DeepGEMM
DeepGEMM is a high-performance General Matrix Multiplication (GEMM) library developed by DeepSeek, designed for deep learning model inference and training scenarios. It is deeply optimized for GPU hardware architectures and supports low-precision data types such as FP8, enabling significant improvements in computational efficiency for large-scale matrix operations. DeepGEMM is a representative implementation of a low-level operator library in AI infrastructure.
Core Facts
DeepGEMM is a high-performance matrix multiplication library designed by DeepSeek specifically for its MoE models, supporting low-precision formats such as bf16 and fp8.
On September 30, 2025, DeepSeek open-sourced a batch of underlying infrastructure components, including operators, a programming language, compute libraries, and communication libraries
Timeline (last 90 days)
In DeepInfer, Ascend and NVIDIA are first-class citizens, with 30 TileLang operators on each side corresponding one-to-one, plus Ascend versions of DeepGEMM and DeepEP.
v5.20.1 fixes the DeepGEMM Mega MoE expert parallelism issue (#49428).
For production environments that depend on a specific torch version or use MoE and multimodal inference, the HunYuan-VL and DeepGEMM related paths should be verified with priority before upgrading.
DeepGEMM is a GPU matrix operation (BLAS) kernel library open-sourced by the DeepSeek team, written in CUDA.
DeepGEMM has garnered approximately 8,525 stars and 1,361 forks on GitHub, and continues to gain popularity at a rate of 363 new stars added per day.
The core positioning of DeepGEMM is "clean and efficient," with readability and efficiency listed side by side as design goals.
Low-level kernel libraries have a high barrier to entry and are typically aimed at engineers and researchers with CUDA development experience, while ordinary application developers are more likely to benefit indirectly through higher-level frameworks.
DeepGEMM corresponds to cuBLAS, responsible for optimizing the matrix multiplication calculations that are core to AI.
The original DeepGEMM is a CUDA kernel library, and the Ascend version provides a relatively familiar interface for Ascend while handling layout and data movement.
The six components have distinct roles: DeepGEMM accelerates general matrix operations, DeepEP handles large-scale cross-device communication, TileKernels provides common vector computation and memory access operators, FlashMLA provides sparse attention operators, and DeepSelect performs efficient data selection.
3 more timeline events
All Facts (13)
DeepGEMM is a high-performance matrix multiplication library designed by DeepSeek specifically for its MoE models, supporting low-precision formats such as bf16 and fp8.
65%VerifiedOn September 30, 2025, DeepSeek open-sourced a batch of underlying infrastructure components, including operators, a programming language, compute libraries, and communication libraries
65%UnverifiedIn DeepInfer, Ascend and NVIDIA are first-class citizens, with 30 TileLang operators on each side corresponding one-to-one, plus Ascend versions of DeepGEMM and DeepEP.
50%UnverifiedFor production environments that depend on a specific torch version or use MoE and multimodal inference, the HunYuan-VL and DeepGEMM related paths should be verified with priority before upgrading.
50%Unverifiedv5.20.1 fixes the DeepGEMM Mega MoE expert parallelism issue (#49428).
50%UnverifiedDeepGEMM is a GPU matrix operation (BLAS) kernel library open-sourced by the DeepSeek team, written in CUDA.
50%UnverifiedThe core positioning of DeepGEMM is "clean and efficient," with readability and efficiency listed side by side as design goals.
50%UnverifiedLow-level kernel libraries have a high barrier to entry and are typically aimed at engineers and researchers with CUDA development experience, while ordinary application developers are more likely to benefit indirectly through higher-level frameworks.
50%UnverifiedDeepGEMM corresponds to cuBLAS, responsible for optimizing the matrix multiplication calculations that are core to AI.
50%UnverifiedThe original DeepGEMM is a CUDA kernel library, and the Ascend version provides a relatively familiar interface for Ascend while handling layout and data movement.
50%UnverifiedThe six components have distinct roles: DeepGEMM accelerates general matrix operations, DeepEP handles large-scale cross-device communication, TileKernels provides common vector computation and memory access operators, FlashMLA provides sparse attention operators, and DeepSelect performs efficient data selection.
50%UnverifiedThe Ascend version of DeepGEMM retains the original API, is developed and validated on the Ascend 950 series, and requires the corresponding CANN toolchain
50%UnverifiedKimi K3 supports simultaneous provision of both Python and Rust frontend support, and integrates DeepGEMM, compressed-tensors quantization checkpoints, and DSpark AR fusion.
50%