Watt Watchers

These days, computing accounts for a significant amount of the world’s total energy consumption. The power consumed by large data centers has been on the rise for many years, but the latest breakthroughs in artificial intelligence have resulted in a sharp uptick recently. But while the massive amount of energy being consumed is a growing concern, we certainly do not want to flip the switch off and go back to a pretechnological world.

Advances in hardware and algorithm design may help to alleviate this problem in the future. Or maybe not. We do not know the future, so we must address this problem as best we can with the tools we have available to us today. Some promising options include software-based methods — such as dynamic voltage and frequency scaling, power-aware scheduling, and power capping —that significantly lower power consumption without negatively impacting computing performance.

However, these software-based solutions have to act very fast to be effective. Without new power measurements coming in at a sub-millisecond frequency, these tools cannot adjust fast enough to maximize the energy efficiency of the system’s computations. Unfortunately, even state-of-the-art power meters generally do not have the accuracy or the speed needed to be useful for this purpose.

A group led by researchers at the Netherlands Institute for Radio Astronomy has developed a very fast and accurate, open source power measurement tool called PowerSensor3. This device was born of necessity — astronomers know a thing or two about supercomputing, to put it mildly. So they built this device that can provide fine-grained measurements of the power consumption of everything from CPUs and GPUs to FPGAs, SSDs, and dedicated AI accelerators.

The tool offers real-time power measurements at an impressive 20 kHz sampling rate, which is a major leap from previous tools, which typically top out at under 3 kHz. This higher sampling rate is essential for profiling fast operations, such as GPU kernel executions, which often occur in sub-millisecond intervals.

Designed with flexibility and modularity in mind, PowerSensor3 consists of a baseboard centered around an STM32F411 Black Pill microcontroller that can support up to four interchangeable sensor modules. Each module is capable of measuring both voltage and current, with optical and magnetic isolation to eliminate interference from nearby components or noisy environments.

The sensor modules use Hall effect sensors (Melexis MLX91221) for current measurement and optically isolated voltage sensors (Broadcom ACPL-C87B), ensuring both safety and precision. A key feature of the modules is the inclusion of remote sensing, which allows voltage to be measured directly at the component rather than the sensor input, avoiding measurement errors due to resistance in cables.

Both the hardware — licensed under CERN-OHL-P v2 — and the supporting firmware and host software — licensed under Apache-2.0 — are freely available. This openness enables wide adoption and customization, which is particularly valuable for researchers, developers, and institutions looking to build more energy-efficient computing platforms without spending thousands on proprietary metering tools. A complete PowerSensor3 kit with three sensor modules can be built for little more than $100 in parts.

As our need for large-scale computing increases in the years to come, tools like PowerSensor3 may offer us a way to maintain performance without burning through our limited energy resources unchecked.