l123_single_gpt · variant overview
trajectories
250
finished
250
correct
133
compiled
185
avg speedup (correct)
43.87x
sweep-wide eval metricsaveraged over correct kernels
outcomescorrect 133 · compile_fail 65 · incorrect 52
avg speedup43.87x (over 133 correct)
avg memory ratio (kernel / ref)0.81x (133 samples)
avg fusion ratio (ref launches / kernel launches)6.35x (133 samples)
avg energy ratio (ref / kernel)2.00x (114 samples)
avg SOL score (0–1)0.653 (130 samples)
avg DRAM util %57.6%
avg occupancy %142.4%
avg max abs error4.10e-03
modelncorrectcompiled avg speedup
gpt-5.5-priority 250 133 185 43.87x