Verification is the moat
Swipe the figure to read it — tap for the full view
- ~23bn tokens
- internal data NVIDIA used to adapt a model for its own engineers
- ChipNeMo, arXiv 2023
- ~1.5%
- of the compute of training the generalist model
- ChipNeMo, arXiv 2023
- 20 blocks
- what AlphaChip was pretrained on — three TPU generations
- Nature, 2021 · DeepMind
- ~100M
- labelled examples one study puts on specialised RTL models
- Chang et al., DAC 2024
- Design data is walled: PDKs are confidential, and public RTL is orders of magnitude short.
- We reward verification, not generation — simulation is the teacher, and it needs no corpus.
- Multi-gate signoff: vacuity-gated proofs, two DRC engines, provenance on every artifact.
Lumicule