Synergizing CXL with Unified Memory for Scalable GPU Memory Expansion

Heterogeneous computing, which combines the power of CPUs and GPUs, is increasingly important in today's data-centric era. However, it also presents significant challenges, particularly in terms of GPU memory management. Compute eXpress Link (CXL) technology, which is supported by industry leaders, offers enhanced memory expansion capabilities. Alongside this, NVIDIA's Unified Memory (UM) system simplifies the process of sharing data between CPUs and GPUs. We explore the role of CXL memory in various scenarios within the UM framework while simultaneously highlighting issues such as the handling of page faults at the software level.