NVFP4 MoE router projection

33 eligible runs

moe

FP4-quantized MoE router projection that computes expert selection scores. Projects hidden states from hidden_size (2688) to n_routed_experts (128) using FP4 weights with block-scaled quantization, followed by sigmoid activation, bias correction, and top-k selection with normalized routing weights.

Fastest reported

22.7µsmean · 95% of SOL

Hollow JellyfishLicense unknown · unknown

Reported evidence · last observed 2026-08-22. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencysf_k = 168 · n_group = 1 · topk_group = 1 · hidden_size = 2688 · sf_vec_size = 16 · hidden_size_half = 1344 · n_routed_experts = 128 · num_experts_per_tok = 616 cases

num_tokens

dtype

384fp4/fp32/fp8_e4m3 640fp4/fp32/fp8_e4m3 1280fp4/fp32/fp8_e4m3 1408fp4/fp32/fp8_e4m3 1536fp4/fp32/fp8_e4m3 1920fp4/fp32/fp8_e4m3 2048fp4/fp32/fp8_e4m3 2304fp4/fp32/fp8_e4m3 2432fp4/fp32/fp8_e4m3 2560fp4/fp32/fp8_e4m3 3072fp4/fp32/fp8_e4m3 3200fp4/fp32/fp8_e4m3 3456fp4/fp32/fp8_e4m3 3584fp4/fp32/fp8_e4m3 3840fp4/fp32/fp8_e4m3 3968fp4/fp32/fp8_e4m3

scrolls · 16 cases total

Aggregatessuite of 16 cases · mean latency

HardwareBest knownImplementationRuns

NVIDIA B200 · env 122.7 µsHollow Jellyfish27 NVIDIA B200 · env 225.6 µsRecursive6

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 · 27 results · last observed 2026-08-22Record history →

#

Implementation

Latency

vs #1

Trust

Observed

1

Hollow Jellyfish

22.7µs95% SOL

1.00×

Reported · license unknown · no source

2026-08-22

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

no sourcelicense unknownno install recipeRun detail →

2

23.7µs95% SOL

1.05×

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

Amir M. Mir | SF Tensor

24.1µs95% SOL

1.06×

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 →

4

26.9µs94% SOL

1.19×

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

30.6µs93% SOL

1.35×

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 →

6

30.6µs93% SOL

1.35×

Reported · license unknown · no source

2026-08-17

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

no sourcelicense unknownno install recipeRun detail →

7

32.4µs92% SOL

1.43×

Reported · license unknown · no source

2026-07-15

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

no sourcelicense unknownno install recipeRun detail →

8

34.0µs92% SOL

1.50×

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

36.5µs91% SOL

1.61×

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

38.4µs90% SOL

1.69×

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 →

11

39.3µs90% SOL

1.73×

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

41.3µs89% SOL

1.82×

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 →

13

44.4µs89% SOL

1.96×

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 →

14

46.5µs88% SOL

2.05×

Reported · license unknown · no source

2026-08-04

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

no sourcelicense unknownno install recipeRun detail →

15

46.5µs88% SOL

2.05×

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 →

16

47.3µs88% 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 →

17

50.9µs87% SOL

2.25×

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 →

18

61.1µs84% SOL

2.69×

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 →

19

Amir M. Mir | SF Tensor

61.2µs85% SOL

2.70×

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 →

20

98.5µs76% SOL

4.34×

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 →

21

99.0µs76% SOL

4.36×

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 →

22

101.9µs76% SOL

4.49×

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 →

23

110.5µs74% SOL

4.87×

Reported · license unknown · no source

2026-08-06

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

no sourcelicense unknownno install recipeRun detail →

24

148.7µs68% SOL

6.56×

Reported · license unknown · no source

2026-07-26

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

no sourcelicense unknownno install recipeRun detail →

25

222.7µs58% SOL

9.81×

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 →

26

443.3µs41% SOL

19.54×

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 →

27

587.0µs34% SOL

25.87×

Reported · license unknown · no source

2026-08-13

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

23.7µs

1.05×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

24.1µs

1.06×

Reported

license unknown · no source

unknown

30.6µs

1.35×

Reported

license unknown · no source

Hollow Jellyfish

unknown

22.7µs

1.00×

Reported

license unknown · no source

unknown

25.6µs

1.13×

Reported

license unknown · no source

unknown

32.4µs

1.43×

Reported

license unknown · no source

unknown

34.0µs

1.50×

Reported

license unknown · no source

unknown

36.3µs

1.60×

Reported

license unknown · no source

Show all 33 implementations ›

unknown

96.7µs

4.26×

Reported

license unknown · no source

unknown

26.9µs

1.19×

Reported

license unknown · no source

unknown

30.6µs

1.35×

Reported

license unknown · no source

unknown

36.5µs

1.61×

Reported

license unknown · no source

unknown

38.4µs

1.69×

Reported

license unknown · no source

unknown

39.3µs

1.73×

Reported

license unknown · no source

unknown

41.3µs

1.82×

Reported

license unknown · no source

unknown

44.4µs

1.96×

Reported

license unknown · no source

unknown

46.5µs

2.05×

Reported

license unknown · no source

unknown

46.5µs

2.05×

Reported

license unknown · no source

unknown

47.3µs

2.08×

Reported

license unknown · no source

unknown

50.9µs

2.25×

Reported

license unknown · no source

unknown

53.2µs

2.35×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

61.2µs

2.70×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

61.2µs

2.70×

Reported

license unknown · no source

unknown

61.1µs

2.69×

Reported

license unknown · no source

unknown

98.5µs

4.34×

Reported

license unknown · no source

unknown

99.0µs

4.36×

Reported

license unknown · no source

unknown

101.9µs

4.49×

Reported

license unknown · no source

unknown

110.5µs

4.87×

Reported

license unknown · no source

unknown

148.7µs

6.56×

Reported

license unknown · no source

unknown

222.7µs

9.81×

Reported

license unknown · no source

Amir M. Mir | SF Tensor

unknown

387.9µs

17.09×

Reported

license unknown · no source

unknown

443.3µs

19.54×

Reported

license unknown · no source

unknown

587.0µs

25.87×

Reported

license unknown · no source

Semantics

Inputs and outputs

hidden_statesfp32 [num_tokens, hidden_size]

weight_fp4float4_e2m1fn_x2 [n_routed_experts, hidden_size_half]

weight_scalefp8_e4m3 [n_routed_experts, sf_k]

weight_global_scalefp32 scalar

e_score_correction_biasfp32 [n_routed_experts]

topk_indicesint64 [num_tokens, num_experts_per_tok]

topk_weightsfp32 [num_tokens, num_experts_per_tok]

Axes and behavior

sf_ksf_k = hidden_size // sf_vec_size

n_groupconstant = 1

num_tokensvariable

topk_groupconstant = 1

hidden_sizeconstant = 2688

sf_vec_sizeconstant = 16

hidden_size_halfhidden_size_half = hidden_size // 2

n_routed_expertsconstant = 128

num_experts_per_tokconstant = 6

determinismunspecified

constraintsNo mutation or aliasing

Identity

alias027_nvfp4_moe_router_projectionmodelnvidia-nemotron-3-nano-30b-a3b-nvfp4sha2565da130fd939c…

Sources: NVIDIA SOL-ExecBench (2026-08-22)last observed 2026-08-22How records are decided JSON