Speaker
Description
Since 2023, HEPScore has been well established in the High‑Energy Physics community as the standard benchmark for WLCG sites. With GPU applications increasingly important in HEP and a growing availability of GPUs in the Grid, the need for a representative GPU benchmark is rising accordingly. This is particularly relevant in view of the High‑Luminosity LHC (HL‑LHC) era, which will substantially increase computing demands across the WLCG, posing challenges in terms of cost, energy consumption, and environmental sustainability.
The HEPiX Benchmarking Working Group has therefore intensified the development of a GPU counterpart, HEPScore4GPU. This effort builds on existing software frameworks and follows the successful HEPScore approach by using real HEP applications as benchmark payloads. GPU benchmarking introduces additional challenges, including heterogeneous architectures, workload‑specific optimizations, and rapid hardware and software evolution.
In parallel, the benchmark suite is extended with lightweight modules to collect server utilization and power metrics, enabling a more comprehensive evaluation of performance and power efficiency across diverse computing environments.
This contribution analyzes these challenges and presents the current benchmarking strategy, the development status, and power‑aware measurements in WLCG.