Study finds AI efficiency tweaks save only 9.1% net energy once rebound is counted

The paper measured training emissions with Carbontracker version 2.4.2. Training GPT-3 produced as much as 590 kgCO2, against 59 kgCO2 for YOLOv5, showing how much model scale drives the footprint.

The authors report that the information technology industry reached a synergy index of 0.88 through AI optimization. Dedicated hardware responded in 1.5 s, three times faster than the cloud. They conclude that large models need optimization to prevent rebound, and that real-time feedback and hardware scheduling are key to carbon neutrality.

The framework has four parts. An energy consumption perception index is introduced and emissions are tracked in real time with Carbontracker version 2.4.2. Multi-task learning predicts energy demand and emissions from multi-source data.

The rebound effect is quantified and corrected using an elasticity coefficient model. Reinforcement learning then optimizes resource allocation under environmental constraints, and a closed-loop feedback mechanism ties the pieces together.

Results show a wide gap between gross and net savings: the 45% efficiency gain from dynamic batch processing contrasts with a 9.1% net energy-saving rate for knowledge distillation. The authors say large-scale models consume a lot of energy and need optimization to prevent rebound.

On deployment, dedicated hardware responded in 1.5 s, three times faster than the cloud. The authors name real-time feedback and hardware scheduling as the keys to achieving carbon neutrality.