Roadmap

What is worth doing next. Nothing here is scheduled.

Fix a crash in the failure simulator. If a second failure lands after a replay but before the next checkpoint, recovery raises an error. The cause is in WorkLedger.invalidate_outcomes (gpu_stack/research/recovery.py) and begin_recovery (gpu_stack/research/recovery_runtime.py).
Fix a mislabeled field. idle_subtracted_gpu_board_energy_j currently holds the raw energy, not the energy with idle power taken out. Compute the real value or rename the field.
Decide what the collective-communication equations are for. Either use them when the calculator estimates training time, or label them clearly as a reference library that the time estimate does not use.
Fix the three economics presets that still look wrong. Facility overhead (PUE) never enters site power. The demand charge is counted twice. A 4-year life is applied to the whole facility instead of to the equipment it fits.
Measure energy on a real GPU with the cumulative energy counter. Sampled power readings are too coarse for the energy questions here.

Navigation Menu