MACE: A Multicore Agentic Co-Design Engine

MACE is an agentic AI system that plans, builds, and verifies multicore OpenPiton hardware on the CHIA framework. It turns a hardware objective into a set of tasks, runs them in parallel, and accepts a change only when Verilator simulation of the RTL passes. MACE reaches OpenPiton through chia_openpiton, a CHIA adapter for OpenPiton’s configure, build, run, and collect operations.

Organization of this Document

The documentation has six parts.

The MACE User Manual covers installing MACE, running the loop and the baselines, and reading the results.

The MACE Requirements Specification lists what MACE must do.

The MACE Design Document describes the MACE loop: planning, parallel task execution, integration, verification, and failure analysis.

The chia_openpiton Adapter describes how the loop builds and runs OpenPiton.

The Supported Cores part covers MACE’s three native OpenPiton cores (Ariane, OpenSPARC T1, and PicoRV32), the generic tile for RV32 cores with no OpenPiton adapter, and how to add another core.

The MACE Evaluation part collects the baselines, results, and cost measurements.