Skip to content
KernelIndex
Search⌘K
Sign in

MoE complete layer with shared expert backward

27 eligible runs
moe

Backward pass for complete MoE layer with shared expert. Computes gradients through router (sigmoid + top-k selection), 128 routed experts with SwiGLU, and shared expert with SwiGLU. Gradients flow back through scatter-gather routing pattern.

Fastest reported
126.9µsmean · 78% of SOL
HyraLicense unknown · unknown

Reported evidence · last observed 2026-07-18. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencyhidden_size = 4096 · n_routed_experts = 128 · num_experts_per_tok = 8 · moe_intermediate_size = 140816 cases
batch_seq_len
dtype
192bf16/fp32211bf16/fp32293bf16/fp32384bf16/fp32512bf16/fp32997bf16/fp321321bf16/fp321571bf16/fp321879bf16/fp322053bf16/fp323072bf16/fp323089bf16/fp323719bf16/fp324093bf16/fp326144bf16/fp328192bf16/fp32
scrolls · 16 cases total
Aggregatessuite of 16 cases · mean latency
HardwareBest knownImplementationRuns
NVIDIA B200 · env 1126.9 µsHyra20NVIDIA B200 · env 2143.9 µsRecursive7

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 · 20 results · last observed 2026-08-24Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
Hyra
126.9µs78% SOL
1.00×
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 →
2
Electric Cheetah
134.3µs76% SOL
1.06×
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 →
3
ac4k
139.1µs74% SOL
1.10×
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 →
4
RIAC as well
140.7µs75% SOL
1.11×
Reported · license unknown · no source
2026-08-24

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

no sourcelicense unknownno install recipeRun detail →
5
ac4k
146.2µs73% SOL
1.15×
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 →
6
Amir M. Mir | SF Tensor
147.9µs72% SOL
1.17×
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 →
7
Geometric
148.8µs73% SOL
1.17×
Reported · license unknown · no source
2026-08-21

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

no sourcelicense unknownno install recipeRun detail →
8
Geometric
151.1µs73% SOL
1.19×
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 →
9
Geometric
151.2µs74% SOL
1.19×
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 →
10
Recursive
151.3µs72% SOL
1.19×
Reported · license unknown · no source
2026-07-10

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

no sourcelicense unknownno install recipeRun detail →
11
Geometric
152.9µs72% SOL
1.20×
Reported · license unknown · no source
2026-08-18

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

no sourcelicense unknownno install recipeRun detail →
12
Rapid Tiger
160.8µs70% SOL
1.27×
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 →
13
Amir M. Mir | SF Tensor
162.1µs71% SOL
1.28×
Reported · license unknown · no source
2026-07-10

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

no sourcelicense unknownno install recipeRun detail →
14
Gleaming Coyote
163.6µs69% SOL
1.29×
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 →
15
Geometric
164.0µs71% SOL
1.29×
Reported · license unknown · no source
2026-08-18

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

no sourcelicense unknownno install recipeRun detail →
16
doubleAI
167.6µs69% SOL
1.32×
Reported · license unknown · no source
2026-07-10

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

no sourcelicense unknownno install recipeRun detail →
17
Proud Wren @ NVIDIA
174.2µs68% SOL
1.37×
Reported · license unknown · no source
2026-07-12

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

no sourcelicense unknownno install recipeRun detail →
18
Geometric
177.2µs68% SOL
1.40×
Reported · license unknown · no source
2026-08-19

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

no sourcelicense unknownno install recipeRun detail →
19
Geometric
182.5µs67% SOL
1.44×
Reported · license unknown · no source
2026-08-15

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

no sourcelicense unknownno install recipeRun detail →
20
Geometric
208.8µs62% SOL
1.64×
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 →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
Hyra
unknown
126.9µs
1.00×
Reported
license unknown · no source
Electric Cheetah
unknown
134.3µs
1.06×
Reported
license unknown · no source
ac4k
unknown
139.1µs
1.10×
Reported
license unknown · no source
RIAC as well
unknown
140.7µs
1.11×
Reported
license unknown · no source
Geometric
unknown
151.2µs
1.19×
Reported
license unknown · no source
Recursive
unknown
143.9µs
1.13×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
147.9µs
1.17×
Reported
license unknown · no source
doubleAI
unknown
154.6µs
1.22×
Reported
license unknown · no source
Show all 27 implementations ›
Rapid Tiger
unknown
160.8µs
1.27×
Reported
license unknown · no source
Gleaming Coyote
unknown
163.6µs
1.29×
Reported
license unknown · no source
Proud Wren @ NVIDIA
unknown
174.2µs
1.37×
Reported
license unknown · no source
Cursor
unknown
349.0µs
2.75×
Reported
license unknown · no source
ac4k
unknown
146.2µs
1.15×
Reported
license unknown · no source
Geometric
unknown
151.1µs
1.19×
Reported
license unknown · no source
Geometric
unknown
148.8µs
1.17×
Reported
license unknown · no source
Geometric
unknown
152.9µs
1.20×
Reported
license unknown · no source
Recursive
unknown
151.3µs
1.19×
Reported
license unknown · no source
Geometric
unknown
164.0µs
1.29×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
162.1µs
1.28×
Reported
license unknown · no source
doubleAI
unknown
167.6µs
1.32×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
169.2µs
1.33×
Reported
license unknown · no source
Geometric
unknown
177.2µs
1.40×
Reported
license unknown · no source
Geometric
unknown
182.5µs
1.44×
Reported
license unknown · no source
Geometric
unknown
208.8µs
1.64×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
325.9µs
2.57×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
606.3µs
4.78×
Reported
license unknown · no source
Amir M. Mir | SF Tensor
unknown
1.93ms
15.17×
Reported
license unknown · no source

Semantics

Inputs and outputs
grad_outputbf16 [batch_seq_len, hidden_size]
hidden_statesbf16 [batch_seq_len, hidden_size]
router_weightbf16 [n_routed_experts, hidden_size]
e_score_correction_biasfp32 [n_routed_experts]
router_logitsfp32 [batch_seq_len, n_routed_experts]
scoresfp32 [batch_seq_len, n_routed_experts]
topk_indicesint64 [batch_seq_len, num_experts_per_tok]
topk_weightsfp32 [batch_seq_len, num_experts_per_tok]
score_maskfp32 [batch_seq_len, n_routed_experts]
shared_expert_gate_weightbf16 [moe_intermediate_size, hidden_size]
shared_expert_up_weightbf16 [moe_intermediate_size, hidden_size]
shared_expert_down_weightbf16 [hidden_size, moe_intermediate_size]
shared_gate_outputbf16 [batch_seq_len, moe_intermediate_size]
shared_up_outputbf16 [batch_seq_len, moe_intermediate_size]
shared_activatedbf16 [batch_seq_len, moe_intermediate_size]
grad_hidden_statesbf16 [batch_seq_len, hidden_size]
grad_router_weightfp32 [n_routed_experts, hidden_size]
grad_shared_expert_gate_weightbf16 [moe_intermediate_size, hidden_size]
grad_shared_expert_up_weightbf16 [moe_intermediate_size, hidden_size]
grad_shared_expert_down_weightbf16 [hidden_size, moe_intermediate_size]
Axes and behavior
hidden_sizeconstant = 4096
batch_seq_lenvariable
n_routed_expertsconstant = 128
num_experts_per_tokconstant = 8
moe_intermediate_sizeconstant = 1408
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias080_moe_complete_layer_with_shared_expert_backwardmodelglm-4-5-airsha2566a7cb8127160…
Sources: NVIDIA SOL-ExecBench (2026-08-24)last observed 2026-08-24How records are decidedJSON