OpenChip is an open framework for building silicon with collaborating engineering agents. Our chip target is autonomous execution of a complete small language model, built with open hardware and carried through to working silicon.
The target chip boots from its own ROM, checks firmware and model images in external Flash, and runs the complete prefill and decode loop on CoralNPU. Tokenization, model execution and text generation run on the chip. A serial terminal provides input and output. Automatic boot and fixture execution work without a connected host.
First-silicon workload: the complete stories260K model in FP32, with a 40-token runtime context. This architecture is a design target; implementation and physical qualification are tracked separately from this system view.
Target contract · Architecture decision · Native IP
Architecture, RTL, verification, software and physical design have separate owners. An orchestrator manages the goal and dependencies; specialists deliver through manifest pull requests, and independent reviewers check behavior, evidence and scope before integration. The canonical role methods work across coding-agent clients. Start with the agent constitution.
The reference blink and tinyNPU designs exercise lint, cocotb simulation and
selected formal jobs. Public CI also checks framework tooling and role boundaries,
and packages the exact tested source after successful main-branch runs.
Verification records the tested scope, coverage denominators
and remaining gaps. These results establish neither full-SoC closure nor tapeout
readiness. Coverage reporting is diagnostic; lifecycle signoff is a manual,
independently reviewed process.
The GitHub agent workflow supports its configured provider's dispatch and review. Methods, a local dispatch ledger and durable state support session resumption. They do not implement a cross-provider scheduler, enforced workspace isolation or background monitoring.
make framework-test # Python, documentation and boundary checks
make agents # inspect available coding-agent clients
make agents-sync # expose canonical skills to supported clients
make lint # reference RTL lint
make sim MOD=npu # tinyNPU cocotb regression
make formal MOD=blink # configured reference formal jobs
make help # other entrypoints, including unimplemented targetsFramework checks need Python 3.9+ and shell/Git tools. Reference hardware checks need the pinned EDA tools and Python dependencies described in CI and toolchain. Imported IP may retain its native build, verification and implementation tools.
| Directory | Purpose | Primary owner |
|---|---|---|
docs/ |
Framework guidance, specifications and architecture decisions | Orchestrator for navigation; architects and verification specialists retain contract ownership |
hw/, sw/ |
Reference RTL, DV, formal, implementation flows and software | Assigned RTL, verification, backend and software specialists |
workloads/, explore/ |
Workload profiles and reproducible architecture cost models/results cited by specs and ADRs | Chief architect with workload specialists |
flow/ |
Shared Make recipes, tool pins, policy thresholds and boundary checks used by builds and CI | Assigned framework/flow author; independent integration review |
provenance/ |
Registered account mappings and public work-item evidence used by attribution/reporting tools | Assigned framework author; independent identity/evidence review |
evidence/ |
Reproducible evidence packages for explicitly scoped engineering claims | Specialist producers; integrator assembles and audits |
scripts/, tools/ |
Framework validation/reporting utilities and agent-client setup | Assigned framework/flow author |
.agents/, .github/ |
Canonical role methods and repository automation | Assigned framework author; integrator reviews |
explore/ retains engineering rationale, provenance/ records attribution, and
flow/ executes shared checks; all have live consumers. Private project state
and experiment receipts stay in the approved local store, outside public history.
The CoralNPU native IP entry now provides pinned source acquisition, native model generation and runtime tooling. The complete-model engineering contract remains a draft and supplies workload evidence for the target SoC. Autonomous boot, real memory/peripheral integration and physical qualification follow their own reviewed contracts. The smaller tinyNPU stays a fast framework regression. See the roadmap.
- Constitution and ownership: startup, independent roles and branch delivery.
- Silicon lifecycle: incremental, scoped signoff from feasibility through production validation.
- Verification: methods, public assets, recorded results and coverage diagnostics.
- CI and toolchain: actual jobs, pins and tested-source artifacts.
- Attribution: registered identities and evidence limits.
- Specifications and architecture decisions: specialist-owned engineering contracts.
Project goals, dispatch state and private chip evidence belong in an approved local store. Reusable state and signoff templates contain no project data.
Licensed under Apache-2.0.