Run2: training loss and games won versus 30×CPU9
Loss is the recorded training loss at exact 50-step intervals, connected by straight lines. CPU9 markers are unsmoothed 30-game waves.
All win-vs-time series
16 native series have 20 one-minute checkpoints; 2 PyTorch-compiled series have 40.
Loss and CPU9 wins by model
left y-axis = loss
·
right y-axis = wins out of 30
· x-axis = post-JIT training time