BIGGER PICTURE Scientific discovery
A chip’s floor plan can be learned
Before a processor can be manufactured, its circuit blocks must fit together with workable connections. AlphaChip turns that crowded layout problem into a sequence of decisions it can improve through practice.

- Design task
- Placement of circuit blocks
- Documented hardware
- TPU-v5 and later generations
- Public starting point
- Code and pretrained checkpoint
How the idea works
Represent circuit connections
A graph describes the blocks and their relationships.
Place and reward
Repeated layout attempts learn useful placement choices.
Check physical quality
Engineering tools evaluate the final manufacturing candidate.
A city with very short streets
Think of a chip floor plan as a dense city: blocks must fit, but the connections between them matter just as much. Long wires and congested routing can undermine a design. The original research therefore optimized placement using estimates of wire length and congestion while respecting density constraints. These are physical engineering requirements, not aesthetic preferences.
Sources: [2]
Learn a placement strategy from practice
AlphaChip places components one after another and receives feedback on the completed layout. Its graph-based representation captures which blocks connect to which. Practice on earlier chip designs gives the system a starting strategy for new ones. Engineers still define the circuit, constraints and validation process; the algorithm searches within that carefully specified problem.
Layouts that reached manufactured hardware
This did not remain a drawing exercise. Google’s project documentation says its reinforcement-learning placements were frozen for manufacture in TPU-v5, and its 2024 account describes use across subsequent TPU generations. The practical pathway is to the computing hardware on which researchers and developers run their workloads. Floorplanning helps engineers reach usable physical implementations.
The score is not the finished chip
The repository distinguishes quick proxy scores from slower engineering checks of timing, power and area. It also recommends training on relevant chip blocks. That boundary matters: a good layout score is not a universal energy-saving percentage, and a processor’s final performance has many causes. Open code and a pretrained checkpoint allow others to investigate the method.
Sources: [3]
The original sources
- Google DeepMindHow AlphaChip transformed computer chip design
- Google research authors / arXivChip Placement with Deep Reinforcement Learning
- Google Research / GitHubCircuit Training: features, validation and production FAQ