Skip to content
KernelIndex
Search⌘K
Sign in

GEMM MXFP4 n4096 k4096

9 eligible runs
gemm

MXFP4 dense GEMM C = A @ B.T (N=4096, K=4096). Inputs A and B are bf16 values pre-snapped to the MXFP4 (E2M1 data + UE8M0 block scale, block size 32) representable grid, so MXFP4 kernels are numerically exact against the high-precision reference.

Source baseline · unbeaten
271.3µsmean
mm_fp4_mxfp4_cudnn_n4096_k4096FlashInfer-Bench baselines · Apache-2.0 · python

Reported evidence · last observed 2026-06-06. The source's designated baseline implementation. Reported by source; not independently reproduced.

No package · vendor the source →View source →Run detail →

Current records

Workloadm = 16 · bf16m = 128 · bf16m = 1024 · bf16

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload m = 1024 · bf16 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 3 results · last observed 2026-06-06Record history →
Estimated floor 15.3 µs · record 17.76× above itestimate, not evidence ›
DRAM 5.24 µs · compute 15.3 µs · compute-bound on B200
every declared tensor crosses HBM exactly once (8,000 GB/s, B200 datasheet)
2·M·N·K with M=1024, N=4096, K=4096 at the dense bf16 peak (2,250 TFLOP/s)
headroom-v1: a lower bound from declared tensors and datasheet peaks. A kernel can sit well above it for good reasons.
#
Implementation
Latency
vs #1
Trust
Observed
1
mm_fp4_mxfp4_cudnn_n4096_k4096baselineFlashInfer-Bench baselines
271.3µs
1.00×
Reported · Apache-2.0 · source
2026-06-06

Measured exactly what you asked. The source's designated baseline implementation. Reported by source; not independently reproduced.

source mirroredApache-2.0no install recipeView source →Run detail →
2
mm_fp4_mxfp4_flashinfer_n4096_k4096baselineFlashInfer-Bench baselines
284.7µs
1.05×
Reported · Apache-2.0 · source
2026-06-06

Measured exactly what you asked. The source's designated baseline implementation. Reported by source; not independently reproduced.

source mirroredApache-2.0no install recipeView source →Run detail →
3
mm_fp4_mxfp4_flashinfer_n4096_k4096baselineFlashInfer-Bench baselines
304.3µs
1.12×
Reported · Apache-2.0 · source
2026-06-06

Measured exactly what you asked. The source's designated baseline implementation. Reported by source; not independently reproduced.

source mirroredApache-2.0no install recipeView source →Run detail →

Scaling by m

100.0 µs200.0 µs300.0 µs161281024m →277.1 µs274.2 µs271.3 µsmm_fp4_mxfp4_cudnn_n4096_k4096293.7 µs286.2 µs284.7 µsmm_fp4_mxfp4_flashinfer_n4096_k4096
mm_fp4_mxfp4_cudnn_n4096_k4096mm_fp4_mxfp4_flashinfer_n4096_k4096best per workload · NVIDIA B200 · CUDA 13.0 · pytorch 2.11.0+cu130 · flashinfer-bench held constant

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
mm_fp4_mxfp4_cudnn_n4096_k4096FlashInfer-Bench baselines
python
271.3µs
1.00×
Reported
Apache-2.0 · source
mm_fp4_mxfp4_flashinfer_n4096_k4096FlashInfer-Bench baselines
python
284.7µs
1.05×
Reproduction-ready
Apache-2.0 · source

Semantics

Inputs and outputs
abf16 [m, k]
bbf16 [n, k]
cbf16 [m, n]
Axes and behavior
kconstant = 4096
mvariable
nconstant = 4096
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgemm_mxfp4_n4096_k4096sha256c13c37b4c44b…
Sources: FlashInfer-Bench (2026-06-06) · Apache-2.0last observed 2026-06-06How records are decidedJSON