Carbon Reduction Plan

Executive Summary

This plan outlines BeeHealthy strategic approach to significantly reduce our carbon footprint, with a primary focus on emissions generated by AI workloads in the public cloud.

We do recognize the increasing energy demands of AI and the shift away from traditional emission sources like business travel or office space, this plan emphasizes optimizing cloud resource utilization, leveraging green cloud infrastructure, and fostering a culture of sustainable AI development as the AI workloads and consumption is growing rapidly.

Baseline Assessment and Monitoring

Establish a comprehensive baseline:

• Cloud Emissions: Utilize cloud provider tools (AWS Carbon Footprint Tool, Azure Emissions Impact Dashboard, Google Cloud Carbon Footprint as we are cloud agnostic) to track and report Scope 2 and Scope 3 emissions specifically from AI services (e.g., machine learning platforms, GPU instances, data storage for AI).
• Traditional Emissions (for context): Continue to monitor office energy consumption (Scope 1 & 2) and business travel (Scope 3) to demonstrate the shift in emission profiles and ensure these areas remain optimized.

Implement continuous monitoring: carbon emission tracking is being integrated into our FinOps and MLOps practices to provide real-time visibility into AI-related emissions as the AI usage and use cases are growing. The emissions are monitored as a part of our FinOps practice and processes.

Cloud Infrastructure Optimization for AI

Right-sizing and efficiency:

• Instance Selection: Prioritize energy-efficient instance types (e.g., newer generation CPUs/GPUs, ARM-based processors where suitable) for AI workloads.
• Processing location: Processing location impacts heavily on the electricity and cooling related CO2 emissions, ensuring using best possible location for processing while ensuring the data residency, privacy and other related compliance requirements
• Auto-scaling and Serverless: Implement aggressive auto-scaling and leverage serverless AI services to ensure resources are only consumed when actively needed, minimizing idle compute capacity standing by.
• Resource Scheduling: Optimize scheduling of AI training and inference jobs to utilize off-peak hours when renewable energy availability might be higher.

Data Storage Optimization:

• Lifecycle Management: Implement intelligent data lifecycle policies to move less frequently accessed AI training data to colder, more energy-efficient storage tiers and using generic, commercial or open sources models where possible in order to avoid model training overhead.
• Data Deduplication and Compression: Employ techniques to reduce the overall storage footprint of AI datasets.

Network Optimization: Minimize data transfer between regions and availability zones, as data egress can contribute to emissions.

Sustainable AI Development and Operations

Model Efficiency:

• Smaller Models: Explore and prioritize the use of smaller, more efficient AI models that achieve comparable performance with less computational power.
• Quantization and Pruning: Implement model optimization techniques like quantization and pruning to reduce model size and inference energy consumption.
• Transfer Learning: Leverage pre-trained models and transfer learning to reduce the need for extensive training from scratch.

Algorithm Selection: Favor algorithms known for their computational efficiency.

Experiment Tracking: Utilize MLOps platforms to track the energy consumption of different model training runs, enabling developers to make informed choices about model efficiency.

Responsible Data Practices for AI Development:

• Data Minimization: Use minimal synthetic data for AI model training, quality assurance in order to use minimal capacity.
• Synthetic Data Principle: Use of synthetic data only in order to respect privacy.

Green Cloud Adoption and Renewable Energy

• Cloud Provider Selection: Prioritize cloud providers and regions that have a strong commitment to renewable energy and offer carbon-neutral or carbon-negative services.
• Carbon-Aware Regions: Deploy AI workloads in cloud regions powered by a higher percentage of renewable energy, where feasible and aligned with data residency requirements.
• Advocate for Green Cloud: Engage with cloud providers to advocate for further investments in renewable energy and transparent reporting of carbon emissions.
• Advocate for Green Coding: Ensure code efficiency by using optimal tooling and by implementing green coding practices.

Employee Engagement and Education

• Training and Awareness: Educate all employees on the environmental impact of their work and best practices for sustainable AI.
• Internal Guidelines: Develop and disseminate internal guidelines for carbon-efficient AI development and cloud resource management.
• Incentivise Green Practices: Recognition for teams and individuals who demonstrate significant contributions to carbon reduction in AI.

Reporting and Transparency

• Regular Reporting: Publish regular internal reports on our cloud AI carbon footprint and progress towards reduction targets.
• External Communication: Communicate our commitment to sustainable AI and our carbon reduction efforts to stakeholders, customers, and the public.

Continuous Improvement

• Stay Informed: Continuously monitor advancements in green computing, sustainable AI research, and cloud provider offerings.
• Iterative Approach: Treat this plan as a living document, regularly reviewing and updating it based on new technologies, best practices, and our evolving understanding of AI’s environmental impact.