#108108th of 242 in ML frameworks and libraries

onednn

Optimized CPU and GPU deep learning primitives library used inside PyTorch and TensorFlow

Stars
4.1k
License
Apache-2.0
Last commit
Nov 2026
Last release
Nov 2026
Language
C++
Stars, last 29 weeks+137
25 Apr 2026: 3,957 stars2 May 2026: 3,960 stars9 May 2026: 3,964 stars16 May 2026: 3,966 stars23 May 2026: 3,967 stars30 May 2026: 3,971 stars6 Jun 2026: 3,981 stars13 Jun 2026: 3,986 stars20 Jun 2026: 3,988 stars27 Jun 2026: 3,992 stars4 Jul 2026: 3,993 stars11 Jul 2026: 3,999 stars18 Jul 2026: 4,005 stars25 Jul 2026: 4,008 stars1 Aug 2026: 4,014 stars8 Aug 2026: 4,020 stars15 Aug 2026: 4,023 stars22 Aug 2026: 4,027 stars29 Aug 2026: 4,033 stars5 Sept 2026: 4,036 stars12 Sept 2026: 4,038 stars19 Sept 2026: 4,043 stars26 Sept 2026: 4,050 stars3 Oct 2026: 4,052 stars10 Oct 2026: 4,057 stars17 Oct 2026: 4,064 stars24 Oct 2026: 4,078 stars31 Oct 2026: 4,083 stars7 Nov 2026: 4,091 stars14 Nov 2026: 4,094 stars

3,957 on 25 Apr 20264,094 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

oneDNN is a C++ library of optimized deep learning operations such as matmul, convolution, attention and normalization for x86-64, AArch64, RISC-V CPUs and Intel GPUs. It detects the ISA at runtime and JIT-generates specialized kernels, with post-op fusion and a Graph API for patterns like SDPA and GQA. The README says most practitioners should use it through a framework (PyTorch, ONNX Runtime, llama.cpp, others) rather than directly.

Who it is for: Framework and runtime developers needing fast CPU/GPU kernels

Strengths

  • Runtime ISA detection with JIT kernels for Intel 64/AMD64, AArch64 and RV64 CPUs
  • Data types from f32 down to f4 and s4, including OCP microscaling formats
  • Graph API fuses larger patterns such as SDPA and GQA; post-ops fuse into matmul and convolution
  • Integrated in PyTorch, TensorFlow, ONNX Runtime, OpenVINO, llama.cpp and others

Weaknesses

  • Low-level C/C++ API; README advises most users to go through a framework instead
  • NVIDIA and AMD GPU backends are experimental; PPC64 and s390x have limited validation
  • GPU engine is off by default and needs a vendor-specific SYCL or OpenCL toolchain
  • Optimized GPU targets listed are Intel only; other vendors are experimental

What it needs

  • GPU optional
  • Needs CMake 3.13+, C++11 compiler, OpenMP or TBB (CPU runtime), SYCL/OpenCL/Level Zero (GPU), Arm Compute Library (optional, AArch64)
  • Models: f32, bf16, f16, f8, f4

What changed

From the release notes. No flag means the note does not say so, not that the upgrade is safe.

Also in ML frameworks and libraries

See all 242
Also in ML frameworks and libraries
RankProjectScore
1transformersModel-definition library for pretrained text, vision, audio and multimodal models169.3k starsApache-2.094 out of 100
2ultralyticsPython library and CLI for training and running YOLO vision models63.5k starsAGPL-3.094 out of 100
3pytorch-image-modelsPyTorch library of image models, pretrained weights and training scripts37.3k starsApache-2.094 out of 100
#4pytorchPython tensor library with GPU support and tape-based autograd105.8k starscustom license93 out of 100
#5supervisionPython toolkit for loading, annotating and counting computer vision detections Live demo ↗ (opens in a new tab)51.7k starsMIT92 out of 100