DeepSeek Adds Huawei Ascend Support in Six Repos, SemiAnalysis Says

SemiAnalysis tested an Ascend 950PR_9579 with 28 cube cores and 128 GiB of HBM. It says the 950PR is designed for prefill and recommendation, with less memory bandwidth and fewer cores than the 950DT. On 1,055 BF16 matrix-multiply shapes from open models, it says Huawei's best kernels reach 99.7% of the ceiling. Below Ascend's compute-to-bandwidth ridge, the median achieved fraction of attainable performance was 41.5%, against 46.8% for NVIDIA's H100. SemiAnalysis noted that some H100 runs were power-limited at about 690 to 698 W of a 700 W limit.

In absolute terms the 950PR still sits somewhere between the A100 and AMD's MI300X, and SemiAnalysis says Huawei's hardware may be more than five years behind an A100-era baseline even as the software is at the frontier. On the software side, it says TileLang's AutoSchedule cut execution time by up to 40% for a single operation on the 950PR, and that Huawei published benchmarks for 16-way expert dispatch over its URMA interconnect in the 950DT superpod.

What Is Not Confirmed About DeepSeek on Ascend

The report carries its own warnings. Of the six repositories, only TileLang was usable at the time of writing, and the CANN 9.2.0 toolkit that DeepSeek requires had not yet shipped. SemiAnalysis also says the partnership could work well for DeepSeek without making Ascend equally usable for everyone else: it lists missing DeepGEMM support and other problems on x86 hosts in vLLM-Ascend 0.26.0rc1.

SemiAnalysis quotes a May transcript of a DeepSeek investor meeting, which it labels unverified, in which Liang Wenfeng says four Huawei cards are needed to match one NVIDIA card. Neither Huawei nor DeepSeek is quoted commenting in the reports Voltgrid read, and a performance-per-total-cost-of-ownership comparison is promised in a later SemiAnalysis article. Voltgrid read the free section of the SemiAnalysis piece and the opening of the SCMP report, both of which sit behind paywalls.