The rapid growth of AI workloads introduces substantial challenges to cloud infrastructure, particularly in memory scalability, bandwidth bottlenecks, and efficient data movement. Compute Express Link (CXL) offers a new paradigm with its high-speed, low-latency, memory-coherent interconnect between CPUs, GPUs, and memory devices. This paper explores leveraging CXL memory controllers to create scalable, flexible, and efficient AI cloud infrastructure. We propose an architecture based on CXL memory pooling and disaggregation, benchmark its performance, and present results demonstrating significant improvements in memory utilization, latency, throughput, and energy efficiency.
Upendra Sharma (Wed,) studied this question.