The government prefers that its companies no longer buy NVIDIA chips

The government led by Xi Jinping wants Chinese tech companies to stop buying integrated circuits from NVIDIA. At the moment, the step taken, according to SCMP, is only informational in nature, but it is directly addressed companies purchasing GPUs for artificial intelligence (AI) to recommend that they buy Chinese-made chips rather than solutions from Jensen Huang’s company.

US administration sanctions prevent NVIDIA from selling its most advanced artificial intelligence GPUs, such as the A100 or H100 chips, to its Chinese customers. These bans forced the US company to develop modifications to its GPUs that were refined enough to meet the requirements of the US Department of Commerce. NVIDIA plans to ship more than a million H20 chips to its Chinese customers by the end of this year. And this is possible because Joe Biden’s government is allowing this GPU to be sold in China.

Either way, NVIDIA is losing competitiveness in China. It’s inevitable. The H20 GPU is, on paper, far less efficient than the most sophisticated AI chips Jensen Huang’s firm currently sells. At the moment, China’s Ministry of Industry and Information Technology has not made any official announcements, but according to SCMP, the administration led by Xi Jinping has limited the purchase of NVIDIA H20 GPUs by Chinese companies since August last year.

Chinese alternatives are already ready

Dozens of Chinese companies are developing their own artificial intelligence equipment. MetaX, Alibaba, Biren Technology, Moore Threads, Innosilicon, Zhaoxin, Iluvatar CoreX, DenglinAI or Vast AI Tech are some of the most important, but one shines brighter than all the rest: Huawei. This company is doing well in the Chinese market (its annual turnover in China alone is about $7 billion), and it is quite happy with the weakening of NVIDIA.

And Huawei has had its own graphics processors for artificial intelligence – Ascend AI chips – for more than five years. During this period of time, it improved them and increased its capabilities with the goal of equaling or even surpass NVIDIA A100 and H100 chips. According to some analysts, such as the Chinese company iFly Tek, the raw power of their GPUs is equal to that of NVIDIA chips, but they are still a step behind if we stick to their performance in a real-world use case.

For Huawei, the current situation represents an excellent opportunity to strengthen its presence in its own country. However, both the company and other Chinese companies face a major challenge that will prevent them from easily capturing the market that NVIDIA has so far dominated: most AI projects being developed inside and outside China are implemented on CUDA (Unified Computing Device Architecture). ). This technology combines the compiler and development tools used by programmers to develop software for NVIDIA GPUs.

Huawei has CANN (Computational architecture for neural networks), which is its alternative to CUDA, but there is a consensus among analysts that this platform is still very far from the capabilities that NVIDIA technology has in the process of training AI models. Considering the circumstances Huawei only has one option If the company wants to gain a foothold in the Chinese AI chip market: refine CANN and offer NVIDIA customers the ability to easily bring their models and data into their own ecosystem.

It seems like that’s exactly what he’s doing. However, this Chinese company has a very valuable ally to help it break through: the Ascend 910C AI GPU. Many Chinese tech companies are looking for alternatives to NVIDIA chips, and Huawei is taking the opportunity to send them samples of its Ascend 910C GPU for testing. This chip is an improved version of the Ascend 910B, which Huawei itself says has power comparable to the NVIDIA A100, so it seems reasonable to predict that the 910C will be even more powerful if possible.

Image | NVIDIA

Additional information | SCMP

In Hatak | No one can beat NVIDIA: it will earn $10,000 million for the 450,000 Blackwell GPUs it plans to produce