Google's Liquid Cooling at Hot Chips 2025
16 days ago
- #liquid-cooling
- #datacenter
- #AI
- Liquid cooling is increasingly used in datacenters due to high power draw and heat output from modern chips, especially for machine learning applications.
- Google has developed a datacenter-scale liquid cooling solution for TPUs, featuring Coolant Distribution Units (CDUs) that span racks and allow maintenance without downtime.
- Google's cooling system uses a split-flow cold plate and bare-die setup for TPUv4, improving heat transfer efficiency compared to previous designs.
- Liquid cooling reduces cooling-related power consumption, with Google's pumps using less than 5% of the power required by air cooling solutions.
- Maintenance challenges like microbial growth and leaks are addressed in datacenters through extensive testing, alert systems, and preventative measures.
- The AI boom is driving the adoption of liquid cooling in datacenters, with companies like Nvidia and Rebellions AI showcasing water-cooled hardware at Hot Chips 2025.