SemiAnalysis’s October 21, 2026 analysis reports specific Huawei Ascend kernel benchmarks and examines a profile of DeepSeek-V4.1-Flash running on 32 Ascend 950DT chips. The results describe individual operations and one model workload, not a general measure of DeepSeek’s end-to-end performance on Huawei hardware.
What the 99.7% Ascend benchmark measures
SemiAnalysis tested an Ascend 950PR_9579 with 28 Cube cores and 128 GiB of HBM, or high-bandwidth memory. Across 1,055 BF16 general matrix multiplication (GEMM) shapes, Huawei’s best stock torch.matmul kernels reached 99.7% of the measured compute ceiling.
That percentage applies to the best kernels in an operation-level test. It is not a measure of DeepSeek-V4.1-Flash throughput or of how much of the chip the complete model keeps busy. A fast matrix multiplication is one piece of a deployed AI workload, which also involves work such as moving data and coordinating operations.
How the selected H100 results compare
For selected compute-bound shapes, SemiAnalysis reported the following median achieved-to-attainable fractions. The metric compares achieved compute with the attainable level for those shapes.
| Accelerator | Median achieved-to-attainable fraction on selected compute-bound shapes |
| Huawei Ascend 950PR | 41.5% |
| NVIDIA H100 | 46.8% |
SemiAnalysis noted that some H100 results were power-limited. These medians describe the selected compute-bound shapes, not a universal ranking across AI workloads.
What the DeepSeek-V4.1-Flash profile reveals
The separate profile, published by Huawei and analyzed by SemiAnalysis, covers one rank of DeepSeek-V4.1-Flash running on 32 Ascend 950DT chips. In its steady-state window, Stream 19 was occupied for 88.8% of the time and accounted for 72.6% of summed device time. Those percentages use different measures: work can overlap across streams, so occupancy and share of summed device time are not interchangeable.
SemiAnalysis also described a 1.32× step-improvement scenario if Cube throughput doubled while measured stream overlap and idle time stayed fixed. That figure is conditional on those assumptions, rather than a measured speedup.
How the software fits the hardware
DeepSeek released Ascend support across TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect on September 30, 2026, according to SemiAnalysis. In its October 21 analysis, SemiAnalysis said TileLang was the only one of the six repositories usable at its writing cutoff. NeoTeo covered the software-support announcement in its earlier report.
TileLang’s AutoSchedule reportedly cut execution time by up to 40% for a single operation on Ascend 950PR, especially at smaller sizes. That result, too, is operation-level. SemiAnalysis identified CANN 9.2.0 and a C++20 compiler as dependencies for DeepGEMM-Ascend. Cloud Codes describes DeepGEMM-Ascend as a specialized kernel library, while CANN is Huawei’s broader software platform.
Cloud Codes’ tutorial explains why Ascend kernels use hardware-specific memory layouts and covers packing FP8 scale factors into 16-bit integer containers. It also discusses a separate 99.8% kernel-ceiling result under cold-L2 profiling, a different benchmark description from SemiAnalysis’s 99.7% result.
For x86 hosts, SemiAnalysis reported limitations in vLLM-Ascend_0.26.0rc1: missing DeepGEMM support, a startup blocker caused by hard-coded ARM headers in a dependency, CPU binding skipped on non-ARM CPUs, and no NPU-plus-x86 continuous-integration coverage.
