⚡️ DeepSeek ported DeepGEMM to Huawei Ascend and squeezed 99.8% out of the hardware DeepSeek released DeepGEMM-Ascend, a port of its library of matrix cores on the NPU Huawei Ascend 950. The API remained the same as the original DeepGEMM, and the cores themselves were rewritten for Ascend features. GEMM in FP4, FP8 and BF16, grouped GEMM for MoE, MQA logits and merged MegaMoE statements are supported. Dense GEMM almost hits the theoretical ceiling: FP4 gives 1701 TFLOPS (98.3% of peak), FP8 861 TFLOPS (99.5%), BF16 431 TFLOPS (99.8%). Inside there is JIT compilation, pipelining on coroutines and a thin wrapper over Ascend matrix instructions. Requires CANN 9.20, torch_npu and Python 3.10+. MIT License. https://github.com/deepseek-ai/DeepGEMM-Ascend @ai_machinelearning_big_data