AMD EPYC 7502 + EPYC 7742 Linux Performance Benchmarks

MKL-DNN for the uninitiated is a software package out of Intel themselves and making use of their Math Kernel Library for deep neural networks. As one would expect, that Intel optimized package easily performs the best on their AVX-512 Xeon Scalable processors... But not always.

For AlexNet and GoogleNet convolution, the EPYC Rome processors were now giving Intel Xeon Scalable a competitive race for this Intel optimized software stack.

Intel also develops the open-source Streaming Video Technology (SVT) video encoders which was fun to see how well they would perform with EPYC. The SVT-AV1 and SVT-HEVC video encoders dominated on Rome with the greater core/thread counts. With the SVT-VP9 encoder it doesn't scale as well so in that case a single Xeon Platinum 8280 with its higher clock speeds came out ahead.

In the neutral x264 and x265 projects, the AMD Rome processors were performing great and a big step up from the previous generation EPYC processors.

The neutral dav1d AV1 decoder also was performing very well with Rome where even a single EPYC 7502 was beating out the Xeon Platinum 8280.