Skip to content
KernelIndex
Search⌘K

NVFP4 attention output projection

20 eligible runs
quantization

FP4-quantized attention output projection (o_proj) that transforms concatenated multi-head attention outputs back to hidden dimension. Uses FP4 block-scaled GEMM with 16 elements per FP8 scale factor. Computes C = (A * scale_a * global_a) @ (B^T * scale_b * global_b) where A is the input activation and B is the weight matrix.

Fastest reported
62.2µsmean · 82% of SOL
Amir M. Mir | SF TensorLicense unknown · unknown

Reported evidence · last observed 2026-07-21. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Source-native comparison · GPU NVIDIA B200 · Workload suite of 16 cases · mean latency · Protocol SOL-ExecBench evaluation stack v1.1 · mean · 14 results · last observed 2026-08-11Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
62.2µs82% SOL
1.00×
Reported · license unknown · no source
2026-07-21

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
2
66.5µs82% SOL
1.07×
Reported · license unknown · no source
2026-08-01

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
3
72.8µs80% SOL
1.17×
Reported · license unknown · no source
2026-07-18

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
4
75.3µs79% SOL
1.21×
Reported · license unknown · no source
2026-07-13

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
5
75.7µs79% SOL
1.22×
Reported · license unknown · no source
2026-07-16

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
6
76.0µs79% SOL
1.22×
Reported · license unknown · no source
2026-08-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
7
90.4µs75% SOL
1.45×
Reported · license unknown · no source
2026-08-05

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
8
122.0µs69% SOL
1.96×
Reported · license unknown · no source
2026-07-13

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
9
129.2µs68% SOL
2.08×
Reported · license unknown · no source
2026-08-03

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
10
141.7µs65% SOL
2.28×
Reported · license unknown · no source
2026-07-13

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
11
146.6µs65% SOL
2.36×
Reported · license unknown · no source
2026-08-11

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
12
160.9µs62% SOL
2.59×
Reported · license unknown · no source
2026-07-22

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
13
168.5µs61% SOL
2.71×
Reported · license unknown · no source
2026-07-22

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
14
458.8µs36% SOL
7.38×
Reported · license unknown · no source
2026-07-22

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
unknown
62.2µs
1.00×
Reported
license unknown · no source
unknown
66.5µs
1.07×
Reported
license unknown · no source
unknown
66.0µs
1.06×
Reported
license unknown · no source
unknown
72.7µs
1.17×
Reported
license unknown · no source
unknown
72.8µs
1.17×
Reported
license unknown · no source
unknown
76.0µs
1.22×
Reported
license unknown · no source
unknown
75.7µs
1.22×
Reported
license unknown · no source
unknown
136.5µs
2.20×
Reported
license unknown · no source
Show all 20 implementations ›
unknown
75.3µs
1.21×
Reported
license unknown · no source
unknown
90.4µs
1.45×
Reported
license unknown · no source
unknown
99.0µs
1.59×
Reported
license unknown · no source
unknown
119.6µs
1.92×
Reported
license unknown · no source
unknown
122.0µs
1.96×
Reported
license unknown · no source
unknown
129.2µs
2.08×
Reported
license unknown · no source
unknown
146.6µs
2.36×
Reported
license unknown · no source
unknown
141.7µs
2.28×
Reported
license unknown · no source
unknown
160.9µs
2.59×
Reported
license unknown · no source
unknown
168.5µs
2.71×
Reported
license unknown · no source
unknown
458.8µs
7.38×
Reported
license unknown · no source
unknown
543.9µs
8.75×
Reported
license unknown · no source

Semantics

Inputs and outputs
hidden_statesfp32 [m, input_dim]
weight_fp4float4_e2m1fn_x2 [hidden_size_padded, 2048]
weight_scale_fp8fp8_e4m3 [hidden_size_padded, sf_k]
weight_global_decode_scalefp32 scalar
outputfp32 [m, hidden_size_padded]
Axes and behavior
bvariable
lvariable
mm = b * l
sf_ksf_k = input_dim // sf_vec_size
input_dimconstant = 4096
hidden_sizeconstant = 2688
sf_vec_sizeconstant = 16
hidden_size_paddedconstant = 2816
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias024_nvfp4_attention_output_projectionmodelnvidia-nemotron-3-nano-30b-a3b-nvfp4sha256d23c6fe94260…
Sources: NVIDIA SOL-ExecBench (2026-08-11)last observed 2026-08-11How records are decidedJSON