Keynote: ML for Systems and Chip Design
summary
Google Brain ML for Systems applies reinforcement learning and graph neural nets to real engineering tasks: device placement for neural nets, graph partitioning, and chip placement. A trained RL policy maps ops to GPUs and reduces runtime and memory copies, reporting 15x speedup in search and tens of percent runtime gains; a partitioner generalizes to million-node graphs; a placement system produces manufacturable layouts and claims chip floorplanning in 6 hours with zero-shot generalization to new designs.
AI-written summary · about
chapters (5)
- 0:00Introduction to ML for Systems
- 2:47Learning to optimize device placement
- 8:48Learning to partition graphs
- 12:19Learning to optimize chip placement
- 19:02Conclusion and future work
related talks
10mChip floorplanning with deep reinforcement learning
01h 53mMachine Learning for Manycore System Design and Optimization
27mKeynote - Google Cloud, Machine Learning
10mKeynote Analysis
36mKeynote: AI, ML, Cloud for ML
53mML BuzzWords demystified
24mKeynote: Next-generation frameworks for Large-scale AI
06mML for Systems and Systems for ML