SLAI T-Rex Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD paper: https://huggingface.co/papers/2607.20145
This is about CloudMatrix 384 with 910C's anon Still nothing about 950's
DeepSeek-V4 post-training on Ascend SLAI T-Rex achieves 34.22% MFU with 2.93x speedup. Specialized for Operations Research, the model outperforms GPT-5.4-Mini by 4 points.
Many users praise full-parameter DeepSeek-V4 training on Huawei Ascend hardware as a credible Nvidia alternative and useful technical reference, while some doubt the SuperPOD performance numbers hold outside the cluster.
Based on 4 visible X reactions from 9 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@_akhaliq full-parameter post-training on ascend hardware is the part everyone glosses over — the real story isn't the model, it's that huawei just demonstrated a credible alternative to the nvidia stack at scale.
@_akhaliq The ablation on pipeline parallelism granularity is a useful reference for anyone trying to squeeze performance out of restricted VRAM setups or specialized accelerators.
@_akhaliq DeepSeek training V4 fully on Ascend instead of chasing H100s is the geopolitical pivot in one paper
@_akhaliq SuperPOD numbers collapse outside the training cluster
SLAI T-Rex Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD paper: https://huggingface.co/papers/2607.20145
This is about CloudMatrix 384 with 910C's anon Still nothing about 950's
Many users praise full-parameter DeepSeek-V4 training on Huawei Ascend hardware as a credible Nvidia alternative and useful technical reference, while some doubt the SuperPOD performance numbers hold outside the cluster.
Based on 4 visible X reactions from 9 accounts; directional sample.
Ask a question below.
Published answers will appear here.
