01
Template-driven generation
Candidates composed from CUTLASS templates and targeted rewrites of the hot kernels.
Kernels · Kernl
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.
Capabilities
01
Candidates composed from CUTLASS templates and targeted rewrites of the hot kernels.
02
Bayesian and RL search over tile sizes, stages, scheduling and layouts.
03
Each kernel is checked for numerical equivalence before it can be benchmarked.
04
Speed measured with Nsight Compute on the exact target architecture.
05
Re-run the same search on Blackwell instead of rewriting by hand.
06
Kernels shipped as versioned artifacts with the recipe that built them.
Candidate ledger
Kernl tries many kernel variants and keeps few. Filter by verdict or by kernel family.
Example run on one H100. Candidates, speedups and reasons are illustrative, not measured results.
How it works
Take the ranked kernels from Profyle.
Propose variants from templates and rewrites.
Simyx ranks; the correctness gate filters.
Measure on the GPU; keep only faster, correct kernels.
Specification
| Item | Detail | Status |
|---|---|---|
| Input | Opportunity map from Profyle, model graph | Pilot |
| Kernel families | Fused attention, GEMM, fused epilogues | Pilot |
| Libraries | CUDA, CUTLASS, cuDNN baselines | Pilot |
| Targets | H100, H200; Blackwell next | Pilot · Planned |
| Output | Kernel artifacts folded into TensorRT engines | Pilot |
| Novel kernel synthesis | Kernels written from scratch, beyond templates | Research |
Place in the loop
It consumes Profyle's bottlenecks and hands winning kernels to Servex for the final engine.
FAQ
Not in v1. It composes and tunes CUTLASS templates and makes targeted rewrites. Fully novel kernel synthesis is a research track.
Outputs are compared to the reference kernel across representative inputs within a tolerance you set. Any kernel outside it is rejected.
No. Kernl re-runs its search when the GPU, driver or model changes and replaces the kernels with new ones.
Attention and GEMM first, because that is where most inference time goes. Other operations follow.
Next step
In the pilot Kernl runs on the top bottlenecks Profyle finds in your model.