Research

1 min read

The optimization space is 10¹⁷ wide. Here's how we search it.

Tiling, layout, fusion, scheduling and precision choices multiply fast.

Vladimir Zeljkovic

Count the knobs on a single transformer block — tile shapes, memory layouts, fusion boundaries, scheduling policy, precision per tensor — and the product runs past 10¹⁷ configurations before you have left the first layer.

Why exhaustive search and pure heuristics both fail

Exhaustive search is obviously out. Fixed heuristics — the traditional compiler answer — are out for a subtler reason: they encode the hardware and model shapes that existed when they were written, and both change every few months.

What we do instead

The harness treats the space as a search problem with a measured objective and a correctness gate. Proposals come from a model that has read the profile; every proposal is compiled, run, and checked against reference numerics before its throughput counts for anything.

Be there when
the next world opens.

Book a call