GroupNorm
42 eligible runs
groupnorm
Group normalization of a batch×channels×hidden tensor with affine channel weight and bias; groups = channels / channels_per_group.
Fastest reported · usable today
17.5µsmedian · 1.75× faster than baseline
LigerLiger-Kernel · BSD-2-Clause · triton
Reported evidence · last observed 2026-02-28. Reported by source; not independently reproduced.
pip install "liger-kernel==0.7.0"Fastest by languagetriton · Liger · 17.5 µs · #1python · Transformers · 30.7 µs · 1.75×
Current records
Workloadbatch = 128 · hidden = 512 · channels = 32 · channels_per_group = 4 · fp32batch = 128 · hidden = 512 · channels = 64 · channels_per_group = 4 · fp32batch = 128 · hidden = 512 · channels = 128 · channels_per_group = 4 · fp32batch = 128 · hidden = 512 · channels = 2048 · channels_per_group = 4 · fp32batch = 128 · hidden = 512 · channels = 256 · channels_per_group = 4 · fp32batch = 128 · hidden = 512 · channels = 512 · channels_per_group = 4 · fp32batch = 128 · hidden = 512 · channels = 1024 · channels_per_group = 4 · fp32
HardwareBest knownImplementationRuns
NVIDIA B200 · env 117.5 µsLiger2NVIDIA B200 · env 255.3 µsTransformers2NVIDIA B200 · env 384.0 µsTransformers2Not measured on H100 for this workload. Challenges →
Source-native comparison · GPU NVIDIA B200 · Workload batch = 128 · hidden = 512 · channels = 32 · channels_per_group = 4 · fp32 · Protocol Liger-Kernel benchmark scripts · median · 2 results · last observed 2026-02-28Record history →
Estimated floor 1.05 µs · record 16.72× above itestimate, not evidence ›
DRAM 1.05 µs · bandwidth-bound on B200
every declared tensor crosses HBM exactly once (8,000 GB/s, B200 datasheet)
no arithmetic formula for this family: bandwidth floor only
headroom-v1: a lower bound from declared tensors and datasheet peaks. A kernel can sit well above it for good reasons.
#
Implementation
Latency
vs #1
Trust
Observed
1LigerLiger-Kernel17.5µs1.00×Reported · BSD-2-Clause · source2026-02-28stale
1.75× faster than the baseline. Measured exactly what you asked. Reported by source; not independently reproduced.
source mirroredBSD-2-ClauseinstallableView source →Run detail →
230.7µs1.75×Reported · Apache-2.0 · source2026-02-28stale
Measured exactly what you asked. The source's designated baseline implementation. Reported by source; not independently reproduced.
source mirroredApache-2.0no install recipeView source →Run detail →
Scaling by channels
Implementations
Implementation
Runtime
Best latency
Evidence
Availability
Semantics
Inputs and outputs
xfloat [batch, channels, hidden]
weightfloat [channels]
biasfloat [channels]
yfloat [batch, channels, hidden]
Axes and behavior
batchvariable
groupsgroups = channels // channels_per_group
hiddenvariable
channelsvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
sha256a5be9b2fb0c7…
Sources: Liger-Kernel benchmarks (2026-02-28) · BSD-2-Clauselast observed 2026-02-28How records are decidedJSON