A couple of images to think about on top of what I already posted on the AI thread about Ascend-910 + MindSpore being 10% better than A100+PyTorch.
Ascend-910+MindSpore is also 10x as reliable. Operated 25 days reliably vs just 2.8 days reliably for A100+PyTorch
The first picture shows Atlas 900 PoD A2 (supports 8000 cards) vs A100 cluster. Also the Atlas 900 PoD A3 (supports 20000 cards) is comparable to H100 cluster. So in terms of performance, I don't really see any problems here in Huawei product
Look the Atlas Cluster from Huawei can do 9B tokens/day vs just 6.8B tokens/day for A100 DGX cluster. It does GPT-175B test better.
So, if the good folks in Washington do decide to ban A800 exports from Nvidia, this might be a good thing. Keep in mind that basically all the smart city data centers in China are picking Ascend GPUs ahead of Nvidia GPUs. Even the initial version of Baidu Erniebot was trained on Atlas cluster.