SemiAnalysis Test: Huawei Ascend Hardware Lags A100 by ~5 Years, DeepSeek Software Near Frontier
SemiAnalysis reports that while Huawei's Ascend 950PR hardware significantly trails NVIDIA's A100, DeepSeek's software optimizations have reached frontier levels, rapidly narrowing the AI chip ecosystem gap through mutual complementarity.
In a report published on October 21, semiconductor research firm SemiAnalysis conducted independent benchmark tests on Huawei's Ascend 950PR chip and analyzed the current state of China's AI software-hardware synergy in light of DeepSeek's recent open-source toolchain updates. The core conclusion is that compared to the NVIDIA A100 baseline, Huawei's hardware performance may lag by more than five years, yet the software optimization tools provided by DeepSeek have already reached the frontier.
SemiAnalysis tested the Ascend 950PR, equipped with 28 cube compute cores and 128 GiB of High Bandwidth Memory (HBM), across 1,055 BF16 matrix multiplication (GEMM) operation shapes. Results showed that Huawei's best-performing kernel achieved up to 99.7% of its theoretical upper limit. This performance was realized using PyTorch's standard torch.matmul function executed by Huawei's own operator library, not DeepSeek's code. This indicates that DeepSeek's extreme optimization in DeepGEMM is not the only way to approach the upper limit.
However, in terms of absolute performance, the Ascend 950PR remains significantly behind. In memory-bandwidth-bound operations, the median ratio of actual achieved performance to theoretically achievable performance for Ascend was 41.5%, compared to 46.8% for the NVIDIA H100. When comparing absolute values, the 950PR falls roughly between NVIDIA's A100, launched in 2020, and AMD's MI300X. This serves as the basis for the report's claim that Huawei's hardware 'lags by more than five years.'
The report highlights a complementary relationship between Huawei and DeepSeek: Huawei possesses a mature hardware ecosystem but lacks a frontier lab model capable of proving its chips can be used for training or inference; DeepSeek has advanced models but faces constraints in accessing compute power. On September 30, DeepSeek released Ascend support for six repositories, including DeepGEMM and FlashMLA. However, at the time of writing, only one of these six tools—the third-party framework TileLang—was actually usable. The other five depended on Huawei's CANN 9.2.0 software suite, which had not yet been officially released, along with specific compiler support.