Test plan
Contributors
Suggest changes
PDFs
In this validation, we performed image recognition training as specified by MLPerf v2.0. Specifically, we trained the ResNet v2.0 model with the ImageNet dataset until we reached an accuracy of 76.1%. The main metric is the time to reach the desired accuracy. We also report training bandwidth in images per second to better judge scale-out efficiency.
The primary test case evaluated multiple independent training processes (one per node) running concurrently. This simulates the main use case, a shared system used by multiple data scientists. The second test case evaluated scale-out efficiency.