Huawei Report Claims DeepSeek V4 Post-Training on Ascend Chips at 34% MFU — Experts Remain Skeptical on Pre-Training
Summary
- • A July 22 arXiv report from a Huawei-led consortium documents full-parameter post-training of DeepSeek V4 (including V4-Pro, 1.6T params) on Huawei Ascend NPU SuperPOD at 34.22% MFU — a 2.93× efficiency gain over the open-source baseline.
- • Critically, the paper covers post-training only and does NOT establish what hardware trained V4 originally; Tsinghua's Liu Zhiyuan says the model may still have been 'mainly trained on Nvidia hardware.'
- • The Ascend 910C cluster comprised at least 1,000 chips (per Shenzhen government/SCMP), delivering ~60% of H100 inference performance; an earlier Trump admin claim of smuggled Blackwell chips was denied by DeepSeek and called 'far-fetched' by Nvidia.
- • The report arrives as Washington investigates Chinese firms' chip access and Xi Jinping's Washington visit approaches, making contested hardware claims carry significant policy weight beyond their technical content.
Details
Huawei Ascend Post-Training Paper
July 22 arXiv report from Huawei-led consortium; full-parameter post-training of DeepSeek V4 family (V4-Flash and V4-Pro 1.6T) on Huawei Ascend NPU SuperPOD; training reported as stable
34.22% MFU, 2.93× Efficiency Gain
Authors compare optimized Ascend system with open-source baseline recipe; report 34.22% model FLOPs utilization and a 2.93-fold efficiency improvement; both measured internally by the Huawei-led team
Post-Training vs. Pre-Training Scope Gap
Paper explicitly covers post-training only and does not address V4's original pre-training hardware; Tsinghua professor Liu Zhiyuan (MIT Technology Review): model 'may still have been trained mainly on Nvidia hardware'; anonymous sources: Chinese chips better at inference than training
Ascend 910C Chip Specifications
Dual-die accelerator using SMIC second-gen 7nm process and Huawei Da Vinci architecture; cluster of at least 1,000 chips per Shenzhen municipal government/South China Morning Post; returns ~60% of Nvidia H100 inference performance in earlier DeepSeek testing
Competing Hardware Origin Claims
June account (Shenzhen gov/SCMP): 1,000+ Ascend 910C chips; senior Trump admin official alleged V4 trained on smuggled Blackwell chips; DeepSeek denied and cited H800 GPUs + Ascend 910Cs; Nvidia called Blackwell claim 'far-fetched'
June Claim Lacked Benchmarks
Original June announcement carried no benchmarks, run duration, direct Nvidia comparison, or cluster efficiency measure; Tom's Hardware called it part of 'a series of dubious claims'; Wednesday's paper adds MFU metric and baseline comparison
Geopolitical Stakes
Washington is investigating Chinese firms' AI chip access; report lands weeks before Xi Jinping's planned Washington visit; Beijing also weighing its own export controls on AI models — making disputed hardware claims carry significant policy implications
Source: TLDR AI (citing July 22 arXiv report, MIT Technology Review, Tom's Hardware, South China Morning Post, Omdia)
What This Means
The Huawei-led arXiv report supplies the first concrete efficiency benchmarks for Ascend-based post-training of a frontier-class model, partially addressing the credibility gap in the original June announcement — but it deliberately stops short of addressing pre-training, where Nvidia's advantage over Ascend is far larger and where the most consequential chip independence questions reside. Expert consensus from Tsinghua's Liu Zhiyuan and anonymous industry sources is that Chinese AI processors remain more viable for inference and fine-tuning than for frontier pre-training, meaning the headline claim is accurate but significantly narrower in scope than it appears. The geopolitical dimension amplifies the stakes considerably: US chip export controls are predicated on limiting China's AI training capability, and these contested claims will directly shape policy debates as Xi Jinping's Washington visit approaches. The report's selective scope — and the absence of independent external verification — means both skeptics and proponents of Chinese AI chip progress will find support in the same document.