Hardware Sensitivity

• End-to-End Training
• FP32 Throughput
• Compute Utilization
• FP32 Utilization
• FP16 Throughput
• Hardware Specifications

End-to-End Training

Before we proceed to actual profiling, we need to verify that the profiled implementation is able to produce the reported validation accuracy. However, training a DNN model to full completion is usually an extremely time-consuming job. Due to time and hardware resource constraints, we only so far produced training curves on the P4000, 1080 Ti, Titan Xp and V100-SXM2 architectures for some models. We will gradually add more training curves.