New System Cuts AI Power Use 36% — Without Slowing Down

Here's a real-world problem: AI-powered devices — like cameras that read license plates, smart doorbells, or factory sensors — need to respond quickly, but they also burn electricity. Most of them are set to use a fixed amount of computing power, even when the task is easy. That's like driving a car at full speed all the time, even on empty streets. It wastes energy.

PRISM, a new system from researchers at TU Darmstadt and others, fixes that. Think of it as a smart thermostat for AI. It constantly checks how difficult each job is — an easy, bright, clear license plate needs less brainpower than a blurry, dark, angled one. PRISM then adjusts two things in real time: which AI model to use (a small fast one or a big accurate one) and how much computing power to give it. All of this happens automatically, in split seconds, without human help.

The results are striking. In a test with over 52,000 real requests on a license plate recognition pipeline, PRISM cut energy use by 36% compared to the best fixed setup. It also used less than half the computing power — all while catching plates just as accurately. For the second stage of the pipeline, it nearly matched the best quality while using less CPU.

Why should you care? Because AI is popping up everywhere — in traffic cameras, delivery drones, smart fridges, and more. If those devices use less electricity, they can run on smaller batteries, cost less to operate, and reduce carbon emissions. PRISM shows that we don't have to choose between speed, accuracy, and efficiency. We can have all three, if the software is smart enough to adapt on the fly.