China's Huawei is stepping up its competition with Nvidia in the AI GPU market. The company has announced DeepSeek-optimized inference support for its Ascend AI GPUs, a move that could reduce reliance on Western technology and boost China's self-sufficiency in AI. This development comes as Nvidia faces challenges in China due to US export restrictions, creating an opportunity for domestic players like Huawei to gain ground.
On January 27th, the same day Nvidia's stock price experienced a significant drop, Huawei announced that the distilled R1 AI model was available for free via its ModelArts Studio platform. This version is specifically adapted for Huawei's Ascend data center GPUs.
Huawei's explicit statement that the R1 model is "Ascend-adapted" underscores the strategic importance of this development. While the exact Ascend GPUs used for ModelArts Studio aren't detailed, industry experts speculate it could be the latest Ascend 910C, which was sampled to customers in September.
The fact that Huawei's GPUs now have explicit support for running the R1 LLM is significant. Previously, Chinese AI firms heavily relied on Nvidia and AMD GPUs for both training and inference. This move by Huawei potentially cuts out a part of the process where Chinese companies depend on Western technology.
According to Yuchen Jin, an AI industry figure, the inference performance of Huawei's 910C achieves 60% of the H100's performance. With handwritten CUNN kernels and optimizations, the performance could be even higher. While the R1 model was initially trained on Nvidia H800 GPUs, the Ascend 910C could potentially be used for training in the future.
US sanctions prevent Nvidia from selling its most powerful data center GPUs, like the H200 and B200, in China. This forces Nvidia to develop new, less powerful models specifically for the Chinese market. As a result, Nvidia's flagship offering for China, the H20, has reduced memory, memory bandwidth, and TFLOPs compared to its top-end cards.
The US government's export restrictions have significantly impacted Nvidia's position in China. In May 2024, Nvidia was reportedly selling the H20 for less than Huawei's Ascend 910B. However, sales of the H20 improved in the latter half of the year, demonstrating the ongoing demand for AI GPUs in China. Read more about Nvidia's AI chip sales in China.
The ability to run Chinese LLMs with high performance on domestic processors marks a significant step towards technological independence for China. If the Ascend 910C or other Chinese GPUs prove capable for both training and inference, the need for Western chips like the H20 would diminish. This aligns with China's broader strategy to strengthen its semiconductor industry and reduce reliance on foreign technology. Learn about Huawei's efforts in chip manufacturing.
This development is a crucial indicator of the shifting dynamics in the global AI landscape, and it highlights the increasing competitiveness of Chinese technology companies. As Huawei continues to innovate and improve its Ascend AI GPUs, the competition with Nvidia in the Chinese market is expected to intensify further.
