The AI landscape in China is heating up, with Huawei making significant strides in challenging Nvidia's dominance. Recently, Huawei announced that its Ascend AI GPUs now support DeepSeek-optimized inference. This development could reduce the reliance of Chinese AI firms on Western technology amid ongoing US sanctions.
On the same day Nvidia's stock prices dropped due to the market recognizing the implications of the Chinese LLM, DeepSeek, Huawei announced that the distilled R1 AI model is available for free via its ModelArts Studio platform. This version is explicitly "Ascend-adapted," meaning it's optimized for Huawei's Ascend data center GPUs.
While Huawei hasn't specified which Ascend GPUs are used for ModelArts Studio, industry experts speculate it could be the latest Ascend 910C. This new GPU began sampling to customers around September, making its inclusion in Huawei's cloud servers a strong possibility.
This is a significant step for Huawei, signaling the company's GPU's practical ability to run LLMs. This explicit support could reduce the need for AI companies to depend on Nvidia and AMD GPUs for both training and inference.
According to AI industry figure Yuchen Jin, the inference performance of the Huawei 910C achieves 60% of the H100's performance. Jin also noted that with handwritten CUNN kernels and optimizations, the performance could be even higher. Although the R1 was trained using Nvidia's H800 chips, the Ascend 910C could later be used for training, too.
The US government's export restrictions have created a challenging environment for Nvidia in China. The sanctions prevent the sale of high-performance processors, forcing Nvidia to develop new, less powerful models specifically for the Chinese market. For example, The H800 was initially launched as an alternative to the banned H100 after the initial GPU export restrictions on China. Later, the H800 and other Nvidia GPUs for the Chinese Market were banned after another round of sanctions.
Nvidia's flagship product for China, the H20, has significantly less memory, memory bandwidth, and TFLOPs than the top-end H200. This has impacted Nvidia’s sales, with the H20 selling for less than Huawei’s Ascend 910B.
Huawei's advancements in AI GPUs could be a turning point for China's technological independence. If Chinese GPUs like the Ascend 910C prove capable of handling both training and inference, the demand for processors restricted by US sanctions, such as the H20, may decrease.
While China still relies on Western chips to some degree, companies like Huawei are actively working towards advancing domestic chip manufacturing capabilities.
This development marks a crucial step in China's pursuit of AI independence, potentially reshaping the global AI landscape.