The New Stack Podcast Google AI Infrastructure PM On New TPUs, Liquid Cooling and More
13 snips
May 13, 2025 Chelsie Czop, Senior Product Manager for AI Infrastructure at Google Cloud, dives into cutting-edge developments in AI hardware. She discusses the impressive new Ironwood TPUs, boasting 42.5 exaflops, and the advancements in liquid cooling, essential for managing heat. Chelsie clarifies the ongoing debate between using TPUs or GPUs, noting significant performance boosts for some users. Moreover, she highlights the collaboration with DeepMind to stay ahead of evolving model architectures and the sustainable innovations shaping the future of data centers.
AI Snips
Chapters
Transcript
Episode notes
Challenges in Scaling TPU Power
- Increasing TPU power demands balancing power, thermal limits, and interconnectivity.
- Performance growth is incremental but leaps significantly compared to early generations.
Liquid Cooling Advances TPU Efficiency
- Google's fourth-generation liquid cooling deployed since 2017 greatly improves efficiency at scale.
- Liquid cooling innovations include visible pipe placement and design tweaks reflecting lessons learned.
Photonics Connects TPU Chips
- TPU pods connect over 9,000 chips using photonics-based interchip interconnect.
- This approach enables high scalability and fast communication within TPU pods.