Search · KernelIndex

Fastest reported

13.2µsmean · 97% of SOL

RIAC as wellLicense unknown · unknown

Reported evidence · last observed 2026-08-10. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Ranked under the source's published protocolWhy comparable?

GPUNVIDIA B200

Workloadsuite of 47 cases · mean latency

ProtocolSOL-ExecBench evaluation stack v1.1 · mean

The query left these dimensions open; this is the most-measured of 2 cohorts. The hardware chips above switch it.

Exact 62 Compatible 188 Supported 5 Other protocols 31sorted byRecommendedNewest

Has source5 License known5Installable0

Claims support; no run on this exact workload yet.

Implementation

Latency

Trust

Observed

—

claude-opus-4-1 / cudafeab9bclaude-opus-4-1-20250805

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

—

claude-opus-4-1_triton_a98005claude-opus-4-1-20250805

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

—

gemini-2.5-pro_cuda_292432gemini-2.5-pro

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

—

gemini-2.5-pro_triton_dorbxsgemini-2.5-pro

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

—

gpt-o3 / cuda7cdf4bgpt-o3

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

MLA fused QKV RoPE splitOperation in the mla-paged-attention familyMLA paged prefill causal h16 ckv512 kpe64 ps1Operation in the mla-paged-attention familyFP8 MLA KV compression projectionOperation in the mla-paged-attention familyFP8 MLA attention output projectionOperation in the mla-paged-attention familyMLA paged decode h16 ckv512 kpe64 ps64Operation in the mla-paged-attention familyMLA paged decode h8 ckv512 kpe64 ps1Operation in the mla-paged-attention family

Missing a kernel here? Evidence submissions open with the contribution beta.

This search as an API call

curlkicomparison key sha256:06e41496846fdedb…JSON

NVIDIA SOL-ExecBench · last observed 2026-08-18 · FlashInfer-Bench · last observed 2025-10-21

Fastest reported

13.2µsmean · 97% of SOL

RIAC as wellLicense unknown · unknown

Reported evidence · last observed 2026-08-10. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Ranked under the source's published protocolWhy comparable?

GPUNVIDIA B200

Workloadsuite of 47 cases · mean latency

ProtocolSOL-ExecBench evaluation stack v1.1 · mean

The query left these dimensions open; this is the most-measured of 2 cohorts. The hardware chips above switch it.

Exact 62 Compatible 188 Supported 5 Other protocols 31sorted byRecommendedNewest

Has source5 License known5Installable0

Claims support; no run on this exact workload yet.

Implementation

Latency

Trust

Observed

—

claude-opus-4-1 / cudafeab9bclaude-opus-4-1-20250805

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

—

claude-opus-4-1_triton_a98005claude-opus-4-1-20250805

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

—

gemini-2.5-pro_cuda_292432gemini-2.5-pro

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

—

gemini-2.5-pro_triton_dorbxsgemini-2.5-pro

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

—

gpt-o3 / cuda7cdf4bgpt-o3

—

— · Apache-2.0 · source

—

Declared support only; no measurement for this workload.

source mirroredApache-2.0no install recipeView source →

Missing a kernel here? Evidence submissions open with the contribution beta.

This search as an API call

curlkicomparison key sha256:06e41496846fdedb…JSON

NVIDIA SOL-ExecBench · last observed 2026-08-18 · FlashInfer-Bench · last observed 2025-10-21

MLA paged decode h16 ckv512 kpe64 ps1

Related

MLA paged decode h16 ckv512 kpe64 ps1

Related