Is the reason DeepSeek-R2 has not been released yet due to technical issues?
According to this information, DeepSeek-R2 is currently being trained using Huawei's Ascend chips, and some problems have been encountered. Although Huawei has sent an engineering team to assist in solving the issues, the Ascend platform has shown insufficient stability, incomplete software and hardware support, and slow communication speed between chips. The engineering team is currently ensuring that AI inference works properly on the Ascend platform.
It is normal to encounter such issues. Switching from NVIDIA's CUDA platform to the Ascend platform requires some compatibility work. After all, DeepSeek-R2 is an upgrade of R1, which was previously trained on NVIDIA chips, with software and hardware optimization and acceleration. Now, switching to a new platform and making modifications is a normal phenomenon. It is similar to how C/C++ language has excellent performance but needs to be adapted for different systems like Windows or Linux. For example, adapting OpenAI's GPT-5 to run on the Ascend platform also requires adaptation, which is nothing surprising.
DeepSeek's adaptation to the Ascend chip is aimed at using domestic chips for AI training/inference, which is a crucial step to break free from foreign constraints. For instance, the H20 chip specially provided by NVIDIA for China has only 15-20% of the performance of the H100 chip. Moreover, it still has a backdoor that can be used to shut down the chip at any time, steal AI technology, or add "dirty data" during training, which would render previous training efforts useless. This is a malicious intention.
Domestic AI chips must become self-reliant; there is no other choice. The current difficulties are temporary, and breakthroughs are just a matter of time.
Original: www.toutiao.com/article/1840433519704074/
Statement: This article represents the views of the author.