Independent studies on how AI workloads use infrastructure, and how that use can be measured, attributed, and improved across workloads, infrastructure, and the interfaces between them. The research supports the working areas of the call for collaboration, and anyone can contribute. Read the call for collaboration and sign →
Each paper is published by its authoring institution. Aitra hosts copies alongside the underlying data and reproduction materials, and links to the original source. The first studies measure the energy cost of AI inference.
Tsinghua University · supported by Linux Foundation Research and the SODA Foundation through its Aitra project
Extends the measurement boundary from steady operation to the full service lifecycle — startup, formal requests and resource release — across 13 models in text chat, code completion and image-text QA.
Tsinghua University
Steady-state energy per output token for 8 models, with a tensor-parallel and batch-size sweep across three production scenarios.