AI summary by TwinMind
Google’s Chip Designing AI
Asianometry
13m 18s
Introduction - ML in Chip Design
- ML hype died down but trend continues, making waves in chip design
- Focus on recent breakthrough: floor planning
- Google applying same AI prowess from Go mastery to chip design
- Paper by Azalea Mirhoseini and Anna Goldie on AI floor planning success
Floor Planning Fundamentals
- First major step in physical design after logic design stage
- Physical designer places logic/memory circuit groupings (netlist) onto chip canvas
- Modern abstractions:
- Tens of millions logic gates → standard cells
- Thousands memory blocks → macro blocks
- Goal: wire up with tens of kilometers of wiring on fingernail-sized chip
Floor Planning Challenges
- Minimize white space (uncovered areas)
- Fixed chip area/size constraints (mobile phones need smaller chips for battery space)
- Wire placement optimization:
- Minimize latency, power consumption, heat
- Intel study: 51% of microprocessor power consumed driving signals through interconnects
- Traditional EDA tools inadequate - designed for millions small cells, not large blocks with little future parameter info
- Analogy: placing doors, hallways, sockets, connections in 10+ million room house without knowing room locations/usage
Complexity Comparison
- Chess: 10^23 states
- Go: 10^360 states
- Chip floor planning: up to 10^9000 states
- Problem grows harder with each new TSMC/Samsung node as netlists get larger
Traditional Approaches
Simulated Annealing
- Uses objective equation to calculate floor plan cost
- Cost based on wire length and area size
- Computer iteratively searches for globally optimal solution
- Flaws:
- Gets stuck at local optimum
- Declares victory too early
- Hill climbing methods help escape local optimums
- Long search times due to placement possibilities
Other Methods
- Compacted floor plans: no modules move left/right, represented as ordered binary tree
- Analytical method: mathematical formula + constraints to calculate smallest floor plan
- Partitioning: break larger problem into smaller circuits, optimize in passes
- Reality: designers use semi-manual potpourri of methods + human intuition
- Slow iteration becomes drag on productivity/speed
Machine Learning Background
- AI term meaningless now (broad blanket term like "cloud")
- ML specific, not interchangeable with AI
- Traditional symbolic AI (1950s): handwritten rules/sets
- Problem: underestimated implicit human knowledge/common sense
- ML approach: let data write rules, enabled by GPU processing power
- Neural networks: feed data, compare against pre-labeled results, minimize error without overfitting
Google's AI Floor Planning Approach
- Treated problem like game (similar to Go)
- Board = chip canvas
- Pieces = netlist blocks
- Win conditions based on floor plan evaluation metrics
- Neural network helps identify best moves/winning percentages
- Designer decides whether to place - network as tool
Training Process
- Set broad parameters: chip netlist metadata, process technology node
- Train neural network with episodes of state/actions/rewards:
- State: chip canvas shown to computer
- Actions: place macro blocks (memory block aggregations) on canvas
- Standard cells/logic circuits placed with standard EDA tool
- Reward: negatively correlated to wire length, congestion, density
- Network learns to maximize final reward through state+action+reward cycles
Results & Performance
- Speed improvement: major advantage
- Google TPU design comparison:
- Human expert: 6-8 weeks (72 hours per iteration for EDA evaluation)
- ML floor planner: 24 hours
- Quality comparison: three methodologies very close (within 2-5% range)
- ML not consistently better than human/EDA tools - sometimes human better, sometimes EDA tool best
- ML performs about as well as humans (based on human-created data)
Google's Chip Strategy
- Apple best chip designer among American tech giants
- Google making rapid strides in chip proficiency
- Recognized own hardware critical for superior cost/performance
- Started in databases, now consumer hardware
- Neural network applied to latest Google TPU design
- Bulking up hiring in Taiwan
- Expect more from Google in chip space
Future Applications
- Beyond chip macro block placement
- Potential use cases:
- Hardware design
- City planning
- Vaccine planning and distribution
- Technology could appear in many more use cases
