Keynote from Microsoft by Pradeep Sindhu

In this monologue, Pradeep Sindhu discusses the fundamentals of high-performance interconnects for AI, emphasizing that AI's hyperparallelism makes interconnects central to the problem. He outlines three dimensions of interconnects: topology, physical layer, and logical layer, focusing on scale-out and scale-up bandwidth latency domains. Sindhu advocates for flat, high-radix topologies due to their ease of implementation and resilience to failures. He also addresses physical layer limitations, such as the bandwidth distance product of transmission mediums and transistor modulation rates, and logical layer considerations, including the importance of the end-to-end principle, bandwidth conservation, and low jitter. Sindhu proposes a straw man architecture with a host QPARE interface and a simplified Ethernet protocol stack, highlighting the need for a reliable packet transport layer and receiver-based congestion control. He concludes by stressing the importance of re-examining networking trade-offs, focusing on fundamental principles, and prioritizing implementation-driven standards over complex, committee-led approaches.