DARPA Backs Project to Measure AI Energy Use at Every Stage of the Workflow

The artificial intelligence boom is creating an energy monster. The indiscriminate integration of machine learning into everything from from our energy grids to our electric toothbrushes – no, really – has created a global race to build up data center capacity as quickly as humanly possible, causing energy demand projections to skyrocket and placing unprecedented strain on the world’s energy grids. By 2030, data center energy consumption is expected to more than double from its 2024 consumption level, which stood at about 420 terawatt-hours (TWh).Proponents of artificial intelligence argue that, eventually, AI will save more energy than it consumes as its integration into nearly every economic sector imaginable makes all kinds of processes more efficient. “Finding new materials, catalysts or processes that can produce stuff more efficiently is the sort of ‘’needle in a haystack’ problem that AI is ideally suited to,” the Financial Times reported last year in an article musing about “How AI might save more energy than it soaks up.”But in the meantime, the sector’s water and energy use is continuing to skyrocket at a time when humanity simply cannot afford it. Even though the world is racing to add clean energy capacity faster than at any previous time in human history, clean energy production growth is simply no match for overall energy demand growth. Artificial intelligence is already causing a massive buildout of natural gas-powered plants and is set to extend the dominance of fossil fuels in the global energy mix. As a result of this demand boom, paired with decades of delayed climate action, surpassing the emissions goals set in the Paris agreement is now an inevitability. So instead of waiting for AI to make the world more efficient, why not make AI itself more efficient? A team of researchers at the University of Michigan Engineering has proposed a new system for designing more efficient AI software. “While computer chips are getting faster, misalignments between software and hardware waste massive amounts of electricity—up to 30% during AI model training alone,” reports Michigan Engineering. “More energy-efficient AI brings down electricity demand, reducing the cost of AI and extending battery life for AI-enabled devices.” As such, researchers from University of Michigan Engineering are partnering with colleagues at the University of Illinois Urbana-Champaign (UIUC) and defense contractor BAE Systems to develop a new efficiency-focused software system with better energy tracking capabilities. One of the biggest challenges facing the AI energy problem today is the extreme opacity of the sector and insufficient measuring and reporting of resource use. Researchers don’t really know how much energy AI is using in the present moment. But we do know that it’s a whole lot.The new concept being proposed would enhance the ability to track energy use with the aim of increasing efficiency. “Rather than relying on expensive, time-consuming trial-and-error testing across thousands of hardware and software configurations, the team’s proposed system will enable developers to estimate energy usage across every stage of an AI workflow—from data processing and model training to real-world deployment,” Michigan Engineering writes. The technology could be particularly important in military and defense settings, “where field equipment and remote computing systems must run complex AI algorithms under strict battery and power limits.” As such, the project, E-CALC: Energy Calculus for ML Lifecycle Optimization, will be funded for two years at $1.83 million by the Defense Advanced Research Projects Agency (DARPA). “AI is becoming part of everyday life, business, science and national security, but it requires enormous and rapidly growing amounts of electricity. If that energy use continues to rise unchecked, it could increase costs, strain power infrastructure, and make advanced AI harder to deploy,” says University of Michigan associate professor of computer science and engineering and principal investigator for E-CALC Mosharaf Chowdhury. “This project tackles a practical question: how can we get widespread benefits of AI with minimum energy?” By Haley Zaremba for Oilprice.comMore Top Reads From Oilprice.comEuropean Gas Tops 101 Euros for First Time Since 2022Japan to Overtake China As World’s Top LNG Importer This YearGasoline Up 34% and Heating Oil Up 74% as Voters Head to the Midterms