MLPerf
The industry-standard benchmark suite for AI training and inference performance; the closest thing to apples-to-apples numbers across accelerator vendors.
Current numbers
Unet3D ≥90%, RetinaNet ≥85%accelerator-utilization thresholds for a valid MLPerf Storage v3.0 training result, by workload; holding this line is the loader's core objective
~every 30 minMLPerf Storage model for 100k accelerators at full utilization: ~one failure per 30 min; illustrative checkpoint workload, not an observed fleet rate