Abstract:
Some media learned from people familiar with the industry chain that DeepSeek's next-generation large model, DeepSeek V5, plans to use Huawei's Ascend 950 chip as the core training computing power, and the project is scheduled to be launched in 2027. It is reported that the core R&D docking team of DeepSeek and Huawei Ascend is currently based in Hangzhou. Both parties have completed a large number of underlying operator adaptation and customized development in the V4 stage, laying the foundation for subsequent model training migration and computing power collaboration.
From the perspective of the cooperation path, in the previous generation of DeepSeek models, Huawei's Ascend computing power was mainly responsible for inference tasks and has not yet entered the model training process. The V5 plan this time will introduce Ascend 950 into the core training system, which means that the cooperation between the two parties is expected to further extend from the inference end to the training end. Compared with inference tasks, large model training places higher requirements on chip computing power, communication interconnection, software ecology and cluster stability.
The industry believes that if DeepSeek V5 finally implements native training based on Ascend 950, it will not only be expected to further verify the performance and adaptability of domestic AI chips in large-scale model training scenarios, but may also bring new commercialization opportunities to the domestic computing power industry chain. However, the relevant cooperation is still reported by the media citing people in the industry chain, and the specific technical plan and implementation progress have yet to be further confirmed by the official.
Comments