Flexible AI System Design and CXL

Presented by Chris Petersen (Meta) & Prakash Chauhan (Meta)
AI models and applications continue to evolve at an extremely rapid pace. This poses significant challenges for developing new AI systems to not only provide sufficient capabilities to keep pace with the evolution of these models, but also to do so efficiently. System architectures need to be able to scale and balance CPU performance, GPU or accelerator compute, memory bandwidth, memory capacity, front-end network bandwidth, and back-end network bandwidth. Solutions that can be easily adapted, evolved, and reconfigured by providing flexibility in system resources can help us keep pace. In this talk, we will explore these challenges in more depth, describe some possible solutions, and show how Compute Express Link (CXL) may be able to help us.