Kernel optimization
Antinuous writes kernel variants, compiles them, and benchmarks each one on the same inputs. It checks every variant's output against a reference before it compares speed.
Use cases
Search kernel variants for a faster version
Each tile size, block size, and memory layout is a separate build and benchmark.
You ask Claude Code/antinuous Here is our matrix multiply kernel. Try different tile sizes, block sizes, and memory layouts, check each result against the current version and report the fastest correct ones.
Claude Code hands it to Antinuous.
- What it does
- Antinuous reads the kernel and its tests from GitHub or an upload.
- It writes variants and compiles each one in its code workspace.
- It checks each variant's output against the reference and drops any that differ.
- It benchmarks the rest on the same inputs and charts the results in a notebook.
- You get
A notebook with each variant, its benchmark results, and its code.
- Write variants
- Check against the reference
- Benchmark
- Keep the fastest
- Try new variants
Who it's for
- Performance engineers
Benchmark many kernel variants on the same inputs.
- ML infrastructure engineers
Compare kernel options before you ship one.
- Compiler engineers
Test code generation settings across many workloads.
Works with
GitHub
arXiv
SlackMore in compute and hardware
Software engineering
Run your kernel benchmarks on Antinuous.
Start with our agent, or connect your own to the same compute, orchestration, and simulation.