Huawei provides Ascend chips and software stack, DeepSeek workloads and optimization: According to SemiAnalysis, neither of them would do it alone, but the Ascend ecosystem does not benefit.
The analysis firm SemiAnalysis (Bryan Shan and Dylan Patel) devotes a study to the relationship between Huawei and DeepSeek. His diagnosis: each one brings exactly what is missing to the other. Huawei provides hardware and software infrastructure for its Ascend accelerators; DeepSeek brings real workloads and low-level engineering know-how to make the most of it.
Ascend 950PR: performance close to the theoretical ceiling
On the tested version of the Ascend 950PR (28 Cube cores and 128 Gio of HBM memory), the matrix multiplications in BF16 reach 99.7% of the theoretical ceiling on 1,055 matrix shapes, and the memory bandwidth is close to that of an H100, until it is exceeded for small messages. DeepSeek has published support for Ascend in six of its deposits (TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect): the TileLang compiler brings up to 40% gain on small operations, FlashMLA reorganizes the key-value cache to reduce memory transfers, and DeepEP passes expert parallelism across several bays thanks to Huawei's URMA interconnection. The analysis also details the execution of DeepSeek V4.1-Flash on 32 chips, which uses speculative decoding and, for the first time at DeepSeek, "Engram" dives.
A limited benefit for the rest of the ecosystem
SemiAnalysis temperes however: this partnership does not necessarily facilitate the adoption of Ascend beyond DeepSeek. In vLLM-Ascend, DeepGEMM does not work with a x86 host, some dependencies contain hardcoded ARM headers, and NPU + x86 continuous integration is missing. The rest of the analysis is for pay subscribers only.
Read the full article on the original source
Go to the source →