Skip to content
KernelIndex
Search⌘K

NVFP4 linear layer

19 eligible runs
quantization

FP4-quantized linear layer with block-wise scaling (16 elements per block). Computes y = x @ W^T where both x and W are quantized to FP4 with FP8 E4M3FN block-wise scale factors. Uses torch._scaled_mm for efficient FP4 matrix multiplication.

Fastest reported
87.2µsmean · 91% of SOL
Amir M. Mir | SF TensorLicense unknown · unknown

Reported evidence · last observed 2026-08-20. 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 2026-08-20Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
87.2µs91% SOL
1.00×
Reported · license unknown · no source
2026-08-20

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

no sourcelicense unknownno install recipeRun detail →
2
88.5µs91% SOL
1.02×
Reported · license unknown · no source
2026-08-02

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

no sourcelicense unknownno install recipeRun detail →
3
89.3µs90% SOL
1.02×
Reported · license unknown · no source
2026-07-17

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

no sourcelicense unknownno install recipeRun detail →
4
89.8µs90% SOL
1.03×
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
105.6µs89% SOL
1.21×
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
107.0µs88% SOL
1.23×
Reported · license unknown · no source
2026-08-09

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

no sourcelicense unknownno install recipeRun detail →
7
108.0µs88% SOL
1.24×
Reported · license unknown · no source
2026-08-08

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

no sourcelicense unknownno install recipeRun detail →
8
116.7µs87% SOL
1.34×
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 →
9
117.1µs87% SOL
1.34×
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 →
10
125.5µs86% SOL
1.44×
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 →
11
175.0µs82% SOL
2.01×
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 →
12
177.3µs81% SOL
2.03×
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 →
13
193.3µs79% SOL
2.22×
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
247.3µs75% SOL
2.84×
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
88.5µs
1.02×
Reported
license unknown · no source
unknown
107.0µs
1.23×
Reported
license unknown · no source
unknown
108.0µs
1.24×
Reported
license unknown · no source
unknown
115.2µs
1.32×
Reported
license unknown · no source
unknown
87.2µs
1.00×
Reported
license unknown · no source
unknown
89.8µs
1.03×
Reported
license unknown · no source
unknown
105.6µs
1.21×
Reported
license unknown · no source
unknown
110.5µs
1.27×
Reported
license unknown · no source
Show all 19 implementations ›
unknown
125.5µs
1.44×
Reported
license unknown · no source
unknown
680.1µs
7.80×
Reported
license unknown · no source
unknown
89.3µs
1.02×
Reported
license unknown · no source
unknown
116.7µs
1.34×
Reported
license unknown · no source
unknown
117.1µs
1.34×
Reported
license unknown · no source
unknown
140.4µs
1.61×
Reported
license unknown · no source
unknown
175.0µs
2.01×
Reported
license unknown · no source
unknown
177.3µs
2.03×
Reported
license unknown · no source
unknown
193.3µs
2.22×
Reported
license unknown · no source
unknown
247.3µs
2.84×
Reported
license unknown · no source
unknown
872.7µs
10.01×
Reported
license unknown · no source

Semantics

Inputs and outputs
xbf16 [batch_size, seq_len, in_features]
weight_fp4float4_e2m1fn_x2 [out_features, in_features_half]
weight_scalesfp8_e4m3 [out_features, sf_k]
weight_global_decode_scalefp32 scalar
outputbf16 [batch_size, seq_len, out_features]
Axes and behavior
sf_ksf_k = in_features // sf_vec_size
seq_lenvariable
batch_sizevariable
in_featuresconstant = 8192
sf_vec_sizeconstant = 16
out_featuresconstant = 8192
in_features_halfin_features_half = in_features // 2
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias020_nvfp4_linear_layermodelnova-70b-nvfp4sha2565d28ac0bfa19…
Sources: NVIDIA SOL-ExecBench (2026-08-20)last observed 2026-08-20How records are decidedJSON