Sparse GEMM fp4 FP8 n2048 k2048
2 eligible runs
gemm
2:4-style structured-sparse mixed FP4(E2M1) x FP8(e4m3) GEMM (N=2048, K=2048). A keeps 2 nonzero E2M1 values per group of 4 K-columns (A_vals: two fp4 codes per byte) at positions given by A_meta (two 2-bit indices per byte). B is dense FP8 e4m3. Reference reconstructs A and matmuls.
Source baseline · unbeaten
889.9µsmean
cuda_sparse_fp4_fp8_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 · fp8_e4m3 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-06-06Record history →
Estimated floor 532 ns · record 1671.29× above itestimate, not evidence ›
DRAM 532 ns · compute 119 ns · bandwidth-bound on B200
every declared tensor crosses HBM exactly once (8,000 GB/s, B200 datasheet)
2·M·N·K with M=2048, N=64, K=2048 at the dense fp8 peak (4,500 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
1889.9µs1.00×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_fp4_fp8_naive_n2048_k2048FlashInfer-Bench baselines
cuda
889.9µs
1.00×
Reproduction-ready
Apache-2.0 · source
Semantics
Inputs and outputs
a_valsint8 [m, g]
a_metaint8 [m, g]
b_fp8fp8_e4m3 [n, k]
cbf16 [m, n]
Axes and behavior
gconstant = 512
kconstant = 2048
mvariable
nconstant = 2048
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliassparse_gemm_fp4_fp8_n2048_k2048sha256d929541e95ae…
Sources: FlashInfer-Bench (2026-06-06) · Apache-2.0last observed 2026-06-06How records are decidedJSON