Page 104 - AI Standards for Global Impact: From Governance to Action

AI Standards for Global Impact: From Governance to Action

Figure 46: Reports announced at the event

14�2 Innovations in environmentally efficient AI

This session explored technological innovations aimed at reducing AI’s environmental footprint
across the hardware-software stack, sharing perspectives on energy-efficient AI models,
neuromorphic computing, sustainable data centre operations, and the role of green energy.

University College London highlighted the daily impact of energy savings and the importance
of optimizing large models, while using small models when appropriate. The importance of
designing for efficiency from the outset was emphasized, with scenarios showing that task-
specific small models can achieve over 90% energy savings compared to large, multipurpose
models. Emerging architectures such as Mixture of Experts, retrieval-augmented generation,
neurosymbolic AI, and brain-inspired designs were discussed as promising pathways toward
sustainability.

Current digital computing architectures pose theoretical and practical limitations for
sustainability, leading participants to consider that spiking neural networks and neuromorphic
computing could be biologically inspired alternatives that drastically reduce energy use while
enhancing reliability.

The "Green AI" movement also emphasizes computational efficiency and transparency. While
large models like PaLM require massive amounts of computing resources, most environmental
impact comes from inference, not from training. The Hebrew University of Jerusalem noted that
inference operations account for 80-90% of all AI computation and are run billions of times per
day. There is a need for the community to report compute budgets and match model complexity
to task difficulty, considering that LLMs are not the solution for every problem.

Google’s end-to-end sustainability strategy includes:

• Model optimization (e.g. quantization, pruning, and knowledge distillation)
• Custom hardware (e.g. Ironwood TPU, 30x more efficient than its 2018 predecessor)

92