinese data centers can handle the higher energy use due to fewer regulatory constraints and lower power costs. Huawei claims its internal tests show that the CloudMatrix outperforms Nvidia H100 platforms on specific model classes, although public benchmarks are scarce. The company’s CANN programming environment and MindSpore framework continue to mature, supporting common model architectures through a translation layer.

Source: tomshardware.com