Wattchmen: Watching the Wattchers -- High Fidelity, Flexible GPU Energy Modeling

A research team led by Brandon Tran has introduced Wattchmen, a breakthrough methodology for precisely measuring and predicting GPU energy consumption. As modern high-performance computing systems become increasingly energy-constrained, understanding application power usage is critical. Wattchmen addresses the limitations of current tools—which are often inaccurate or outdated—by constructing a per-instruction energy model using diverse microbenchmarks. This enables fine-grained energy consumption breakdowns for applications, achieving a mean absolute percent error (MAPE) of just 14% on NVIDIA V100 GPUs, compared to 32% for AccelWattch and 25% for Guser.

Wattchmen demonstrates remarkable flexibility across GPU architectures and cooling systems. It maintains similar accuracy on water-cooled V100s (15% MAPE) and extends effectively to newer NVIDIA architectures, including air-cooled A100 (11%) and H100 (12%) GPUs. The methodology was validated across 16 diverse workloads spanning GPGPU computing, graph analytics, HPC, and machine learning. Most impressively, when applied to real applications like Backprop and QMCPACK, Wattchmen's insights enabled developers to achieve energy reductions of up to 35%, demonstrating immediate practical value for energy-constrained computing environments.

The research, accepted to the prestigious 2026 International Conference on Supercomputing (ICS '26), represents a significant advancement in hardware energy modeling. By providing accurate, architecture-aware energy attribution at the instruction level, Wattchmen empowers developers and system administrators to make informed optimization decisions. This capability is increasingly vital as data centers face growing pressure to reduce their massive energy footprints while maintaining computational performance for AI training and scientific simulations.