DeepGEMM-Ascend
An open-source matrix multiplication kernel library released by DeepSeek for Huawei Ascend chips, ported from the original CUDA-based DeepGEMM. It supports multiple numeric formats including BF16, FP8, and FP4, and encapsulates low-level engineering details such as Ascend fractal layout conversion, operand reuse, and pipelining.
Timeline (last 90 days)
DeepSeek released the Ascend version of DeepGEMM (DeepGEMM-Ascend) on September 30, 2026.
DeepGEMM-Ascend kernels move matrix tiles from global memory into L1 cache and then into L0 cache for computation, and use fractal layout indexing when loading into L1 cache.
The conversion helper functions in the DeepGEMM-Ascend library pack scaling factors into 16-bit integer containers in pairs, arranged in MN-major layout.
The DeepGEMM-Ascend library supports BF16 (16-bit), FP8 (8-bit), and FP4 (4-bit) numeric formats.
DeepGEMM-Ascend provides specific operations such as MQA Logit (handling attention scores in multi-query attention) and MegaMode (serving mixture-of-experts MoE models).
Despite the repository claiming full API compatibility, the DeepGEMM-Ascend implementation explicitly marks the grouped GEMM interface as unimplemented.
DeepGEMM-Ascend was developed and validated based on the Ascend 950 series, requiring CANN 9.20 and Torch NPU.
The performance numbers reported by DeepGEMM-Ascend do not establish a match or victory over NVIDIA systems, nor do they represent whole-model throughput.
DeepGEMM-Ascend's target is explicitly and exclusively Huawei Ascend, and it does not establish a universal software layer applicable to all chip vendors.
All Facts (9)
DeepGEMM-Ascend provides specific operations such as MQA Logit (handling attention scores in multi-query attention) and MegaMode (serving mixture-of-experts MoE models).
60%UnverifiedDeepGEMM-Ascend kernels move matrix tiles from global memory into L1 cache and then into L0 cache for computation, and use fractal layout indexing when loading into L1 cache.
50%UnverifiedThe conversion helper functions in the DeepGEMM-Ascend library pack scaling factors into 16-bit integer containers in pairs, arranged in MN-major layout.
50%UnverifiedThe DeepGEMM-Ascend library supports BF16 (16-bit), FP8 (8-bit), and FP4 (4-bit) numeric formats.
50%UnverifiedDespite the repository claiming full API compatibility, the DeepGEMM-Ascend implementation explicitly marks the grouped GEMM interface as unimplemented.
50%UnverifiedDeepGEMM-Ascend was developed and validated based on the Ascend 950 series, requiring CANN 9.20 and Torch NPU.
50%UnverifiedThe performance numbers reported by DeepGEMM-Ascend do not establish a match or victory over NVIDIA systems, nor do they represent whole-model throughput.
50%UnverifiedDeepGEMM-Ascend's target is explicitly and exclusively Huawei Ascend, and it does not establish a universal software layer applicable to all chip vendors.
50%UnverifiedDeepSeek released the Ascend version of DeepGEMM (DeepGEMM-Ascend) on September 30, 2026.
50%