The GitHub repository describes the project as a matrix multiplication kernel library and records an OSS release commit dated September 30, 2026. The supplied release material lists support for FP4, FP8 and BF16 GEMM, grouped GEMM for MoE, MQA logits and fused MegaMoE operators. It reports dense GEMM results of 1,701 TFLOPS for FP4, 861 TFLOPS for FP8 and 431 TFLOPS for BF16, or 98.3%, 99.5% and 99.8% of peak, respectively. Those performance numbers are project-reported and have not been independently verified here. The listed runtime requirements are CANN 9.20, torch_npu and Python 3.10+; the repository specifies the MIT license.
Raw SignalSep 30, 20261 min read1 source
DeepGEMM-Ascend: DeepGEMM-Ascend: clean and efficient matrix multiplication kernel library for Huawei Ascend…
In one minute
DeepSeek published DeepGEMM-Ascend, an open-source port of its matrix-kernel library targeting Huawei Ascend NPUs.