Sparse GEMM NVFP4 n2048 k2048
2 eligible runs
gemm
2:4 structured-sparse NVFP4 x NVFP4 GEMM (N=2048, K=2048). A keeps 2 nonzero E2M1 values per 4-col group (A_vals) at positions (A_meta) with UE4M3 per-16 block scale (A_scale); B is dense NVFP4 (E2M1 + UE4M3 per-16 scale). Hand-written CUDA baseline.
Source baseline · unbeaten
1.54msmean
cuda_sparse_nvfp4_naive_n2048_k2048FlashInfer-Bench baselines · Apache-2.0 · cuda
Reproduction-ready 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 = 64 · int8 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-06-06Record history →
Estimated floor 304 ns · record 5063.3× above itestimate, not evidence ›
DRAM 304 ns · bandwidth-bound on B200
every declared tensor crosses HBM exactly once (8,000 GB/s, B200 datasheet)
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
11.54ms1.00×Reproduction-ready · 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
cuda_sparse_nvfp4_naive_n2048_k2048FlashInfer-Bench baselines
cuda
1.54ms
1.00×
Reproduction-ready
Apache-2.0 · source
Semantics
Inputs and outputs
a_valsint8 [m, g]
a_metaint8 [m, g]
a_scaleint8 [m, k_blocks]
b_fp4int8 [n, k_half]
b_scaleint8 [n, k_blocks]
cbf16 [m, n]
Axes and behavior
gconstant = 512
kconstant = 2048
mvariable
nconstant = 2048
k_halfconstant = 1024
k_blocksconstant = 128
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliassparse_gemm_nvfp4_n2048_k2048sha2565ccea92f1de0…
Sources: FlashInfer-Bench (2026-06-06) · Apache-2.0last observed 2026-06-06How records are decidedJSON