#DeepSeek United Huawei, and moved an entire set of AI training and inference low-level components onto Ascend—operator development, matrix computation, cross-card communication, Attention, and data filtering, basically matching the stack it previously did for NVIDIA GPU.
The key is TileLang: in V4 training, a large number of operators have been implemented with it. V4 has been adapted to Ascend 950, and Huawei claims its own chips participated in part of the training.
Huang Renxun said back in April that if DeepSeek were to release first on Huawei chips, it would be a “bad outcome” for the United States—what he’s worried about is that the model begins optimizing around non-U.S. hardware.
The key is TileLang: in V4 training, a large number of operators have been implemented with it. V4 has been adapted to Ascend 950, and Huawei claims its own chips participated in part of the training.
Huang Renxun said back in April that if DeepSeek were to release first on Huawei chips, it would be a “bad outcome” for the United States—what he’s worried about is that the model begins optimizing around non-U.S. hardware.