As the University of Birmingham explores the transformative potential of Artificial Intelligence (AI) to enhance our academic and administrative functions, it is crucial to prioritise the sustainability of these technologies to mitigate their environmental impact. AI promises to enhance efficiency, improve decision-making, and foster innovative solutions in both education and research. It has the potential to personalise learning experiences, streamline administrative processes, and bolster advanced research through data analysis. However, these advantages are accompanied by a considerable environmental footprint.
Understanding AI’s Environmental Impact
The environmental impact of AI is complex. It encompasses the substantial resources needed for the development, training, deployment, and maintenance of AI models, as well as the infrastructure required to support these processes. Data centres, the backbone of AI servers, consume vast amounts of electricity and water, generate electronic waste, and rely on critical minerals and rare elements often sourced unsustainably. The resource-intensive process of training large AI models requires thousands of megawatt-hours of electricity, leading to substantial carbon dioxide emissions. As the adoption of AI technologies increases, addressing these environmental challenges is vital for a sustainable future.
Note: The AI image above was created by the DALL-E image AI model, each image that is created by this AI model is estimated to consume 2.2 g CO2e per creation. Likely significantly less than if it were created by a person.
Microsoft as a Sustainable Partner
In selecting Microsoft Azure as our cloud provider, the University of Birmingham aligns with cutting-edge technology that supports a greener future. Microsoft has demonstrated a strong commitment to sustainability, aiming to be carbon negative by 2030 and to remove all the carbon it has emitted since its founding by 2050[i]. These ambitions extend beyond carbon offsetting to actively removing more carbon from the atmosphere than the company emits. Azure’s data centres employ advanced cooling techniques, energy-efficient hardware, and intelligent energy management systems to optimise energy use[ii]. Additionally, Microsoft collaborates with partners to deliver sustainable solutions and invests in technologies that reduce environmental impact, such as AI for environmental monitoring[iii].
However, it is important to critically assess these claims. The Green Web Foundation[iv], notes that despite Microsoft’s sustainability efforts, the company’s emissions have increased by approximately 45% over the last three years. Additionally, Microsoft has not met its own supply chain decarbonisation targets.
Toward Sustainable AI Practices
As we begin to explore integrating AI into our academic and administrative operations, it is vital to consider strategies that could minimise its environmental footprint. Here are some key approaches we might consider:
- Responsible AI Use: Evaluate the necessity of AI for certain tasks, recognising that less resource-intensive methods may often be preferable. When employing AI, choosing the right models, optimising code, and ensuring efficient resource use are critical to responsible application. Just because we can doesn’t mean we should.
- Leveraging Efficient AI Models: Utilising pre-trained models requires much less computational power, as they eliminate the need for extensive training. When customisation is needed, fine-tuning these models to our specific tasks helps maintain efficiency. Additionally, ‘mini models’, like GPT-4o-mini, stand out for their ability to deliver performance with much lower energy requirements[v].
- Optimising Infrastructure: Using Azure’s features like auto-scaling and serverless computing ensures AI workloads run efficiently, using resources only when necessary and reducing idle times. Such practices not only contribute to our sustainability efforts but also improve operational efficiency[vi].
- Azure OpenAI’s Batch Processing: Batch processing minimises the need for continuous, real-time processing, which can be more resource-intensive. This allows for more efficient use of computational resources, leading to energy savings and a reduction in operational costs by up to 50% compared to standard processing. It is particularly useful for tasks like large-scale data analysis, content generation, and document review, where handling multiple requests simultaneously can significantly improve efficiency[vii].
- Monitoring Environmental Impact: Services like Azure Carbon Optimisation are invaluable for tracking the carbon footprint of our AI projects, enabling continuous refinement towards eco-friendlier outcomes[viii].
The Impact of Thoughtful AI Application
Implementing sustainable AI practices can have a major positive impact on reducing energy consumption and optimising resource use:
- Energy Efficiency: Adopting smaller, more efficient AI models can lead to significant reductions in energy consumption. Techniques like model distillation, capable of shrinking AI models by up to 90%, demonstrate that optimisation can lead to substantial energy savings without sacrificing performance. These smaller models require significantly less computational power and storage, and also result in lower running costs and faster processing times[ix].
- Resource Optimisation: Utilising features such as auto-scaling and serverless architectures can lead to considerable energy and cost savings by dynamically adjusting resource consumption to match demand. This ensures that resources are employed only when necessary, minimising idle times and preventing energy wastage. Integrating batch processing further enhances these strategies by optimizing the timing and scale of computational tasks, thereby improving our capability to use energy and resources more judiciously.
- Sustainability Insights: Monitoring the carbon footprint of AI workloads in Azure is crucial for identifying inefficiencies and implementing improvements, leading to significant reductions in emissions. This proactive approach in understanding and mitigating the environmental impact of our AI applications is essential for cultivating a culture of sustainability at the university.
Conclusion
By adopting these strategic approaches, the University of Birmingham could be well-positioned to make significant advances in the realm of sustainable AI. Our commitment to environmental stewardship is reflected in our selection of partners like Microsoft Azure, and our efforts to optimise AI models and infrastructure for greater efficiency, and our ongoing commitment to the meticulous monitoring and enhancement of our sustainability practices. As we move forward, we should remain focused on discovering and implementing innovative methods to further bolster the sustainability of our AI initiatives. This ensures our dual objectives of contributing positively to the environment and advancing our academic and administrative goals are met with equal dedication.
[i] Microsoft will be carbon negative by 2030
[iii] Microsoft corporate responsibility – sustainability
[iv] The Green Web Foundation AI environmental impact report
[v] GPT-4O Mini vs GPT-4O: A Comprehensive Comparison of AI-Language Models
[vi] How green is your cloud? Guide to building a more sustainable stack
[vii] Getting started with Azure OpenAI batch deployments
[viii] What is Azure carbon optimization
[ix] Bigger Not Always Better as OpenAI Launch New GPT-4o mini