Skip to content
KernelIndex
Search⌘K

NVFP4 multimodal embedding projection

22 eligible runs
quantization

FP4-quantized multimodal embedding projection that merges text embeddings with vision features. Takes text token embeddings and vision token embeddings, applies FP4-quantized linear projections to align dimensions, and merges them based on vision token mask positions. Uses block-wise FP4 quantization (16 elements per block) for the projection weights.

Fastest reported
92.4µsmean · 85% of SOL
Amir M. Mir | SF TensorLicense unknown · unknown

Reported evidence · last observed 2026-08-14. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyvocab_size = 151936 · hidden_size = 8192 · sf_vec_size = 16 · vision_hidden_size = 3584 · num_vision_scale_blocks = 224 · vision_hidden_size_half = 179216 cases

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload suite of 16 cases · mean latency · Protocol SOL-ExecBench evaluation stack v1.1 · mean · 16 results · last observed 2026-08-23Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
92.4µs85% SOL
1.00×
Reported · license unknown · no source
2026-08-14

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

no sourcelicense unknownno install recipeRun detail →
2
96.8µs84% SOL
1.05×
Reported · license unknown · no source
2026-07-31

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

no sourcelicense unknownno install recipeRun detail →
3
98.0µs84% SOL
1.06×
Reported · license unknown · no source
2026-07-14

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

no sourcelicense unknownno install recipeRun detail →
4
103.8µs83% SOL
1.12×
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 →
5
104.1µs83% SOL
1.13×
Reported · license unknown · no source
2026-08-23

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

no sourcelicense unknownno install recipeRun detail →
6
106.4µs82% SOL
1.15×
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 →
7
106.4µs82% SOL
1.15×
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 →
8
106.5µs82% SOL
1.15×
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 →
9
109.9µs82% SOL
1.19×
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 →
10
114.0µs81% SOL
1.23×
Reported · license unknown · no source
2026-07-31

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

no sourcelicense unknownno install recipeRun detail →
11
118.1µs81% SOL
1.28×
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 →
12
128.5µs79% SOL
1.39×
Reported · license unknown · no source
2026-07-14

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

no sourcelicense unknownno install recipeRun detail →
13
196.1µs70% SOL
2.12×
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 →
14
290.2µs61% SOL
3.14×
Reported · license unknown · no source
2026-07-23

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

no sourcelicense unknownno install recipeRun detail →
15
291.8µs61% SOL
3.16×
Reported · license unknown · no source
2026-07-25

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

no sourcelicense unknownno install recipeRun detail →
16
439.2µs50% SOL
4.75×
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
92.4µs
1.00×
Reported
license unknown · no source
unknown
96.8µs
1.05×
Reported
license unknown · no source
unknown
97.5µs
1.06×
Reported
license unknown · no source
unknown
103.8µs
1.12×
Reported
license unknown · no source
unknown
104.1µs
1.13×
Reported
license unknown · no source
unknown
106.4µs
1.15×
Reported
license unknown · no source
unknown
106.5µs
1.15×
Reported
license unknown · no source
unknown
114.0µs
1.23×
Reported
license unknown · no source
Show all 22 implementations ›
unknown
118.6µs
1.28×
Reported
license unknown · no source
unknown
384.7µs
4.16×
Reported
license unknown · no source
unknown
98.0µs
1.06×
Reported
license unknown · no source
unknown
106.4µs
1.15×
Reported
license unknown · no source
unknown
109.9µs
1.19×
Reported
license unknown · no source
unknown
118.1µs
1.28×
Reported
license unknown · no source
unknown
128.5µs
1.39×
Reported
license unknown · no source
unknown
136.0µs
1.47×
Reported
license unknown · no source
unknown
193.7µs
2.10×
Reported
license unknown · no source
unknown
196.1µs
2.12×
Reported
license unknown · no source
unknown
290.2µs
3.14×
Reported
license unknown · no source
unknown
291.8µs
3.16×
Reported
license unknown · no source
unknown
439.2µs
4.75×
Reported
license unknown · no source
unknown
1.35ms
14.63×
Reported
license unknown · no source

Semantics

Inputs and outputs
text_input_idsint64 [batch_size, seq_len]
text_embedding_weightbf16 [vocab_size, hidden_size]
vision_featuresfp32 [batch_size, num_vision_tokens, vision_hidden_size]
vision_proj_weight_fp4float4_e2m1fn_x2 [1, hidden_size, vision_hidden_size_half]
vision_proj_weight_scalefp8_e4m3 [1, hidden_size, num_vision_scale_blocks]
vision_proj_weight_global_scalefp32 scalar
vision_proj_biasbf16 [hidden_size]
vision_token_maskbool [batch_size, seq_len]
merged_embeddingsbf16 [batch_size, seq_len, hidden_size]
Axes and behavior
seq_lenvariable
batch_sizevariable
vocab_sizeconstant = 151936
hidden_sizeconstant = 8192
sf_vec_sizeconstant = 16
num_vision_tokensvariable
vision_hidden_sizeconstant = 3584
num_vision_scale_blocksnum_vision_scale_blocks = vision_hidden_size // sf_vec_size
vision_hidden_size_halfvision_hidden_size_half = vision_hidden_size // 2
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias008_nvfp4_multimodal_embedding_projectionmodelqwen3-vl-32b-instruct-nvfp4sha25632e7d3f31c9a…
Sources: NVIDIA SOL-ExecBench (2026-08-23)last observed 2026-08-23How records are decidedJSON