ETTR · Effective Training Time Ratio
The share of wall-clock time a training job spends making forward progress rather than restarting, checkpointing, or waiting on failures; the honest utilization metric for training clusters.
Current numbers
~2 mincheckpoint-and-restart overhead required to hold ETTR ~0.9 on a 100,000-GPU run at RSC-2-like failure rates
~2 min eachcheckpoint interval and restart overhead each needed to hit ETTR 0.9 on a 100K-GPU run (RSC-2-class failure rate)
0.70 → 0.93modeled ETTR (goodput) for a 16k-GPU run moving from 60-min to 5-min checkpoint interval