owned this note   changed 2 years ago  
Published Linked with GitHub

Building an open source silicon ecosystem - Aaron Cunningham

video

  • Google's cloud computing services was built on the concept of clustering thousands of low-cost computers with general-purpose CPUs. Paper: Barroso, L. A., Dean, J., & Holzle, U. (2003). Web search for a planet: The Google cluster architecture. IEEE micro, 23(2), 22-28.
  • But Moore's law slows down and the demand for Google is scailing, eg, more youtube videos to be compressed
  • Therefore, Google is moving to custom silicon for specific tasks like compression. Google's TPU is an example as well
  • Open source
    • PDKs: specs to be followed so chips can be manufactured by a foundry. They often have strict licenses, which makes universities (especially in the US) hard to collaborate and thus limits the reproductivity of researches
      • find open source PDKs and partner up with foundaries, eg, Skywater
    • silicon toolchain
      • Google has been identifying open source projects that address process steps in a design flow and supports their researchers via academic grants. The goal has been to integrate these into an end-to-end toolchain, filling gaps via open source contributions from Google software engineers and industry partners
      • high level synthesis (.x): XLS
      • synthesis (.v): yosys
      • place and route (.lef/def): OpenROAD & OpenLane
      • GDS streaming and DRC (.gds): Magic & Klayout
  • Open multi project wafer (MPW) shuttle experiment
    • to check if these tools work
    • as Google only design a few chips annually, they accepts 40 open source designs per run from the community and manufacture them for free
    • 8 shuttles since lanching in 2020
      • received over 600 submissions from 19 countries
    • raise awareness and drive adoption of the open source toolchain
    • challenges
      • flaw in MPW-1 harness compromised all submissions
      • documentation needs improvement
      • not all shuttles have returned yet (6-9 months wait + delay)
    • some submissions
      • #007: YiFive
      • #005: FABulous
      • #024: Zero to ASIC Course tinytapeout.com
      • #027: Space Shuttle
    • impact on education: students (especially in the US) often don't have real-world experience in silicon until they graduate, and this project gives them an opportunity
  • Google's next steps
    • help others to replicate their successes with MPW programs
    • help others to avoid the traps they found along the way
    • fund more reproducible silicon research
    • advance high-level synthesis for accelerators, hardware/software co-design
Select a repo