PortentDevRequest access
Co-design engine · Pillar C

Finding faster implementations, with every speedup traced to one change

Optimization is a search over the space of implementation choices. Portent spawns sub-agents that generate candidates, filters them through a cascaded gate, evolves the survivors with a diversity-preserving search, and cross-checks answers across models, all under a strict one-variable-per-change rule.

The design space

Each implementation choice is a dimension

Block size, warps per block, tile shape, pipeline depth, algorithm choice: each is a dimension. A configuration is a path across all of them. The engine searches this space instead of hand-guessing a single point in it. The highlighted path is a current best; the faint ones are candidates still in play.

block sizewarps / blocktile Mtile Npipeline stagesalgorithm
Evolutionary search + MAP-Elites

A diversity archive keeps the search broad

Candidates are bred and filtered across multiple islands. A MAP-Elites archive keeps the best candidate found in each behavioral niche, so the search preserves diversity and keeps exploring instead of collapsing onto one local optimum. Each cell is a niche; the brightest is that niche's best.

occupancy →arithmetic intensity →
Cascaded evaluation gate

Cheap checks before expensive ones

A candidate has to pass parse, then compile, then a correctness check before it is ever benchmarked. The expensive measurement runs only on the few that get that far, so anything measured is already known to be correct.

generatedmanyparsecompilecorrectnessbenchmarkfew

illustrative funnel · proportions, not a specific run

Multi-LLM consensus

Models cross-check before an answer counts

Frontier models answer independently and cross-check each other. A synthesizer resolves disagreements into a single answer, and that answer still has to clear the deterministic gate before it counts. No single model's guess is trusted on its own.

model Amodel Bmodel Csynth-esizeronevalidated
The bandit

Effort shifts toward what's working

A multi-armed bandit tracks which strategy is producing improvements and shifts more of the budget toward it, while still spending some on the others so a strategy that pays off later isn't cut off early.

greedyevolverewritefuseexploit ▲◂ still exploring
One-variable discipline

Every change touches exactly one layer, so every speedup is attributable

A run that bundles three edits tells you nothing about which one helped. Portent changes one variable at a time and re-measures. A gain is credited to a single layer, backed by a before/after on the same harness. A speedup is never credited to a bundle of changes made at once.

one variable changed, one number measured

Sample results

Every result ships with its provenance

A co-design result takes this shape: a value plus the exact commit, image, command, and timestamp behind it. Until a live run lands, the value reads as a dash and the card is marked a placeholder.

matmul kernel vs referenceprovenance-stamped
×
commit
0000000
image
substrate/codesign@sha256:…
cmd
substrate codesign --layer L0 --reps 30
at
pending first live run
note
placeholder · not a real run
candidates evaluatedprovenance-stamped
cmd
substrate codesign --report
note
placeholder · populated by a real search
single-variable attributionprovenance-stamped
1layer / change
cmd
substrate codesign --enforce-one-variable
note
invariant, not a measurement; enforced by the harness