Compare runs
cohort sha256:5a002fb84475ea12… · source-native profile
ranking-v1
no ranks: not comparable
Comparison identity
operation
kernelbench-l3-50-reluselfattention
workload
n_embd = 768 · batch_size = 16 · max_seqlen = 1024 · fp32 · 1862e86b
protocol
KernelBench timing scripts · 547725b0
environment
NVIDIA H100 · 2750ae38
correctness policy
52ff5e65
metric
latency mean (ns)
Context · 10 fields›
architecture
sm_90
CUDA
unknown
driver
unknown
framework
unknown
samples
100
evidence
reported
license
MIT
source
available
status
passed
observed
2026-03-05
No winner: these runs didn't measure the same thing. The rows above show what differs. Why comparable?
Add run ›
Export ›
Markdown CSV JSON
torch.compile (inductor) · sha256:7171da132ecbb… · observed 2026-03-05