Blog

One 300-step ResNet-18 run on a T4, observed three ways. In that TraceML 0.3.4 run, whole-process wall time increased 2.0% and the summary named the bottleneck in one line. torch.profiler wrote a 70.7 MB trace you read yourself, and cProfile named the JPEG decode but could not see the idle GPU.

I moved the same BERT-large training recipe across an L40S, A100 SXM, and H100 SXM on RunPod. The raw throughput ranking was expected. The useful part was the TraceML fingerprint: input wait and H2D copy were negligible, all three runs were compute-bound, and the H100 result pointed to a simple batch-size follow-up.