Kernels · Kernl

Kernels tuned on the GPU they will run on

Kernl generates and auto-tunes CUDA and CUTLASS kernels for the bottlenecks Profyle finds, checks each one for numerical correctness, and keeps only variants measured faster on real hardware.

Kernel searchExample
Bottleneck
Attention, decode
Candidates
212 generated
Benchmarked
18 after Simyx
Kept
4 kernels
Step time
−25.8 ms
Example. Values are illustrative.

Capabilities

What Kernl does

01

Template-driven generation

Candidates composed from CUTLASS templates and targeted rewrites of the hot kernels.

02

Auto-tuning search

Bayesian and RL search over tile sizes, stages, scheduling and layouts.

03

Correctness gate

Each kernel is checked for numerical equivalence before it can be benchmarked.

04

Real-GPU benchmark

Speed measured with Nsight Compute on the exact target architecture.

05

Re-tune per generation

Re-run the same search on Blackwell instead of rewriting by hand.

06

Reusable artifacts

Kernels shipped as versioned artifacts with the recipe that built them.

Candidate ledger

Filter an example kernel search

Kernl tries many kernel variants and keeps few. Filter by verdict or by kernel family.

Showing 8 of 8 example candidates
  1. C-014Fused attention, FlashAttention-style128×64 tiles, 2-stage pipelineAttention was memory-bound; fusion cuts HBM traffic.−14.2 msKept
  2. C-021Split-K GEMM on o_projsplit 4Predicted gain on small batch; measured slower.+2.1 msRejected
  3. C-033CUTLASS GEMM, FP8 inputs256×128 tile, persistentTensor-core utilization rose from 58% to 81%.−6.4 msKept
  4. C-038Fused RMSNorm and residual addone kernel instead of twoRemoves a launch and a round trip to memory.−1.9 msKept
  5. C-041Attention, 64×64 tiles3-stage pipelineFailed numerical check: max error above tolerance.fails checkRejected
  6. C-047Rotary embedding fused into QKVfused epilogueFaster on prefill, slower on decode; net worse at p95.+0.6 msRejected
  7. C-052Grouped GEMM for MLPSwiGLU fused epilogueFewer launches during decode.−3.3 msKept
  8. C-058Paged KV attention, larger pagespage 64 tokensNo measurable change; kept the simpler baseline.±0.1 msRejected

Example run on one H100. Candidates, speedups and reasons are illustrative, not measured results.

How it works

Inside a Kernl run

  1. 01Instant

    Read the map

    Take the ranked kernels from Profyle.

  2. 02Minutes

    Generate

    Propose variants from templates and rewrites.

  3. 03Minutes

    Prune and check

    Simyx ranks; the correctness gate filters.

  4. 04Hours

    Benchmark and keep

    Measure on the GPU; keep only faster, correct kernels.

Specification

Inputs, outputs and hardware

ItemDetailStatus
InputOpportunity map from Profyle, model graphPilot
Kernel familiesFused attention, GEMM, fused epiloguesPilot
LibrariesCUDA, CUTLASS, cuDNN baselinesPilot
TargetsH100, H200; Blackwell nextPilot · Planned
OutputKernel artifacts folded into TensorRT enginesPilot
Novel kernel synthesisKernels written from scratch, beyond templatesResearch

FAQ

Questions engineers ask

Does Kernl write CUDA from scratch?

Not in v1. It composes and tunes CUTLASS templates and makes targeted rewrites. Fully novel kernel synthesis is a research track.

How is correctness checked?

Outputs are compared to the reference kernel across representative inputs within a tolerance you set. Any kernel outside it is rejected.

Do I have to maintain the kernels?

No. Kernl re-runs its search when the GPU, driver or model changes and replaces the kernels with new ones.

Which operations are covered?

Attention and GEMM first, because that is where most inference time goes. Other operations follow.

Next step

Find out what your kernels leave behind

In the pilot Kernl runs on the top bottlenecks Profyle finds in your model.