Last updated: 1 March 2024 at 09:53

Introduction

By taking a more sustainable approach, we can ensure that the production and consumption of entertainment is not only enjoyable, but also responsible and sustainable. Behind every digital device, there is a complex supply chain that use energy- and water-intensive data centres to store and process information, as well as the extraction of raw materials, manufacturing, transportation, and disposal. This complex ecosystem of devices, data centres, algorithms, and networks is responsible for a significant amount of greenhouse gas emissions.

Visualisation of the 'world behind the screen' for a gaming value chain.

It's clear that there is a need for Createchs to take a more sustainable approach to physical and digital production and distribution. This means not only reducing the carbon footprint of devices and data centres, but also rethinking the design of products and services to minimise environmental impact. For instance, energy-efficient devices and servers, renewable energy sources, and responsible e-waste management can all help to reduce the carbon footprint of the Createch ecosystem.

Understanding data centre energy efficiency

Energy efficiency in data centres is a critical aspect of reducing the environmental impact of digital technologies, as data centres consume significant amounts of energy to power and cool the servers that store and process vast amounts of data. Improving energy efficiency in data centres not only helps reduce greenhouse gas emissions but also lowers operational costs. Here are some key strategies for enhancing energy efficiency in data centres:

1. Cooling systems: Efficient cooling systems are essential to manage the heat generated by servers. Some effective cooling techniques include using outside air for cooling, liquid cooling, hot and cold aisle containment, and optimising the data centre layout to facilitate better airflow.
2. Server virtualisation: Virtualisation technology enables multiple virtual servers to run on a single physical server, maximising the use of available resources and reducing the number of servers needed. This results in lower energy consumption and less heat generation.
3. Server consolidation: Data centre operators can reduce energy consumption by consolidating underutilised servers, either by decommissioning them or combining their workloads onto more efficient hardware.
4. Energy-efficient hardware: Investing in energy-efficient servers, power supplies, and storage systems can significantly reduce energy consumption. ENERGY STAR-certified equipment and high-efficiency power supplies are examples of energy-efficient hardware options.
5. Power usage effectiveness (PUE): PUE is a widely used metric that compares the total energy consumed by a data centre to the energy consumed by its IT equipment. Monitoring and improving PUE can help data centre operators identify inefficiencies and implement energy-saving measures.
6. Demand management: Implementing demand management strategies, such as scheduling non-critical tasks during periods of lower energy demand, can help balance the load on the data centre and reduce energy consumption.
7. Renewable energy: Data centres can further reduce their environmental impact by sourcing energy from renewable sources, such as solar, wind, or hydroelectric power.

Data centre actions for Createch companies

Createch businesses range from those that have their own data centres to those that work off a laptop and third-party hosting services only. Most use a mix of both. In all cases there are several steps to ensure that data centres are operated sustainably and have minimal environmental impact:

1. Set sustainability goals: Establish and communicate clear sustainability goals related to computing, server and data centre operations, such as reducing energy consumption, increasing the use of renewable energy, and minimising waste.E-waste management: Ensure the responsible disposal and recycling of electronic waste, including decommissioned servers and other hardware, by partnering with certified e-waste recyclers.
2. Select sustainable data centre providers: Choose data centre providers that prioritise energy efficiency, use renewable energy sources, and have a strong commitment to sustainability.
3. Energy-efficient hardware: If you are building your own data centre (even if it is a single server), then use energy-efficient processors, servers, storage systems, and networking equipment to reduce energy consumption and heat generation in the data centre. When working with conventional and cloud data centre providers, make service choices based on the underlying hardware to minimise environmental impact while optimising the performance and cost of your services and products.
4. Virtualisation and consolidation: Implement server virtualisation and consolidation strategies to maximise resource utilisation and reduce the number of physical servers required, resulting in lower energy consumption.
5. Optimise software performance: Design and optimise software to consume fewer resources, which can help reduce energy consumption and extend the lifespan of hardware. Use energy tracking software modules in your code to measure or estimate likely impact in terms of emissions, and to model alternative coding strategies.
6. Monitor and measure hardware energy efficiency: In the case of owned servers and data centres, track energy consumption using smart plugs, energy bills or energy tracking software. Third parties usually provide information concerning their energy consumption on your behalf that allows you to regularly track energy efficiency metrics, such as proportion of carbon-free energy from renewables and data centre Power Usage Effectiveness (PUE). Use this information to identify evaluate equipment and supplier performance, and to implement improvements.
7. Demand management: Balance computing loads by scheduling non-critical tasks during periods of lower energy demand and avoiding energy-intensive operations during peak hours.
8. Collaboration and knowledge sharing: Participate in industry initiatives and collaborate with other Createch businesses to share best practices and drive innovation in sustainable data centre management.

By implementing these strategies, Createch businesses can promote sustainable data centre practices, reduce their environmental impact, and contribute to a more responsible digital ecosystem.

Energy consumption in machine learning

Energy consumption in machine learning, particularly for Large Language Models (LLMs) and generative AI models (such as OpenAI' ChatGPT and DALL-E), can be substantial due to the computational resources required for both training and inference. This because their large neural networks demand significant processing power to generate their text, image, audio and video outputs.

