DDR5 Memory Channel Scaling Performance With AMD EPYC 9004 Series

For some of the Intel oneDNN neural network library benchmarks there were was scaling up to 12 memory channels while for the smaller test cases there wasn't a measurable impact.

TensorFlow ResNet-50 performance did benefit going up to 12 memory channels if using a large enough model or batch size.

For Neural Magic's DeepSparse software it depended upon the particular AI model for the benefits to see (or not) from the additional memory channels.

OpenVINO with the models tested generally saw just small performance improvements from the higher memory channels but largely depends upon the models being utilized.