DeepSeek delays next AI model due to poor performance of Chinese-made chips

DeepSeek, a Chinese AI firm, has announced a delay in the rollout of its anticipated model due to complications encountered while training it with Huawei's chips. This situation underscores the difficulties China faces in its ambition to replace US technology with domestic alternatives. Following the release of its R1 model in January, DeepSeek was urged by government officials to utilize Huawei's Ascend processor instead of relying on Nvidia's systems. However, the startup faced ongoing technical hurdles during the training phase of its R2 model with the Ascend chips. As a result, it opted to use Nvidia chips for training while reserving Huawei's technology for inference tasks. The challenges with the Ascend chips were cited as the primary factor leading to the postponement of the model's launch from May, causing DeepSeek to fall behind its competitors. In training, the model learns from extensive datasets, while inference involves applying the trained model to generate predictions or responses, such as those used in chatbots. This predicament highlights the ongoing gap between Chinese chips and their US counterparts in crucial applications, revealing the obstacles in China's quest for technological independence. Recently, reports indicated that Beijing has requested local tech companies to rationalize their purchases of Nvidia's H20, aiming to encourage a shift toward Huawei and Cambricon alternatives. Insiders have noted that Chinese chips often struggle with issues such as stability, slower inter-chip communication, and less effective software when compared to Nvidia products. Despite efforts, including a team of engineers from Huawei assisting DeepSeek, the startup found it challenging to achieve a successful training session with the Ascend chip.