Huawei AI chips edge out Nvidia H800, hinting progress over US tech controls

Huawei Ascend AI chips are reportedly overtaking Nvidia H800, flagging a win on US tech controls. The company has recently said that its chips are still a generation behind foreign technologies, although the latest report points out something else.

A technical paper revealed that Huawei AI chips outperformed Nvidia H800 used in the DeepSeek R1 model. The credit goes to the all-new AI CloudMatrix 384.

Nvidia CEO Jensen Huang has applauded Huawei’s CloudMatrix many times. But it’s the first ever occasion when the company itself shared its AI supernode benefits.

SiliconFlow, a Chinese AI startup, and Huawei researchers have written this paper and defined the key details and characteristics of the new AI supernode technology.

CloudMatrix uses 384 Ascend 910C NPUs (Neural Processing Units) and 192 Kunpeng server CPUs. These SoCs are interconnected via a unified bus that delivers ultra-high bandwidth and low latency, enhancing the overall performance.

The company believes that CloudMatrix 384 is a major solution that can overcome the US tech control measures and reshape the foundation of the AI infrastructure.

(Image Credits: Huawei)

The paper further describes that CloudMatrix 384 enables Ascend chips to surpass some most significant AI technologies of the world, used in the DeepSeek R1. It can even play wonders for data centers and manage computing infrastructure.

Huawei AI supernode has a throughput of 6688 tokens per second per NPU for a 4000-token prompt length. Here, tokens refer to the basic unit of LLMs. Meanwhile, the token length affects cost, time, and AI models’ understanding capabilities.

The paper findings reveal that CloudMatric scored 1943 tokens per second per NPU for a 4000-length, resulting in more efficient use of Ascend AI chips. These metrics eventually surpassed that of Nvidia’s chips used for LLMs.

According to the latest report, the technical paper went official earlier this week. Zuo Pengfei, the main author of this paper, shared a post on Quora on Wednesday, saying

“… fully and transparently showcases Huawei CloudMatrix’s comprehensive technology stack. This aims to help the industry fully understand the capabilities of domestic Ascend NPUs. The paper also aims to build confidence within the domestic technology ecosystem in using Chinese-developed NPU to outperform Nvidia’s GPUs.

|| source ||