USM: A first look
23 September 2025
Paul Bartholomew investigates the cost of using USM compared with controlling data movements manually.
Major GPU vendors have recently introduced Unified Shared Memory (USM) programming models that simplify accelerator programming by presenting a single memory space that is accessible to both the host and device. With appropriate hardware support, this can reduce the effort required to port codes to, or develop them for, accelerators.
As part of the ARCHER2 project “Porting x3d2 to AMD GPUs”, I have been investigating the performance impact of different optimisations to control host-device data transfers using OpenMP.
In my article USM: A first look, I compare these with using Unified Shared Memory available on modern HPC GPUs and conclude that the simpler approach, based on USM, achieves nearly the same performance as the best performing solution using manual memory management.
Links
Read Paul's article: USM: A first look
ARCHER2 Embedded CSE (eCSE) projects support the improvement of software on ARCHER2, the UK national HPC service operated by EPCC. See: https://www.archer2.ac.uk/ecse