Compute and hardware

Benchmark kernels, code changes, and hardware designs across many variants, and map the supply chains behind them.

A processor running kernel variants, where a few variants run faster than in the last round

Fields in this area

A GPU die with twelve compute units coloured by how busy they are, a cache band, and memory stacks at each side

Kernel optimization

Benchmark many kernel variants and keep the fastest correct ones.

A code change with removed lines in red and added lines in green, beside a branch graph

Software engineering

Test several code changes in parallel before you merge one.

A circuit board with chips, connectors, capacitors, copper traces, and a measured signal trace

Hardware engineering

Simulate and check many hardware design variants.

A container ship loaded with stacks of coloured containers, a few slots empty

Supply chain research

Map the suppliers and parts behind a product.

Things you can ask

Pick a field and watch Antinuous start on a request.

/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.

Search kernel variants for a faster versionSee how it works

How the work runs

The usual steps for work in this area, from the first search to what you get back.

Code, specs, and parts lists
- Repositories
  Agents read the source, tests, recent changes, and the inputs to benchmark.
- Schematics
  They read schematics, netlists, specs, and datasheets.
- Parts lists
  They start from your bill of materials or supplier sheet.
A source file open in an editor, read line by line
Variants on many computers
- Kernel variants
  Each variant tries different tile sizes, block sizes, and layouts.
- Candidate changes
  Each candidate code change runs in its own workspace.
- Circuit simulation
  Every component choice goes through the same circuit simulation.
- Thousands at once
  When massively parallel computation can speed up the work, the agent starts a swarm that runs thousands of them at once.
A matrix multiply split into tiles, with finished tiles in green and the tile in progress in orange
Checked and timed
- Against the reference
  Each kernel variant is checked against the reference, then timed.
- Full test suite
  The full suite runs on every candidate change.
- Supply risks
  Parts with a single supplier or recent changes are flagged.
Candidate code changes branching from main, each with its test results passing, failing, or still running
Diffs and benchmarks
- Passing changes
  A report lists which changes pass, with a diff for each.
- Notebook
  A notebook keeps the code and results for each variant.
A code diff beside run times for each variant, the fastest correct ones in green