DeepGEMM Ascend Edition
An open-source BLAS-level GEMM operator library released by DeepSeek, designed specifically for Huawei Ascend NPUs. It supports multi-precision matrix multiplication including FP4, FP8, and BF16, includes MoE-dedicated operators, and is capable of driving hardware utilization close to physical limits.
Timeline (last 90 days)
DeepSeek open-sourced DeepGEMM Ascend edition, a BLAS-level GEMM operator library specifically designed to run on Huawei Ascend hardware.
On the Huawei 950DT chip, DeepGEMM Ascend edition achieves a hardware utilization rate of 99.8% for BF16 multiplied by BF16.
DeepGEMM Ascend edition achieves a hardware utilization rate of 98.3% for FP4 multiplied by FP4.
DeepGEMM Ascend edition supports matrix multiplication in multiple precision combinations including FP4, FP8, and BF16.
DeepGEMM Ascend edition approaches the hardware limit through a thin abstraction layer encapsulating Ascend's underlying matrix interfaces, along with hardware asynchronous data loading and diagonal pipeline tiling.
A key trade-off is that too heavy an abstraction layer becomes a performance killer, while too thin a layer is unfriendly to continuous integration.
Currently, optimization projects for Ascend in frameworks such as vLLM have not yet integrated DeepGEMM.
All Facts (7)
DeepGEMM Ascend edition supports matrix multiplication in multiple precision combinations including FP4, FP8, and BF16.
50%UnverifiedOn the Huawei 950DT chip, DeepGEMM Ascend edition achieves a hardware utilization rate of 99.8% for BF16 multiplied by BF16.
50%UnverifiedDeepGEMM Ascend edition achieves a hardware utilization rate of 98.3% for FP4 multiplied by FP4.
50%UnverifiedCurrently, optimization projects for Ascend in frameworks such as vLLM have not yet integrated DeepGEMM.
50%UnverifiedDeepGEMM Ascend edition approaches the hardware limit through a thin abstraction layer encapsulating Ascend's underlying matrix interfaces, along with hardware asynchronous data loading and diagonal pipeline tiling.
50%UnverifiedA key trade-off is that too heavy an abstraction layer becomes a performance killer, while too thin a layer is unfriendly to continuous integration.
50%UnverifiedDeepSeek open-sourced DeepGEMM Ascend edition, a BLAS-level GEMM operator library specifically designed to run on Huawei Ascend hardware.
50%