Janus
Janus is provably-correct, deterministic ML compute IP: one small memory-rich cell, formally verified in Lean and tiled into a mesh. The specification, the proofs, and the RTL that goes to silicon are one machine-checked artifact — a datasheet that builds cannot lie about the chip.
- Correctness is a theorem. The RTL is proved to refine
the instruction set — no caveats, no
sorry. Integration and verification start from a proof, not a datasheet. - Determinism is a theorem. No caches, no speculation, a static schedule — worst-case timing is analysable as a theorem. Built for safety-critical and hard-real-time.
- Reproducibility is a theorem. Block-float arithmetic, formally specified: every inference is bit-exact and carries its own proof. Certifiable ML, not just fast ML.
- Efficient where GPUs bleed. Resident weights and a balanced fabric target the latency- and memory-bound regimes — batch-1 decode, long context — where a GPU discards most of its FLOPs.
- Auditable end to end. Open tools, open process, one source from theorem to gate. No black box at any layer.
- Scales by tiling. One proved cell, replicated across the mesh and defect-tolerant by construction. Prove once, compose everywhere.