Skip to content
Conference

Exploring Unified Shared Memory on AMD MI300A through Real-World Use Cases

Sep 2026 · eScience · pp. 79-87 · 0 citations · 31 references

Abstract

Unified Shared Memory (USM) has emerged as a promising programming model for simplifying data management in heterogeneous CPU-GPU HPC systems with the AMD MI300A APU. However, its performance implications in real-world HPC applications remain insufficiently understood. In this work, we go beyond microbenchmarking by evaluating the impact of USM on performance and programmability in production-level HPC workloads from fields including multigrid solver, time series mining, and plasma physics simulation. Our results show that USM can achieve performance comparable to, or even outperform, explicit memory management while simplifying code development. We conduct an application-driven study using multiple real-world case studies, integrating USM into existing HIP-based implementations and comparing against explicitly managed memory. We analyze execution behaviour through detailed profiling, considering factors such as data locality, communication-computation overlap, workload distribution, asynchronous execution, and reduced precision. We further demonstrate that USM is sensitive to data locality and memory access patterns, while also showing that USM can improve performance in scenarios where eliminating explicit data transfers between CPUs and GPUs reduces communication overhead. Overall, this work provides practical insights into using USM effectively on AMD MI300A APU, bridging the gap between programmability and high performance in real HPC applications.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.