Although DeepSeek's hardware facilities have not released details, it is generally believed that a large number of NVIDIA AI chips are used, including H100, H800, H20 and other different models. However, according to the latest exposure, DeepSeek has also verified Huawei's latest AI chip - Ascend 910C.
Ascend 910C was exposed later in 2024. It is said that it has been supplied in batches to some customers, including Alibaba, Baidu, and Tencent. The first batch of about 70,000 units, with an average of only about 20,000 yuan each.
The message shows,Ascend 910C is manufactured using SMIC's 7nm process, with dual chip integrated packaging. The number of transistors reaches 53 billion, and the overall localization rate has reached about 55%.
It can replace NVIDIA H100 and is also used for large-scale AI training and inference. It performs well under different data types such as FP8, FP16, FP32, and FP64.
According to the latest statement,Measured data from the DeepSeek team shows that Huawei’s Ascend 910C performs unexpectedly well in AI inference, reaching about 60% of that of the NVIDIA H100 chip.
Furthermore,Through handwritten CUNN kernel and optimization, the performance of Ascend 910C can be further improved.
It is said thatDeepSeek supports Huawei Ascend chips from day one, and independently maintains the PyTorch warehouse, which can convert CUDA to CUNN with just one line of code. The potential for performance optimization is also huge, and higher performance can be achieved through customized optimization.
However, it should be noted that what is currently known is that the Ascend 910CAI has excellent inference performance, butAI training performance may still be unsatisfactory.
Shengteng 910