Grouped GEMM FP8 fp4 m contiguous g4 n4096 k2048
2 eligible runs
gemm
M-contiguous grouped mixed FP8xFP4 GEMM (4 groups, N=4096, K=2048), DeepGEMM SM100. Tokens of all groups are concatenated along M; m_indices gives the group id per row (-1 = padding). A is FP8 (float8_e4m3fn, per-token x128 float32 scale); B is per-group FP4 (packed E2M1 int8 + UE8M0 per-32 float32 scale). Reference dequantizes per group and matmuls; the kernel is exact against it.
Source baseline · unbeaten
22.4µsmean
deepgemm_m_grouped_fp8_fp4_contiguous_g4_n4096_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 = 960 · fp32/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 3.58 µs · record 6.25× above itestimate, not evidence ›
DRAM 2.88 µs · compute 3.58 µs · compute-bound on B200
every declared tensor crosses HBM exactly once (8,000 GB/s, B200 datasheet)
2·M·N·K with M=960, N=4096, 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
122.4µs1.00×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
deepgemm_m_grouped_fp8_fp4_contiguous_g4_n4096_k2048FlashInfer-Bench baselines
python
22.4µs
1.00×
Reported
Apache-2.0 · source
Semantics
Inputs and outputs
a_fp8fp8_e4m3 [m, k]
a_scalefp32 [m, k_a_blocks]
b_fp4int8 [g, n, k_half]
b_scalefp32 [g, n, k_b_blocks]
m_indicesint32 [m]
cbf16 [m, n]
Axes and behavior
gconstant = 4
kconstant = 2048
mvariable
nconstant = 4096
k_halfconstant = 1024
k_a_blocksconstant = 16
k_b_blocksconstant = 64
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgrouped_gemm_fp8_fp4_m_contiguous_g4_n4096_k2048sha256889c192a89e0…
Sources: FlashInfer-Bench (2026-06-06) · Apache-2.0last observed 2026-06-06How records are decidedJSON