AI Data Centers Learn to Conserve Energy from Bitcoin Miners
Alexa and Google Assistant may provide instant answers to our queries, but behind the scenes, their power-hungry data centers are consuming massive amounts of electricity. In a recent experiment, Luxor Energy and Bentaus demonstrated that AI data centers can learn to conserve energy by curbing their power draw during peak hours.
The companies used a single Nvidia B200 chip, which normally draws 14.3 kilowatts of power, but the software controlled it to reduce its consumption to around 25% in just half a second without affecting performance.
This 'demand response' technique was pioneered by Bitcoin miners, who could shut down their operations during peak hours and receive payment for doing so. The same principle can be applied to AI data centers, which currently account for an estimated 9.5% of US electricity consumption and are expected to reach up to 15.3% by 2030.
The success of this experiment raises hopes that large-scale energy conservation is possible in the AI industry. However, it also highlights the challenges facing grid planners, who must balance the increasing demand for power with the constraints on transmission lines and infrastructure upgrades.