l5_hw_translation_all_final_gpt · variant overview
trajectories
64
finished
59
correct
10
compiled
32
avg speedup (correct)
2.07x
sweep-wide eval metricsaveraged over correct kernels
outcomesincorrect 22 · compile_fail 17 · correct 10 · error 10 · in_progress 5
avg speedup2.07x (over 10 correct)
avg memory ratio (kernel / ref)1.34x (10 samples)
avg fusion ratio (ref launches / kernel launches)7.07x (10 samples)
avg energy ratio (ref / kernel)— (0 samples)
avg SOL score (0–1)0.293 (8 samples)
avg DRAM util %16.6%
avg occupancy %76.1%
avg max abs error6.25e-03
modelncorrectcompiled avg speedup
FW-GLM-5-1 16 1 9 1.71x
Llama-4-Maverick-17B-128E-Instruct-FP8 16 0 0 0.00x
gpt-5.5-priority 16 8 13 2.33x
grok-4.3 16 1 10 0.31x