Tilessa🔗

Every computer you have ever used asks for your trust. Its datasheet describes what the chip is supposed to do; whether it actually does — every cycle, every corner case — is a promise backed by testing, and testing can only ever sample. The history of hardware lives in the gap between the two: errata sheets, security surprises, processors that computed the wrong answer once in a billion operations.

This book is an experiment in closing that gap. It describes Tilessa, a small parallel computer built for machine-learning inference, and the book itself is written in a language a computer can check. Every definition in it is real, running code; every theorem is machine-verified each time the book is built. If a claim in this book were false, the book would fail to build, and you would not be reading it.

The current architecture is Nano-first. The first object to understand is the Nano core: a tiny, Turing-complete MAC engine. Around it sits a memory-rich tile, because sustained inference is usually limited by where weights, KV cache, and scale metadata live rather than by a peak multiplier count. BFP/MX conversion is therefore modeled as a placement choice: edge-only, tile-boundary, per-bank, or PE-local. The conservative default is integer Nano PEs, scale metadata in tile SRAM/fabric, and conversion at the tile boundary or mesh edge until measurement proves a more aggressive placement is worth the timing and proof cost.

Older tessera material remains in the book as the historical proof stack. It is still useful: the mesh, boundary, strap, and pipeline chapters show how the project reached executable proofs and routed silicon. But the design spine now runs Nano core → workload-tuned tile → measured evidence.