BLUF

This article explains the complexities of chip design and how machine learning allied with an AI might improve efficiency in ‘floorplanning’.

Summary

This article makes the following points:
  • Floorplanning is the fundamental step in the design of microchips.
  • ‘Floorplanning’ involves the placement of standard cells and macroblocks onto silicon.
  • Number of components per microchip doubles every two years.
  • Component placement affects microchip performance.
  • Engineers produce the design with the assistance of software.
  • Evolving designs cause the engineer to restart the ‘floorplanning’.
  • Chip design takes days or weeks to complete.
  • Previous attempts to automate ‘floorplanning’ have failed.
  • A new machine–learning system was trained to floorplan successfully.
  • The system was able to design working floorplans within hours.
  • The new system was able to produce working designs on its first attempt.
  • The AI’s method replicates a method designers moved away from previously.
  • AI designs look vastly different from human design.
The technology has been incorporated into Google’s next-generation AI processors.