Huawei unveils next-gen Ascend 920 AI chip, a Chinese alternative to NVIDIA H20 AI GPU

Huawei unveils the Ascend 920 AI chip, the successor to Ascend 910C: expected to follow the market gap left by NVIDIA H20 AI GPUs in China.

Huawei unveils next-gen Ascend 920 AI chip, a Chinese alternative to NVIDIA H20 AI GPU
Comment IconFacebook IconX IconReddit Icon
Gaming Editor
Published
1 minutes & 30 seconds read time

As an Amazon Associate, we earn from qualifying purchases. TweakTown may also earn commissions from other affiliate partners at no extra cost to you.

TL;DR: Huawei's new Ascend 920 AI chip, built on SMIC's 6nm process, delivers 900 TFLOPs and 4TB/sec memory bandwidth, outperforming its predecessor and rival NVIDIA's restricted H20 GPU. Integrated in the CloudMatrix AI server, it offers 30-40% higher performance, positioning Huawei as a strong competitor in AI hardware.

Huawei just unveiled its new Ascend 920 AI chip, the successor to its China-exclusive Ascend 910C -- which is being used in a new "CloudMatrix" AI server -- within hours of new US export restrictions on NVIDIA's H20 AI GPU entering the country.

Huawei unveils next-gen Ascend 920 AI chip, a Chinese alternative to NVIDIA H20 AI GPU 83

In a new report from DigiTimes, we're learning that Huawei's new Ascend 920 is a direct competitor to the NVIDIA H20 AI GPU, with a focus on the supply chain being entirely in-house with this new chip. Huawei's new Ascend 920 will be made on SMIC's 6nm process, which is an upgrade over the Ascend 910C which was fabbed on 7nm at SMIC.

We are expecting to see 900 TFLOPs of compute performance, 4TB/sec of memory bandwidth through its HBM, and a chiplet design that is similar to what we've seen with the Ascend 910C. This should all lead into around 30-40% more performance from the new Ascend 920 over 910C, meaning the Ascend 920 is a better AI chip than the now even-more-restricted H20 AI GPU from NVIDIA.

In related news, Huawei's current-gen Ascend 910C was used in a 384-chip cluster in China dubbed "CloudMatrix" which offers more AI performance than NVIDIA's beasty (and not available in China) GB200 NVL72 AI server cabinet.

It will be interesting to see how much better a "CloudMatrix 2" system could get, with new and improved Ascend 920 AI chips inside, as we should expect an additional 30-40% more performance out of an Ascend 920-powered CloudMatrix, which should keep up with NVIDIA's next-gen AI servers... albeit with far more power consumption (3.9x more power used on CloudMatrix than GB200 NVL72).