Optimizing your AI/ML workloads for sustainability

Up to 90% of your model's carbon footprint comes from inference. Learn key strategies to right-size workloads and slash your environmental impact in the cloud.

#1about 3 minutes

Understanding the carbon footprint of large AI models

The increasing size and complexity of models like GPT-4 result in a significant carbon footprint, with training a single model consuming more energy than a lifetime of car usage.

#2about 3 minutes

Reducing emissions with the cloud's shared responsibility model

Migrating workloads to a cloud provider like AWS can reduce energy usage by up to 80%, operating under a shared responsibility model where AWS manages the cloud's sustainability.

#3about 5 minutes

Optimizing the ML lifecycle starting with problem framing

Begin the ML lifecycle sustainably by using purpose-built hardware, pre-trained models from marketplaces, and managed AI services to avoid redundant computation.

#4about 6 minutes

Implementing sustainable data processing and storage strategies

Reduce your workload's environmental impact by using tiered storage with lifecycle policies, efficient compression algorithms, and optimized file formats like Parquet.

#5about 2 minutes

Selecting purpose-built hardware for ML workloads

Improve energy efficiency by selecting specialized silicon for different ML phases, such as AWS Trainium for training, Inferentia for inference, and Graviton processors for general workloads.

#6about 3 minutes

Adopting sustainable practices for model development

During model development, define acceptable performance criteria to prevent over-training, choose energy-efficient algorithms, and use pre-trained models to reduce computational waste.

#7about 5 minutes

Optimizing the high-cost model deployment and inference phase

Since deployment accounts for 90% of ML costs, focus on right-sizing inference environments by smoothing traffic peaks with queues and negotiating flexible service level agreements.

#8about 4 minutes

Applying the AWS Well-Architected Framework for sustainability

Use the sustainability pillar of the AWS Well-Architected Framework to get recommendations, such as choosing regions with higher renewable energy usage to lower your carbon footprint.

#9about 2 minutes

Measuring and tracking your workload's carbon footprint

Actively monitor your environmental impact using tools like the AWS Customer Carbon Footprint tool and normalize metrics to track efficiency gains as your workload scales.

#10about 4 minutes

Applying AI and ML to solve global sustainability challenges

Leverage AI/ML for positive environmental impact by using open datasets like the Amazon Sustainability Data Initiative to address challenges in conservation, climate risk, and the circular economy.

#11about 6 minutes

Real-world case studies of ML in environmental conservation

Explore how organizations use ML on satellite imagery to monitor oceans for oil spills and deploy ML at the edge with rugged devices to protect forests and endangered species.

#12about 1 minute

A call to action for building sustainable technology

Take action by starting sustainability conversations, exploring open data initiatives like ASDI, and applying the AWS Well-Architected Framework to your own projects.