Liquid Cooling Impact on Backend Network Design for AIML Systems

Learn how liquid cooling impacts AI/ML backend network design in this 15-minute conference talk from Don Barnetson of Credo at Open Compute Project. Explore how reduced interconnect lengths enable new possibilities in optical vs copper interconnect implementations, leading to significant cost savings of $1000 per GPU, power reductions of over 15W per GPU, and 10x improvement in network reliability. Examine data center architecture, rack elevation considerations, and interconnect co-design principles that unlock these optimization opportunities. Cover key topics including backend network design fundamentals, networking protocols, optical systems, link flap challenges, rack density optimization, and zero flap AACS implementations.