Case studies

These measured investigations show how TraceML observations relate to wall-clock training behavior. The complete write-ups and reproduction code stay with the executable examples in the TraceML repository.

Measured investigations

ResNet-18 input pipeline

A single-T4 run was input-bound while JPEG decoding happened synchronously. Changing only the input pipeline reduced wall-clock step time by 43.8% and changed the TraceML verdict from input-bound to compute-bound.

Read the complete case study and reproduce it

RF-DETR Nano training

On real COCO batches, one T4 sustained about 18.6 images/s and four-T4 DDP sustained about 60.5 images/s, or 3.26x weak scaling. TraceML showed that input waiting remained small while backward time increased under DDP.

Read the complete case study and reproduce it

RF-DETR non-JPEG release regression

RF-DETR 1.11.0 increased median native step time by 14.78% on a controlled large-PNG workload. TraceML observed a 23.74% increase in input wait while measured GPU compute became faster. RF-DETR 1.11.1 restored step time and input wait to within 0.5% of the 1.10.1 baseline.

Read the complete case study and reproduce it

Reproduction packages

The LeRobot v3 image-loading regression package provides pinned before-and-after revisions, a runnable ACT workload and an analysis notebook. It does not yet publish a measured headline result.