DeepSeek's Huawei-Chip Training Claim Finally Gets Its Benchmarks, and Its Doubters

Back to Devoured - July 24, 2026

DEVOURED

DeepSeek's Huawei-Chip Training Claim Finally Gets Its Benchmarks, and Its Doubters

AI infrastructurehardwarellm Implicator

A Huawei-led consortium claims to have achieved 34.22% model FLOPs utilization by training DeepSeek-V4 on Ascend 910B chips, challenging NVIDIA's hardware dominance.

What: Researchers trained the DeepSeek-V4 model using a cluster of Huawei Ascend 910B GPUs, reporting a Model FLOPs Utilization (MFU) of 34.22%. This benchmark aims to validate that Chinese domestic hardware can effectively handle large-scale LLM training, though some industry observers remain skeptical about the reported efficiency and scalability compared to NVIDIA H100 clusters.

Why it matters: The ability to achieve high training efficiency on non-NVIDIA silicon is a critical strategic milestone for organizations facing international GPU export restrictions.

Deep dive

The study utilized a cluster of Huawei Ascend 910B AI processors to fine-tune the DeepSeek-V4 model.
Researchers reported an MFU of 34.22%, a metric representing the ratio of observed performance to peak theoretical compute performance.
The methodology involved full-parameter post-training, which is computationally expensive compared to standard fine-tuning.
The performance data challenges the assumption that large-scale LLM training is tethered to NVIDIA's CUDA-based ecosystem.
Critics have questioned the comparability of these metrics to standard NVIDIA-based training runs given differences in software stacks and networking interconnects.

Decoder

MFU (Model FLOPs Utilization): A standard benchmark that measures the ratio of the actual number of floating-point operations (FLOPs) performed per second to the maximum theoretical FLOPs capacity of the hardware.
Ascend 910B: A high-performance AI accelerator chip developed by Huawei, designed to serve as an alternative to NVIDIA GPUs for large language model training.

Original article

A Huawei-led consortium has released a technical report that documents full-parameter post-training of DeepSeek's V4 family on Huawei Ascend chips, measuring the run at 34.22% model FLOPs utilization.

Explore all 60 articles in today's digest