DeepGEMM Ascend
An open-source general matrix multiplication acceleration component released by DeepSeek for the Huawei Ascend platform. It is fully compatible with the DeepGEMM API, supports BF16, FP8, and FP4 formats, and targets near-peak hardware performance.
Timeline (last 90 days)
The open-source list includes TileLang, DeepGEMM Ascend, DeepEP Ascend, TileKernels, FlashMLA, and DeepSelect, all hosted on GitHub under the MIT license.
DeepGEMM Ascend was first released on 2026.09.30, supports Ascend 950 devices, is fully compatible with DeepGEMM's API, supports BF16, FP8, and FP4 matrix multiplication, and uses the same package name as the original version.
DeepGEMM Ascend requires the Ascend 950 series, CANN 9.20 toolkit, torch_npu, Python 3.10 or above, and C++20 format support.
The DeepSeek community has open-sourced two underlying acceleration projects targeting the Ascend platform: DeepEP Ascend and DeepGEMM Ascend
The MegaMoE operator places communication and computation into a single operator and pipelines them block by block, using 1 AIC core working in coordination with 2 AIV cores
All Facts (5)
DeepGEMM Ascend requires the Ascend 950 series, CANN 9.20 toolkit, torch_npu, Python 3.10 or above, and C++20 format support.
50%UnverifiedThe MegaMoE operator places communication and computation into a single operator and pipelines them block by block, using 1 AIC core working in coordination with 2 AIV cores
50%UnverifiedThe open-source list includes TileLang, DeepGEMM Ascend, DeepEP Ascend, TileKernels, FlashMLA, and DeepSelect, all hosted on GitHub under the MIT license.
50%UnverifiedDeepGEMM Ascend was first released on 2026.09.30, supports Ascend 950 devices, is fully compatible with DeepGEMM's API, supports BF16, FP8, and FP4 matrix multiplication, and uses the same package name as the original version.
50%UnverifiedThe DeepSeek community has open-sourced two underlying acceleration projects targeting the Ascend platform: DeepEP Ascend and DeepGEMM Ascend
50%