Zeus

Zeus is an open source software library for measuring and optimising the energy consumption of deep learning workloads. Developed by the ML.ENERGY Initiative, it helps developers, researchers and infrastructure operators understand and reduce the energy used to train and run AI models. Its methods are based on peer reviewed research published at venues including USENIX NSDI, ACM SOSP and NeurIPS. Zeus is released under the Apache 2.0 licence, supporting efforts to improve the environmental sustainability of AI systems and to report their resource use more consistently.

The measurement component records the energy and power consumption of specific parts of a workload, such as a training step, an epoch or an inference request. Users can embed it in existing Python code or run it as a command line tool. Zeus supports NVIDIA and AMD GPUs, CPUs, DRAM, Apple Silicon and NVIDIA Jetson devices through a common interface, so that results can be compared across hardware. Measurements can be exported to Prometheus for continuous monitoring in production, and the same approach underpins the ML.ENERGY Benchmark and Leaderboard, which compare the energy consumption of publicly available AI models.

The optimisation component adjusts settings such as GPU power limits, batch size and GPU frequency to reduce energy use while taking account of the trade off with training time. Some optimisers profile a workload during its initial runs and then automatically select the most energy efficient configuration. Others target large model training across multiple GPUs by reducing energy spent on computations that do not affect overall completion time. A separate daemon allows hardware level controls to be applied without granting the main application elevated system privileges. Documentation, worked examples and a ready to use Docker image are available to support adoption.