©2021 Reporters Post24. All Rights Reserved.
DeepSeek R1 AI model has gathered attention within a few days of launch and a major credit goes to Huawei chips. Even though the LLM trained on Nvidia’s AI SoCs, it uses Ascend 910C for inference. This move shows China’s support for Huawei.
Nvidia and Huawei have been battling in the AI chip field for years. Due to the US regulations, Nvidia is unable to sell advanced chipsets in China. Yet, it still has a major foothold in the Chinese AI semiconductor industry for its foreign-based techs.
Huawei on the other hand is gradually improving its chipset capabilities. The phone maker recently built the Ascend 910C processor which can par with the Nvidia H100.
The US chipmaker is still a strong challenger in front of Huawei. However, the Chinese OEM’s latest move in the AI battle shows that it is ready to counter its rivals.
DeepSeek R1 which has gained a lot of popularity in the market for its cost-effective and advanced tech properties, uses Huawei Ascend 910C AI chips for inference. Ascend 910C is slightly inferior to Nvidia H100 in performing basic learning tasks. However, it can offer lower energy costs and higher operational efficiency than H100.
R1 used Nvidia H100 chips for training but its inference is powered by Ascend 910C processors. Something more catchy is that despite using a moderately powerful AI processor, the China-made AI model is surpassing U.S. opponents like OpenAI ChatGPT, Google Gemini, and other top LLMs (large language models).
One of the AI analysts – Yuchen Jin said inference performance on Ascend 910C achieves 60% of the H100’s performance from developers’ experience. With hand-written CUNN kernels and optimizations, the performance is higher.
DeepSeek’s shift towards Huawei over Nvidia shows that China is choosing more domestic products than foreign ones. Although the country isn’t ready to eliminate Western chips completely. Yet, Huawei’s progress in the chip field shows that it is all set to challenge Nvidia’s foothold in the Chinese AI market.
