Scale Out Networks and Scale Up Architectures with CPO

https://www.youtube.com/watch?v=vmXzRsWW9S0

Summary

TL;DR — The presentation discusses the evolution of network architectures from scale-up to scale-out, highlighting the limitations of copper interconnects and the advantages of optical solutions like Co-Packaged Optics (CPO). CPO enables denser, more powerful compute clusters by moving optical interconnects closer to the processing units, leading to significant improvements in efficiency, reliability, and power consumption.

Key points

• Traditional scale-up architectures are limited by copper interconnects, necessitating a shift towards optical solutions for higher density and performance.

• Co-Packaged Optics (CPO) allows for optical interconnects to be integrated directly into the compute package, enabling scale-up within the rack and even the row.

• CPO significantly improves data center efficiency by reducing power consumption, latency, and enabling longer training jobs without link flaps, leading to higher overall training efficiency (e.g., 90% improvement on a 24,000 GPU cluster).

• Broadcom is advancing CPO technology through multiple generations, from Gen 1 (Humboldt) to Gen 3 (Davidson), focusing on increased bandwidth (200 gig per lane), reduced power, and improved reliability.

• The industry is working collaboratively to develop a robust CPO ecosystem, including standardized components like ELSFP and addressing manufacturing challenges.

• CPO offers a unified protocol (Ethernet) for both scale-up and scale-out, simplifying network design and management.

Takeaway — Co-Packaged Optics (CPO) is a critical technology for enabling the next generation of high-performance computing by overcoming the limitations of copper interconnects and delivering substantial gains in efficiency and reliability.