Intel MKL-DNN Deep Neural Network Library Benchmarks On Xeon & EPYC

This week Intel released MKL-DNN 1.1 as their open-source deep learning library. They also rebranded the software project as the "Deep Neural Network Library" (DNNL) though its focus remains the same. I ran some initial benchmarks on MKL-DNN/DNNL 1.1 on AMD EPYC and Intel Xeon hardware for reference.

We've been running benchmarks of MKL-DNN since learning about it at the start of the year and already used it for Xeon Cascadelake and EPYC Rome benchmarking, but curious about the performance of the new release, some fresh tests were carried out.

On the dual EPYC 7601, EPYC 7742, and Xeon Platinum 8280 servers the MKL-DNN/DNNL 1.1 release was benchmarked using some of the popular DNNL drivers and also varying data type configurations.










With the standard f32 data type, it was exciting to continue to see the EPYC 7742 2P trading blows with the Xeon Platinum 8280 for this Intel-optimized deep learning package. The EPYC 7742 2P performance was a great deal faster than the previous generation EPYC 7601 2P, showing substantial generational performance improvements with Rome. In some of these Intel DNNL tests, the EPYC 7742 outperformed the Xeon Platinum 8280.






When running the tests with u8s8f32 for AVX-512 with the Cascadelake server, the Intel performance to no surprise was at a big advantage... Not really a surprise considering DNNL is Intel-optimized and catered towards AVX-512 / DL-BOOST at that. But at least for the f32 results, fun seeing the Naples to Rome improvement and trading blows with Cascade Lake.

Those wanting to see how their own Linux systems perform with the Intel Deep Neural Network Library "DNNL" can simply run phoronix-test-suite benchmark mkl-dnn to reproduce this Intel software benchmark on your own system(s).