DARPA amps up effort to make AI power-conscious (www.theregister.com)

DARPA has launched ML2P (Mapping Machine Learning to Physics), a roughly two-year, $5.9M effort to make energy a “first-class” metric in ML by directly mapping model performance to physical energy use measured in joules. The program will fund teams to develop methods, tools and simulations that quantify the energy-per-performance tradeoff across the ML lifecycle, with deliverables (code, algorithms, docs, tutorials) required under a permissive open-source license. The goal is to move beyond accuracy-only optimization and enable principled, repeatable comparisons of how much utility models return for each joule spent—critical for battery-powered, edge and battlefield deployments. Technically, ML2P targets granular, end-to-end accounting that captures hardware, workload type, location and operating conditions—factors that hobble many prior, incomplete energy estimates for closed-source models. DARPA structures the program as a 24-month run (months 1–6 for experiment setup; months 7–24 for data collection) with a parallel government test-and-evaluation team validating results. If successful, ML2P could establish a gold-standard simulation/benchmark for energy-aware ML, accelerate energy-conscious model and hardware co‑design, and improve reproducibility and transparency across the community. Proposals are due Dec. 8, and DARPA hopes the outputs will be widely adopted for both research and deployment.