Accurate and Convenient Energy Measurements for GPUs: A Detailed Study of NVIDIA GPU's Built-In Power Sensor

The Graphical processing unit (GPU), initially developed for graphics rendering, has emerged as the go-to accelerator for high throughput and parallel workloads, spanning scientific simulations to AI, thanks to its performance and power efficiency. Given that 6 out of the top 10 fastest supercomputers in the world use NVIDIA GPUs and many AI companies each employ 10,000's of NVIDIA GPUs, an accurate understanding of GPU power consumption is essential for making progress to further improve its efficiency. Despite the limited documentation and the lack of understanding of its mechanisms, NVIDIA GPUs' built-in power sensor, providing easily accessible power readings via the nvidia-smi interface, is widely used in energy efficient computing research on GPUs. Our study seeks to elucidate the internal mechanisms of the power readings provided by nvidia-smi and assess the accuracy of the power and energy consumption data gathered from this method. We have developed a suite of micro-benchmarks to profile the behaviour of nvidia-smi power readings and have evaluated them on over 70 different GPUs from all architectural generations since power measurement was first introduced in the 'Fermi' generation of GPU. We have identified several unique and unforeseen problems in terms of power/energy measurement using nvidia-smi, for example on the A100 and H100 GPUs only 25% of the runtime is sampled for power consumption, during the other 75% of the time, the GPU can be using drastically different power and nvidia-smi and results presented by it are unaware of this. This along with other findings can lead to a drastic under/overestimation of energy consumed, especially when considering data centres housing tens of thousands of GPUs. We proposed several good practices that help to mitigate these problems. By comparing our results to those measured from an external power-meter, we have reduced the error in the energy measurement by an average of 35% and in some cases by as much as 65% in the test cases we present. We have encapsulated our learning, our micro-benchmark and measurement good practice into a Python library for public use.