Huawei's Big Step Forward

Huawei's Ascend AI chip has surpassed Nvidia GPU performance in the DeepSeek R1 model test, showing that the Chinese tech giant is looking to break the US monopoly.

According to SCMP , a new technical report shows that Huawei's Ascend AI processor has surpassed the performance of Nvidia GPUs in the DeepSeek R1 artificial intelligence model test. This result marks an important step forward for the Chinese company in the context of facing technology control orders from the US.

The paper, co-authored by researchers from Huawei and AI infrastructure startup SiliconFlow, details the architecture of the CloudMatrix 384 data center. It is an “AI supersystem on a chip” specifically designed to handle massive workloads from large language models (LLMs).

CloudMatrix 384 consists of 384 Ascend 910C processors and 192 Kunpeng CPUs, connected together by a unified bus system with low latency and high bandwidth. This architecture is expected by Huawei to reshape the foundation of AI infrastructure.

The CloudMatrix-Infer system built on this platform shows outstanding performance when deploying the DeepSeek R1 model. In the prefill phase, the system achieved a throughput of 6,688 tokens/second per NPU when processing a 4,000-token request from a user. This performance is equivalent to 4.45 tokens/second per TFLOPS (trillion operations per second).

During the decoding phase, CloudMatrix recorded a throughput of 1,943 tokens/second per NPU, with a response time of less than 50 milliseconds per token. This performance of 1.29 tokens/second per TFLOPS is higher than Nvidia's SGLang system using H100 GPUs and also the H800 GPU platform running DeepSeek R1.

According to Zuo Pengfei, the lead author of the study and a member of Huawei's "Genius Youth" program, the aim of the report is to help the industry fully understand the capabilities of China-developed Ascend NPU. On the Zhihu platform, he affirmed that the report is a clear demonstration of Huawei CloudMatrix's technological capabilities.

This is the first time Huawei has disclosed technical details about the Ascend 910C AI accelerator. The report also reiterates Huawei founder Ren Zhengfei’s recent remarks that Ascend chips are still “a generation behind” their US counterparts. However, he asserted that the application of methods such as “stacking and clustering” has helped Huawei achieve computing performance comparable to today’s most advanced AI systems.

Meanwhile, Nvidia CEO Jensen Huang also admitted in an interview with CNBC that AI is a parallel processing problem and they will need more processors if they don't have enough computing power.

Source: https://znews.vn/buoc-tien-lon-cua-huawei-post1562828.html