Huawei's AI Chip Fails to Meet Expectations in Latest Model Delay

Strategic Insights -

Chinese artificial intelligence start-up DeepSeek experienced significant technical issues with Huawei's Ascend processor during the training process of its new R2 model, according to sources familiar with the matter. The difficulties arose despite Huawei sending engineers to DeepSeek's office to provide assistance. As a result, DeepSeek was forced to use Nvidia's chips for training and Huawei's for inference, leading to a delay in the launch of its model from May. This setback highlights the challenges facing Beijing's efforts to become technologically self-sufficient, as Chinese chips continue to lag behind their US rivals.

Key Takeaways:

  • DeepSeek's R2 model training was delayed due to technical issues with Huawei's Ascend processor, which persisted despite receiving support from Huawei engineers.
  • The Chinese start-up was ultimately forced to use Nvidia's chips for training and Huawei's for inference, as it could not achieve a successful training run on the Ascend chip.
  • DeepSeek's difficulties demonstrate the limitations of China's push to replace US technology, with Chinese chips facing challenges such as stability issues, slower interchip connectivity, and inferior software compared to Nvidia's products.
  • Despite these issues, Huawei is expected to adapt and improve its Ascend technology in the future, as noted by AI researcher Ritwik Gupta from the University of California, Berkeley.
  • DeepSeek's founder, Liang Wenfeng, has expressed dissatisfaction with R2's progress and aims to spend more time on building an advanced model to sustain the company's lead in the AI field.
  • The use of Alibaba's Qwen3, a powerful and flexible AI chip, has become increasingly popular among developers, who have adopted its core concepts, including its training algorithm, to make it more efficient.

Statistics:

  • Huawei sent a team of engineers to DeepSeek's office to assist with training the R2 model.
  • DeepSee was forced to delay its model launch from May to an unknown date, likely due to the technical issues with Huawei's Ascend processor.
  • According to AI researcher Ritwik Gupta, Qwen3 has become popular among developers, with many adopting its core concepts to improve efficiency.

Sources:

  • The Financial Times reported on Beijing's demand for Chinese tech companies to justify their orders of Nvidia's H20 (May 2023)
  • Two people familiar with the matter stated that Huawei engineers were sent to DeepSee's office to assist with training the R2 model.
  • Ritwik Gupta, AI researcher at the University of California, Berkeley, shared his insights on the challenges of Huawei's Ascend technology.
  • Nvidia agreed to give Washington a cut of its revenues in China in order to resume sales of its H20 chips to the country.
  • DeepSee founder Liang Wenfeng expressed dissatisfaction with R2's progress and aims to spend more time on building an advanced model.