For example, during 2023, ChatGPT handled around 200 million requests per day and each required around 3Wh to process (based on an article by Alex de Vries, 'The growing energy footprint of artificial intelligence', published in Joule Vol. 7 Issue 10 in 2023). This equates to over 500MWh per day (the same as a small european city) and equal to around 200tonnes of CO2e emissions (based on a grid emission factor of 0.4 kg CO2e per kWh).

Several factors contribute to the high energy consumption in machine learning, and Createch companies can make choices to minimise the impact:

1. Model complexity: Generative models often have millions or billions of parameters, which require substantial computational resources for processing. The more complex the model, the higher the energy consumption during training and inference.
2. Training duration: Training large models can take days or even weeks, depending on the model size and the dataset. This prolonged processing time leads to increased energy consumption.
3. Frequent updates and fine-tuning: AI models often require frequent updates and fine-tuning to maintain accuracy and performance. Each update can consume significant energy resources, especially if the model is large.
4. Hardware requirements: Training and running AI models usually require powerful hardware, such as graphics processing units (GPUs) or tensor processing units (TPUs). These specialised hardware components can consume significant amounts of energy during operation.

Training AI models is typically performed in third party data centres or cloud providers so it is important to ensure that your machine learning workloads are hosted in environmentally responsible facilities. This means selecting providers that prioritise energy efficiency, use renewable energy sources, and have a strong commitment to sustainability.

Whether you operate your own hardware or use third-party data centres and clouds or a mix, strategies to reduce energy consumption Createch companies are similar:

1. Green data centres: Ensure that your machine learning workloads are hosted in environmentally responsible facilities. If using third parties, partner with data centre providers that prioritise energy efficiency, use renewable energy sources, and have a strong commitment to sustainability (e.g. through certification to ISO 50001).
2. Dynamic resource allocation: Implement dynamic resource allocation strategies that adjust computing resources based on the workload demand, reducing energy consumption during periods of low usage.
3. Model efficiency: Develop more efficient models by using techniques such as weight pruning, quantisation, or knowledge distillation, which can reduce the number of parameters and computations needed without sacrificing performance.
4. Hardware optimisation: Optimise hardware usage by selecting energy-efficient processors, using hardware accelerators designed for AI workloads, or leveraging cloud-based services that optimise resource utilisation (e.g., scheduling workloads at times when carbon free energy in the mix is highest).
5. Transfer learning: Use pre-trained models and fine-tune them on specific tasks rather than training models from scratch. This approach can save both time and energy by leveraging existing model weights and architectures.
6. Energy-aware algorithms: Research and develop energy-aware machine learning algorithms that balance performance with energy consumption, prioritising energy efficiency without compromising model quality.
7. Monitoring and reporting: Regularly monitor energy consumption and other sustainability metrics associated with machine learning workloads. Share progress with stakeholders and set targets for continuous improvement.
8. Collaboration and knowledge sharing: Engage with the broader AI and sustainability communities to learn about best practices, share insights, and collaborate on initiatives that advance sustainable machine learning.
9. Sustainable AI applications: AI applications are prone to well-publicised issues such as discrimination between different groups of people, a lack of transparency concerning algorithms and data, copyright infringement and excessive energy consumption. These issues need to be addressed during design and development. It is important to establish design principles for developers to facilitate development and deployment AI applications with positive environmental and social impacts. , and build in fairness, accessibility, data privacy, clear data ownership and lower emissions. An example of these principles are those promoted by Microsoft or the OECD.

For developers there are a range of strategies that can be employed to minimise energy consumption:

• Include energy consumption factors in model architecture decisions
• Decide whether you need to offer a service 24hrs x 7days per week x 365 days per year
• Train models on hardware and in regions where the energy mix is cleaner

• Use tools such as energydashboard to identify times of the day when the energy mix is cleaner (“lower carbon intensity”)

• Take advantage of times of the day when the  energy mix is cleaner
• Train models on hardware optimised for ML tasks
• Debug on small scale training sets
• Store data in low-energy storage
• Use pre-training of models where possible
• Once model training plateaus, stop it!
• Use simple, small models when they are sufficient (e.g., don’t used the latest, biggest model just because you read about it in an academic paper)
• Similarly, distil large models into smaller production models
• Optimise models for accuracy AND energy consumption
• Applications should be analysed and tuned on the target hardware
• Understand that for most ML computations, there is an optimal number of cores that will minimise carbon footprint (i.e. using more cores may not speed execution, but will increase carbon footprint)
• Pay attention to the activation function: selection of an activation function can greatly influence the time a model takes to train
• Set a budget for the model i.e. limit how much to allow a machine learning model to emit during its lifetime

Focusing on energy efficiency in machine learning, means that developers can contribute to more sustainable AI development practices and help reduce the environmental impact of machine learning technologies.

This information is brought to you by the Centre for Sustainable Design (CfSD) at the University for the Creative Arts in the UK. CfSD was established in 1995 in Farnham, Surrey, UK and is based within the Business School for the Creative Industries (BSCI). The Centre has led and participated in a range of high-quality research projects and has organised hundreds of conferences, workshops and training courses in Europe. CfSD works with partners in Europe, Asia, and North America to deliver high quality results.

Disclaimer.

Copyright© 2023 The Centre for Sustainable Design