AlphaChip
A deep reinforcement learning method that places blocks on a chip floorplan, pre-trained on blocks from earlier chips and then applied to new designs. Google says its layouts have been used in every TPU generation since 2020, including TPU v5e, v5p and Trillium, and in Google's Axion CPUs.
An RL agent lays out blocks of Google's TPUs, the accelerators that Google says lie at the heart of Gemini, Imagen and Veo, and it gets better and faster as it pre-trains on blocks from earlier chip generations. Better TPU layouts make the hardware for the next models cheaper and faster; Google has not stated that AlphaChip is trained on the TPUs it lays out, so the loop runs through the wider AI stack rather than back into AlphaChip alone.
- RL agent lays out TPU blocks
- Better layouts make AI hardware cheaper
- Agent pre-trains on earlier chip generations
- RL agent lays out TPU blocks
- Better layouts make AI hardware cheaper
- Agent pre-trains on earlier chip generations
Why it is a road to recursion
It is the longest-running deployed case of AI laying out the accelerators that train AI, and its lead authors later founded Ricursive Intelligence to push this loop toward recursion.
Evidence
Google reports that AlphaChip generates layouts that match or beat human experts in hours rather than the weeks or months of human effort, and that its layouts have been used in every TPU generation since 2020.
Related loops
More hardware and chips →- AI system writes accelerator hardware
- Qwen runs on the new accelerator
- Qwen finds optimizations for the accelerator
- Optimizations improve next accelerator generation
- repeat
- AI systems would design chips
- Chips would train more capable AI
- That AI designs the next chips
- repeat