Green AI techniques for reducing energy consumption in AI systems
Sunawar Khan, Naila Sammar Naz, Tehseen Mazhar, Muhammad Usman Tariq, Tariq Shahzad, Sghaier Guizani, Habib Hamam
Abstract
This systematic review synthesizes current evidence on energy-reduction techniques across algorithmic, hardware, and infrastructure layers of AI systems. Model compression and knowledge distillation (e.g., DistilBERT) deliver ∼60% faster inference with ∼40% fewer parameters while retaining ∼97% of baseline performance. Low-precision computation (quantization) yields up to ∼50% energy reductions, and architecture-level strategies—such as neural architecture search and depthwise-separable convolutions in MobileNetV2—significantly lower compute and memory demand. Specialized accelerators (TPUs) and neuromorphic hardware further improve efficiency, while data-center measures (advanced cooling, virtualization, renewable integration) reduce system-level consumption. For generative-AI workloads, distillation, quantization, efficient architectures, and accelerator-optimized inference remain the primary pathways to lowering both training and inference energy. Across studies, recurring gaps include inconsistent energy-metric reporting, limited standardized benchmarks, and a dominant focus on accuracy over efficiency. Regulatory progress is uneven: the EU has introduced stronger transparency requirements, whereas comparable obligations are not yet global. Review limitations include heterogeneous methodologies and incomplete transparency artifacts, which restrict cross-study comparability. Future research directions include algorithm–hardware co-design, neuromorphic methods, energy-harvesting AI devices, improved data-center operations, and explainable-AI tools to support reliable, energy-aware deployment at scale.