Cloud-Hardware Co-Design for Memory Bandwidth-Bound HPC Workloads: Performance and Characteristics of Azure HBv5 Virtual Machines

A majority of modern High-Performance Computing (HPC) workloads are memory bandwidth-bound, with their performance constrained by limited access to the main memory. While High Bandwidth Memory (HBM) has enabled multi-terabyte-per-second sustained bandwidths on accelerator devices such as GPUs, CPUs have traditionally lagged due to legacy memory subsystem designs. Despite the rise of accelerated computing, the majority of HPC applications still depend on CPUs. Integrating HBM directly with CPUs offers a promising path to address their bandwidth bottlenecks, although earlier efforts have struggled with limited capacity, insufficient architectural support, and high software complexity.