GEMM MXFP4 out NVFP4 n2048 k2048
2 eligible runs
gemm
NVFP4-output GEMM: C = A @ B.T quantized to NVFP4 (N=2048, K=2048). Inputs A, B are bf16 pre-snapped to the MXFP4 grid (so the FP4 GEMM is exact); output is packed E2M1 (C_fp4) plus UE4M3 per-16 block scale (C_scale, 128x4 swizzle, global scale 1).
Source baseline · unbeaten
354.7µsmean
mm_fp4_out_nvfp4_flashinfer_n2048_k2048FlashInfer-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.
Current records
Not measured on H100 for this workload. Challenges →
Source-native comparison · GPU NVIDIA B200 · Workload m = 256 · bf16 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-06-06Record history →
Estimated floor 1.18 µs · record 300.69× above itestimate, not evidence ›
DRAM 1.18 µs · compute 954 ns · bandwidth-bound on B200
every declared tensor crosses HBM exactly once (8,000 GB/s, B200 datasheet)
2·M·N·K with M=256, N=2048, K=2048 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
1354.7µs1.00×Reported · Apache-2.0 · source2026-06-06
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 →
Implementations
Implementation
Runtime
Best latency
Evidence
Availability
mm_fp4_out_nvfp4_flashinfer_n2048_k2048FlashInfer-Bench baselines
python
354.7µs
1.00×
Reproduction-ready
Apache-2.0 · source
Semantics
Inputs and outputs
abf16 [m, k]
bbf16 [n, k]
c_fp4int8 [m, n_half]
c_scaleint8 [m, n_blocks]
Axes and behavior
kconstant = 2048
mvariable
nconstant = 2048
n_halfconstant = 1024
n_blocksconstant = 128
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgemm_mxfp4_out_nvfp4_n2048_k2048sha25656ad21e4eff2…
Sources: FlashInfer-Bench (2026-06-06) · Apache-2.0last observed 2026-06-06How records are decidedJSON