What Happens When GPUs Talk?

Modern AI is often described in terms of bigger models and faster GPUs. But if GPUs are getting faster every year, why does training large AI models still take so long? What happens when hundreds or thousands of GPUs need to work together? And why do some of the world's most powerful accelerators spend time waiting instead of computing?

The answer lies in communication.

In this talk, we'll uncover the hidden infrastructure that powers large-scale AI systems and follow the journey of data as it moves across memory, devices, networks, and distributed systems. How do GPUs exchange information? What makes moving data so challenging? And why has communication become one of the biggest bottlenecks in modern computing?

Along the way, we'll build an intuitive understanding of concepts like collective communication, RDMA, and the Linux networking stack, without requiring any prior AI or networking knowledge.

What You'll Walk Away With :

Why adding more GPUs doesn't always make AI systems faster.

Why the cost of moving data is catching up with the cost of computing it.

How communication became a first-class problem in modern computing.

Where RDMA and modern networking fit into the AI story.

Why open-source infrastructure sits beneath nearly every large-scale AI system.