Paper Title: AI Memory at Meta: Challenges and Potential Solutions

Paper Abstract: Memory requirements for AI workloads are growing at a rate that is outpacing the underlying memory technologies. In this presentation, we will discuss AI workloads at Meta, the memory challenges we observe, and their impacts on system architectures. We will then compare several potential solutions and their requirements including CXL memory and memory accelerators. Finally, we will close with a call to action to help us address these challenges in the long term!

Paper Author: Chris Petersen, Hardware Systems Technologist, Facebook

Author Bio: Chris Petersen is a Hardware Systems Technologist leading hardware technology and infrastructure roadmaps at Facebook. Chris has been designing and building servers, storage, and datacenter solutions for over 17 years. He has worked on a wide variety of architectures, including scale-up, scale-out, and accelerators. Chris is a board member of NVM Express (NVMe), and a board member of Compute Express Link (CXL). He has evangelized CXL, NVMe, storage, and accelerator solutions at multiple conferences, including the Open Compute Summit, NVM Workshop, SNIA Summit, and Flash Memory Summit.