Abstract:
AI scaling is entering a new chapter. Progress is increasingly constrained not only by the availability of compute, but also by the cost of moving data: across GPUs, between processors and memory, and throughout ever-larger AI clusters. At the heart of this challenge lies a fundamental limitation of today's interconnect technologies. Copper links are power-efficient and reliable, but only over very short distances. Conventional optical links provide greater reach and bandwidth, but at the cost of higher power, complexity, and lower reliability.
Recent advances in low-power, high-bandwidth optical interconnects, including our work on microLED-based technologies, offer a path beyond this trade-off. Their impact could extend far beyond improving today's networks. Efficient optical connectivity could relax the physical boundaries that currently shape AI infrastructure, enabling a shift from tightly integrated, power-hungry GPU packages and racks toward smaller compute modules, shared memory pools, and flexible optical fabrics. Compute, memory, and network resources could then be pooled, composed, and reconfigured around the needs of individual workloads, paving the way toward a truly disaggregated AI datacenter.
Realizing this vision, however, will require sustained collaboration across the optics, networking, computer architecture, and AI-systems communities. In this talk, I will examine the opportunities and challenges ahead and suggest some research directions for the networking and broader systems communities.
Bio:
Paolo Costa is a Partner Research Manager at Microsoft Research and an honorary faculty member at Imperial College London. His research brings together hardware, optics, networking, and application-level expertise to take a cross-stack approach to developing next-generation data-center interconnects and architectures, leveraging the availability of emerging hardware and optical technologies.
Online meeting link:
