Huawei challenges Nvidia with powerful new AI accelerator card
The Atlas 350 can deliver 1.56 petaflops of FP4 computing power, a 2.8-times improvement over Nvidia’s H20 chip, an executive says
Chinese tech giant Huawei Technologies has launched the Atlas 350 accelerator card for inference, boasting higher computing power for artificial intelligence applications and better performance than US rival Nvidia’s H20, as AI rapidly advances into the agentic era.
The Atlas 350 AI accelerator card would be powered by Huawei’s latest Ascend 950PR chip, designed to deliver enhanced computing power and storage for AI inferencing, said Ma Haixu, a vice-president at Huawei, at the firm’s China Partner Conference on Friday.
The card delivered 1.56 petaflops of FP4 computing power, a 2.8-times improvement over Nvidia’s China-tailored H20 chip, according to Zhang Dixuan, head of Huawei’s Ascend computing business. FP4 refers to low precision in computing, which allows the Atlas 350 accelerator to move data faster.
The accelerator card aimed to match or even exceed its industry peers in AI inference for search recommendation, multimodal generation and large language models, Zhang said.
An accelerator card is a hardware unit dedicated to a specialised task and designed to be integrated into a server.
The launch came as the US-sanctioned company ramped up efforts in AI computing infrastructure powered by its self-developed semiconductors, after achieving chip breakthroughs in recent years, including its Ascend AI chips, without relying on American technology.