Achieving a sustainable future for AI

1. Emphasize data quality over data quantity. Smaller datasets require less energy for training and have lighter ongoing compute and storage implications, thereby producing fewer carbon emissions. Studies show that many of the parameters within a trained neural network can be pruned by as much as 99%, yielding much smaller, more sparse networks.
2. Consider the level of accuracy truly needed to solve for your use case. For instance, if you were to fine-tune your models for a lower accuracy intake calculation, rather than compute-intensive FP32 calculations, you can drive significant energy savings.
3. Leverage domain-specific models and stop re-inventing the wheel. Orchestrating an ensemble of models from existing, trained datasets can give you better outcomes. For example, if you already have a large model trained to understand language semantics, you can build a smaller, domain-specific model tailored to your needs that taps into the larger model’s knowledge base, resulting in similar outputs with much more efficiency.
4. Balance your hardware and software from edge to cloud. A more heterogenous AI infrastructure, with a combination of AI computing chipsets that meet specific application needs, will ensure you save energy across the board, from storage to networking to compute. While edge device SWaP (size, weight, and power) constraints require smaller, more efficient AI models, AI calculations closer to where data is generated can result in more carbon-efficient computing with lower-power devices and smaller network and data storage requirements. And, for dedicated AI hardware, using built-in accelerator technologies to increase performance per watt can yield significant energy savings. Our testing shows built-in accelerators can improve average performance per watt efficiency 3.9x on targeted workloads when compared to the same workloads running on the same platform without accelerators. (Results may vary.)  
5. Consider open-source solutions with libraries of optimizations to help ensure you’re getting the best performance from your hardware and frameworks out of the box. In addition to open source, embracing open standards can help with repeatability and scale. For example, to avoid energy-intensive initial model training, consider using pre-trained models for greater efficiency and the potential for shared/federated learnings and improvements over time. Similarly, open APIs enable more efficient cross-architecture solutions, allowing you to build tools, frameworks, and models once and deploy everywhere with more optimal performance.