Physix Frontier · News Briefing Card (HuggingFace · Nov 19, 2026)

Study: FLOPs runtime estimates off by hours, new model cuts error to 1.71h

KEY FACTS

  • Dharma-AI trained a runtime prediction model on 272 real SFT tasks.
  • Those tasks consumed roughly 7,382 H200 GPU-hours in total.
  • Traditional FLOPs-based estimation reached only R²=0.31, with a mean absolute error of 12.41 hours.
  • A CatBoost model using only information available before task dispatch reached R²=0.98.
  • That model's mean absolute error is 1.71 hours, with a median error of 0.51 hours.

KEY DATA

272Production SFT tasks
~7,382 GPU-hoursTotal GPU time
12.41 hoursFLOPs estimate mean absolute error
1.71 hoursCatBoost model mean absolute error

PHYSIX OBSERVATION

Scheduling by back-of-the-envelope FLOPs carries errors of a dozen-plus hours, and the idle and queueing costs on shared clusters are badly underestimated. The value of this work is not how complex the model is, but that pre-dispatch information alone can push the error down to the hour level, showing that engineering-side features are worth more investment than theoretical compute. For teams renting GPUs, it is time to change how they schedule.

Source: HuggingFace report