Skip to content
KernelIndex
Search⌘K
Sign in

GEMM n2048 k2048

43 eligible runs
gemm

General matrix multiply (GEMM) C = A @ B.T. Captured from Llama-3.2-1B attn.o_proj.

Source baseline · unbeaten
10.7µsmean
torch / matmul756deaFlashInfer-Bench baselines · Apache-2.0 · python

Reported evidence · last observed 2026-04-20. The source's designated baseline implementation. Reported by source; not independently reproduced.

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

Current records

Workloadm = 56 · fp1643 cases
m
dtype
1fp162fp164fp167fp168fp1615fp1616fp1624fp1632fp1635fp1640fp1648fp1656fp1664fp1670fp1672fp1680fp1688fp1696fp16104fp16112fp16120fp16128fp16136fp16144fp16152fp16160fp16168fp16176fp16184fp16192fp16200fp16208fp16216fp16224fp16232fp16240fp16248fp16256fp16972fp162053fp162379fp168192fp16
scrolls · 43 cases total

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload m = 56 · fp16 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-04-20Record history →
Estimated floor 1.08 µs · record 9.91× above itestimate, not evidence ›
DRAM 1.08 µs · compute 209 ns · bandwidth-bound on B200
every declared tensor crosses HBM exactly once (8,000 GB/s, B200 datasheet)
2·M·N·K with M=56, 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
1
torch / matmul756deabaselineFlashInfer-Bench baselines
10.7µs
1.00×
Reported · Apache-2.0 · source
2026-04-20

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

20.0 µs40.0 µs14248825620538km →15.3 µs12.3 µs11.6 µs11.0 µs10.9 µs10.8 µs10.7 µs10.8 µs10.8 µs10.8 µs10.7 µs10.7 µs10.7 µs10.7 µs11.5 µs11.5 µs10.7 µs10.7 µs10.9 µs10.7 µs10.7 µs10.7 µs10.7 µs10.7 µs10.7 µs10.8 µs10.8 µs10.7 µs10.8 µs10.7 µs10.8 µs10.8 µs10.8 µs10.7 µs10.8 µs10.8 µs10.8 µs10.7 µs10.8 µs14.8 µs19.6 µs20.9 µs51.2 µstorch / matmul 756dea
torch / matmul 756deabest per workload · NVIDIA B200 · CUDA 13.0 · pytorch 2.11.0+cu130 · flashinfer-bench held constant

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
torch / matmul756deaFlashInfer-Bench baselines
python
10.7µs
1.00×
Reported
Apache-2.0 · source

Semantics

Inputs and outputs
afp16 [m, k]
bfp16 [n, k]
cfp16 [m, n]
Axes and behavior
kconstant = 2048
mvariable
nconstant = 2048
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgemm_n2048_k2048modelllama-3-2-1bsha2560f0f71d7a672…
Sources: FlashInfer-Bench (2026-04-20) · Apache-2.0last observed 2026-04-20How records are decidedJSON