Huawei Launches AI Chip With 2.87x Nvidia H20's Power

Edgen Stock·Mar 22 2026, 00:56

Share to

Share to

Copy link

Key Takeaways

Huawei has unveiled its Atlas 350 AI accelerator card, a direct competitor to Nvidia's offerings in China. The new hardware, powered by the self-developed Ascend 950PR chip, demonstrates significant performance gains and signals China's accelerating push for technological self-sufficiency in the critical AI sector.

Huawei's Atlas 350 card, powered by the new Ascend 950PR chip, delivers 2.87 times the computing power of Nvidia's H20, marking a significant domestic alternative.
The new chip is China's first commercial product to support FP4 inference, a low-precision format that boosts multimodal generation speeds by up to 60%.
Seven hardware partners, including iSoftStone, are already launching servers based on the Atlas 350, indicating rapid commercialization and adoption within China's tech ecosystem.

Atlas 350 Delivers 2.87x Nvidia H20's Computing Power

Huawei has launched its Atlas 350 AI accelerator card, built on the new Ascend 950PR processor, directly challenging Nvidia's position in the Chinese market. Announced at the Huawei China Partner Conference 2026, the company stated the single-card's computing power is 2.87 times greater than that of Nvidia's H20, a chip specifically designed for export to China.

The Atlas 350 achieves a formidable 1.56 petaflops of FP4 computing power, a low-precision format that speeds up data processing. This makes it China's first commercial product to support FP4 inference, which improves efficiency for tasks like content recommendation and video generation. The card also features 112GB of high-bandwidth memory (HBM), 1.16 times more than the H20, and boosts multimodal generation speeds by 60%. While its performance is superior, its power consumption is also higher at 600W, 1.5 times that of the H20.

Ecosystem Mobilizes for Commercial Rollout

The Ascend 950PR platform is moving immediately into commercial use, with seven of Huawei's core partners announcing server products based on the Atlas 350. These partners include Kunlun, Huakun Zhenyu, and iSoftStone, a publicly traded company. iSoftStone introduced its "A860 A5" AI server, a 6U machine capable of hosting eight Atlas 350 cards for large language model training and inference.

Further cementing its ecosystem integration, AI firm iFlytek announced its new Spark large model will be fully adapted for Huawei's Ascend 910 and 950 series chips. This rapid adoption by both hardware and software players signals a concerted effort to build a robust domestic AI infrastructure independent of foreign technology, targeting high-demand scenarios like short-form video, e-commerce, and advertising recommendations.