Simple Explanation (ELI5)

It’s like measuring not only how fast a car goes but also how much fuel it uses: the updated HEPscore benchmark now tells you both the speed and the energy efficiency of the supercomputers that simulate particle collisions.

Key Advantages & Current Limits

🟢 Pros: gives a direct energy‑efficiency metric (HEPscore/Watt) without requiring admin rights. 🟢 Pros: works across Intel, AMD, and ARM platforms via RAPL or fallback sensors/tools. 🔴 Cons: accuracy relies on the presence and quality of RAPL or system sensors. 🔴 Cons: adding power measurement adds a small computational overhead to the benchmark run.

Future Outlook & Takeaways

In the next 3‑5 years the energy extension will become a HEP computing standard, steering procurement toward greener hardware and enabling mandatory carbon‑footprint reporting for internationally funded projects. HEPscore/Watt metrics will be woven into workload‑management tools to auto‑optimize energy use while preserving scientific throughput.

How Can the HEPscore Benchmark Become Greener?

Article

100%

The world of particle physics relies on massive computing centers to simulate subatomic collisions. The HEPscore benchmark measures how fast these systems solve typical physics problems, but it never accounted for the energy they consume. Now researchers have added an extension that records power usage without opening servers or needing admin rights.

The trick lies in the RAPL (Running Average Power Limit) interface built into modern CPUs: it reads the chip's power draw directly. If RAPL isn’t available, the software switches to alternative methods, like reading system sensors or using external monitoring tools. To check accuracy, the team compared software data with hardware measurements taken from the power supply, finding a strong correlation across multiple platforms.

The outcome is an enriched HEPscore report featuring metrics such as HEPscore/Watt, showing not only speed but also energy efficiency. This helps research groups pick greener hardware and plan resources more sustainably, meeting the HEPiX Benchmarking Working Group’s requirements. In practice, supercomputers can now be compared not just for raw performance, but also for their ecological footprint.

Bonus

100%

The energy extension of HEPscore leverages the RAPL interface, introduced by Intel with the Sandy Bridge microarchitecture and later adopted by AMD. RAPL provides three main counters: Package, Core, and DRAM, each with a temporal resolution of a few microseconds. The software reads these registers via MSR (Model Specific Register) using privileged instructions, but the implemented library encapsulates the calls so that no root permissions are required, thanks to a user‑level driver.

When RAPL is unavailable, the system falls back to alternatives such as lm‑sensors on Linux or the Power Management API on Windows, which extract data from voltage and current sensors integrated on the motherboard. If neither is present, an external wattmeter (e.g., Yokogawa WT310) can be connected and timestamps synchronized with the HEPscore workload.

Validation was performed on three hardware classes: an Intel Xeon Scalable cluster, an ARM64 node with Neoverse CPUs, and an AMD EPYC server. For each platform, standard benchmarks (event generation, detector simulation) were run and energy values compared with wattmeter readings. The mean absolute error was below 3 %, demonstrating software measurement reliability.

Energy data integration occurs at the end of the benchmark, generating a JSON file with fields “energy_joules”, “duration_seconds”, and “hepscore_per_watt”. This format is compatible with HEPiX Working Group analysis tools, enabling cross‑platform comparisons and informed decisions on hardware procurement, workload optimization, and reducing the carbon footprint of high‑energy physics data centers.