Fused add RMSNorm h5120
10 eligible runs
rmsnorm
Fused Add + RMSNorm with hidden_size=5120 for Qwen3 14B. Epsilon is fixed at 1e-6.
Source baseline · unbeaten
3.30µsmean
flashinfer / wrapper7f6051FlashInfer-Bench baselines · Apache-2.0 · python
Reported evidence · last observed 2026-03-26. The source's designated baseline implementation. Reported by source; not independently reproduced.
Current records
Workloadbatch_size = 1 · bf16batch_size = 6 · bf16batch_size = 32 · bf16batch_size = 37 · bf16batch_size = 1174 · bf16
Not measured on H100 for this workload. Challenges →
Source-native comparison · GPU NVIDIA B200 · Workload batch_size = 32 · bf16 · CUDA 12.8 · Framework pytorch 2.9.1+cu128 · Protocol flashinfer-bench · mean · 2 results · last observed 2026-03-26Record history →
Estimated floor 83 ns · record 39.62× above itestimate, not evidence ›
DRAM 83 ns · 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
13.30µs1.00×Reported · Apache-2.0 · source2026-03-26stale
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 →
23.42µs1.04×Reported · Apache-2.0 · source2026-03-24stale
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 batch_size
Implementations
Implementation
Runtime
Best latency
Evidence
Availability
Semantics
Inputs and outputs
hidden_statesbf16 [batch_size, hidden_size]
residualbf16 [batch_size, hidden_size]
weightbf16 [hidden_size]
outputbf16 [batch_size, hidden_size]
Axes and behavior
batch_sizevariable
hidden_sizeconstant = 5120
determinismunspecified
constraintsNo mutation or aliasing
Identity
Sources: FlashInfer-Bench (2026-03-26) · Apache-2.0last observed 2026-03-26How records are decidedJSON