AI summary by TwinMind
Google's Open Source Hardware Dreams
Asianometry
9m 59s
Google's Open Hardware Strategy & Foundry Partnerships
- Google partnered with skywater, e-fabulous, and Global foundries to enable custom silicon design fabrication
- Skywater partnership: open sourced PDKs for 130nm process node
- PDK (Process Design Kit): design rules, physical limitations, simulators, IP libraries, design checkers
- Received as many GitHub stars in one week as hardware projects like riosis and chisel did over entire lifetimes
- Working on open sourcing skywater's 90nm fully depleted silicon on insulator process node
- Specialty process using insulator to prevent electron burrowing through gate
- August 2022: open sourced Global foundries 180nm process node PDK
- Enables applications for fabricating open source designs
- Significant milestone given Global foundries' industry prominence
Business Strategy Behind Open Source Hardware
- No significant open source silicon ecosystem exists for data center/consumer applications
- Google's internal/external hardware design (e.g., Google GPU) relies heavily on proprietary IP
- Proprietary IP contributes to exponentially increasing hardware design costs
- Goal: build world-class open source hardware design community to bend cost curve
- At Google's scale (billions of chips), replacing proprietary IP with open source alternatives provides significant financial/performance benefits
- Benefits extend beyond Google to entire ecosystem
Strategic Implementation Plan
- Focus on ease of use rather than ultimate performance
- No intention to compete with A-series Apple chips or Nvidia H100 GPUs
- 130nm process cannot achieve cutting-edge performance
- Enable ordinary users to download tools, design chips, share designs without expensive EDA software or NDAs
- Two-step execution:
- Open up and unify software/tool chains for hardware design
- Open sourcing PDKs from semiconductor foundries
- Backing tools like open Lane (automated flow bundling yosys, magic, open road)
- Takes RTL design abstraction to foundry-ready GDSII file
- Create on-ramps and incentives for talent adoption
- Tutorials and cloneable starter projects
- Free shuttle fab program for real chip fabrication
- Open up and unify software/tool chains for hardware design
130nm Process Capabilities & Applications
- Process debuted 2001-2002, approximately 18-20 years old
- Historical example: IBM PowerPC 970 chips
- 118 square millimeter chip with 58 million transistors
- Powered Power Mac G5 (introduced by Steve Jobs June 2003)
- Clock speeds 1.6-2 GHz
- Modern applications: IoT projects, sensor systems, microcontroller spaces
- Particularly suitable for low power applications
Featured Project: University of Michigan Neural Network Accelerator
- Ultra low power compute-in-memory neural network accelerator using analog principles
- Stores model weights as threshold voltages in on-chip volatile memory
- Linear multiplication: access transistor receives neural network input as voltage, output current maps to multiplication of input and weight
- Built using Sonos 130nm process
- Demonstrates real-world capability of 130nm technology
Academic Adoption & Open Source Hardware Evolution
- Strong academic uptake for open source hardware design
- Professors want to share code for work replication without NDA requirements for closed source EDA software
- Example: Brazilian team presented hardware accelerator for AV-1 video encoding
- Used open source tools (open Lane, skywater 130nm PDK)
- Open and royalty-free format
- Parallel to Linux/GCC evolution: started as "toys" before becoming business-critical
- Semi-analysis report: Apple converting dozen+ non-customer facing cores from ARM to RISC-V
Getting Started Resources
- IPython/Jupyter notebooks maintained by Google team
- Google Collaboratory: run notebooks in browser without installation
- Learning approach: modify beginner projects, observe effects
- Build custom silicon with Google website for advanced projects
- Shuttle program for fabrication opportunities
