Michigan’s $1.83m project targets the energy cost of the whole AI workflow

A University of Michigan-led project announced on 27 October will try to reduce AI energy consumption by optimising the entire computational workflow. The $1.83 million, two-year DARPA contract supports E-CALC, a collaboration involving Michigan, the University of Illinois Urbana-Champaign and BAE Systems. Its ambition is a reduction of up to 50%; that is a target for the research.[1]

The attractive idea is to stop treating a model’s training run or an individual response as the whole energy story. Preprocessing, training and inference interact. A choice that makes one stage cheaper can move work into another, so the best local setting need not produce the lowest total bill.[1]

Power is a rate; energy is the total

The distinction can be shown without a proprietary benchmark. In a hypothetical constant-power calculation, a 400-watt workload running for one hour consumes 0.4 kilowatt-hours. Lowering it to 300 watts would use 0.3 kilowatt-hours if it still finished in an hour. If completion instead took 90 minutes, consumption would rise to 0.45 kilowatt-hours. Lower instantaneous power would then mean more energy for the task.

These are illustrative calculations, not measurements of E-CALC or a particular accelerator. They explain why a useful optimiser must account for both watts and completion time. It also needs a consistent definition of completion: producing a weaker model more cheaply is a different trade-off from delivering the same required result with less energy.

Michigan’s existing Zeus software makes that trade-off explicit. Its power-limit optimiser profiles configurations and allows selection based on energy, time, a combined cost or a maximum-slowdown constraint. Integration tells the optimiser where training steps and epochs begin and end; changing GPU power limits also requires appropriate system privileges.[2]

The Zeus measurement tools separately let developers delimit a section of computation and record elapsed time and device energy. That is more useful for comparing alternatives than a single headline power reading. It can also make the boundary of a claim visible: measured device consumption is not automatically every electrical load associated with a service.[3]

The value of a common accounting boundary

E-CALC plans measurement and optimisation tools spanning the AI lifecycle, with open-source outputs. The university describes a move away from repeated trial-and-error configuration towards a more systematic understanding of how choices affect consumption.[1]

For an operator, the most informative future demonstration would hold the required output constant and disclose the full comparison: which stages were counted, how long they ran and whether any work was shifted elsewhere. A saving during training is commercially different from a saving on every later inference request; neither should obscure the other.

There is also a scale question. As a simple scenario, halving energy per task while tripling task volume would still raise total task energy by 50%. Efficiency and absolute demand are separate quantities. A successful E-CALC result could improve the former without guaranteeing the latter falls.

That is why this project is worth following before its percentage target becomes a result. Its potentially durable contribution is a method for making AI’s energy trade-offs comparable, so a team can explain what it saved, where it saved it and what performance it retained.

Report an error or suggest a correction