The Financial Times reported that DeepSeek was forced to delay the release of its new R2 large language model due to performance problems with Huawei Ascend processors, “highlighting the limits of Beijing’s push to replace US technology,” according to the FT.
The publication cited three sources who said Chinese government officials encouraged DeepSeek to use Ascend chips instead of NVIDIA processors after the release early this year of DeepSeek’s R1 model, which rocked the AI world.
“But the Chinese start-up encountered persistent technical issues during its R2 training process using Ascend chips,” the FT reported, “prompting it to use NVIDIA chips for training and Huawei’s for inference, said the (sources).”
The news comes amidst other reports that NVIDIA and AMD GPUs – the NVIDIA H20 and AMD MI308 – currently allowed for export to China after a recent US export rule change, are intended to deliver roughly similar power as China’s homegrown GPUs. Also, there are reports that Chinese authorities are asking companies planning to justify their use of American GPUs.
The FT article cited sources who said the Huawei processors have stability problems, “slower inter-chip connectivity and inferior software compared with Nvidia’s products.”
In an attempt to resolve the problems, a team of Huawei technicians worked with DeepSeek on site, but that a successful training run could not be achieved on the Ascend processors, sources said.
The FT said DeepSeek’s R2 model could be released “in the coming weeks.”
