oneDNN

About

• DNN performance primitives library, with C/C++ API
• Optimized for Intel hardware, with runtime detection of ISA and JIT code-gen
• https://github.com/oneapi-src/oneDNN
• Apache 2.0 license.
• Formerly MKL-DNN and DNNL
• See oneDNN-AArch64 for 64-bit Arm support

Version

• See oneDNN Versions

Relationship with Intel MKL

Version	Relevance
0.21.x	Could use Intel MKL for faster GEMMs on older platforms.
1.x	Onwards - Does not use MKL anymore - MKLDNN_USE_MKL is not applicable anymore.

Prebuilt containers with frameworks and oneDNN

Containers	Versions
Intel Docker Hub images	TF 2.8.0

See Also

• cuDNN

Resources

• Best performance flags for oneDNN - https://software.intel.com/content/www/us/en/develop/articles/maximize-tensorflow-performance-on-cpu-considerations-and-recommendations-for-inference.html