Sustainable approaches and innovation: how to reduce the ecological footprint of AI?

Given its significant environmental impact, the AI and data center industry is actively seeking innovative strategies to reduce its environmental footprint - with a focus on energy efficiency, water conservation and resource optimization. These solutions address the challenges described in the previous section.

Streamlining data centres

Rethinking traditional approaches to cooling

One example is rethinking traditional approaches to cooling. Google has conducted a study that suggests that excessive cooling of components, particularly hard disk drives (HDDs), can be counterproductive. Too low temperatures can lead to mechanical and electrical problems, which paradoxically increases the failure rate.

These findings support the trend increasing operating temperatures in data centres. Google, for example, operates some of its data centres at temperatures up to 27 °C, which contributes to higher energy efficiency without negatively impacting equipment reliability.

Free cooling - cooling using natural conditions

Taking advantage of natural climatic conditions offers significant potential for reducing energy consumption:

• Free cooling by air: It uses cool outdoor air to cool data centres, reducing the need for conventional air conditioning - especially in colder areas.
• Free cooling water: It uses naturally cold water sources, such as lakes or seawater, to cool servers. For example, google data centre in Hamina, Finland uses seawater for cooling. Studies show that using these techniques can reduce cooling costs by up to 67 % compared to traditional methods.

Waste heat recovery

In addition to efficient cooling, the focus is also on the smart use of waste heat from data centres:

• Central heating: Capture and use waste heat for heating residential and commercial buildings in the area. For example, Meta's data centre in Odense, Denmark, will annually provide 100 000 megawatt hours heat to the local community.
• Agricultural use: Excess heat can be used for heating of greenhouses and support year-round crop production.
• Industrial processes: Waste heat can be used for industrial purposes, for example for drying wood pellets.

Responsible water management and the transition to renewables

Wider corporate commitments and investments also play an important role:

Target: water positivity: Technology companies such as Microsoft and Google have committed to making 2030 will be "water positive" - that is, they will return more water than they use.

• For example, Microsoft is investing in projects replenishing water in vulnerable regions, while Google focuses on improving water management across its services.

Integration of renewable energy sources: Data centres are increasingly using renewable energy sources - solar and wind - to reduce the carbon footprint associated with electricity generation.

• For example, Microsoft has entered into contracts to supply more than 900 MW renewable energy for its data centres in Ireland, while Google has signed an off-take agreement 100 MW energy from the Moray West wind farm in Scotland for its UK operations.
• Companies should try to influence the energy mix in the very regions where data centres operate. The aim is to increase the overall sustainability of the infrastructure and reduce dependence on fossil fuels (by using nuclear, hydro, solar and wind).

Optimising AI models

In addition to hardware and infrastructure, the optimization of AI models itself plays an important role, which can significantly reduce resource consumption:

Quantization

Reducing the accuracy of model computations (e.g. from 32-bit to 8-bit), which reduces computational power and energy consumption and carbon footprint - without significantly affecting the quality of the results.

Ztráta přesnosti vah po dekvantizaci

Model distillation

Training smaller, effective models (students) to imitate the behavior of larger models (teachers). The result is models with high accuracy and significantly lower resource requirements.

Destilace velkého modelu do menšího

MoE architecture (Mixture-of-Experts)

It uses only selected parts of the model according to the specific task, which reduces the number of calculations and energy consumption.

MoE architektura

Prompt caching

It allows you to store and reuse frequently repeated parts of a prompt, thus significantly reducing latency and computation costs.

For example, OpenAI implemented prompt caching in its APIs, which led to 50% reduce costs and process prompts faster.

Prompt caching

Pruning

Removing less important neurons or connections in the model, resulting in smaller model size and lower computational power requirements.

Studies show that pruning can reduce size, and thus energy, by up to 90 % with minimal loss of model performance (when pruning is done carefully).

Prořezávání modelu (neuronů i synapsí)

Speculative Decoding

This method speeds up text generation by having a smaller and faster "draft" model suggest several tokens up front, which are then verified and possibly modified by the larger "verification" model.

Thanks to parallel token processing, inference is significantly accelerated without the need to retrain the model.

Přijímání a odmítání navrhnutých tokenů vyšším modelem

vLLM

An open-source library that optimizes inference of large language models using the PagedAttention algorithm, which efficiently manages memory by partitioning keys and values into smaller blocks.

In this way, it reaches up to of a 24-fold increased throughput over traditional libraries without having to change the model architecture.

Optimalizace inference LLMs

Individual action and social influence

While systemic change and technological innovation are key to reducing the environmental impact of AI, we as individuals and as a society also have an important role to play. Each of us can contribute to a more sustainable digital ecosystem through our approach:

• Responsible use of AI: Before using an AI tool, let's consider whether a simpler solution is not enough. Let's cluster queries, optimize prompts and limit demanding tasks like unnecessary image and video generation.
• Use efficient and local tools: For less demanding tasks, we can use AI tools running directly on our device, eliminating the need for remote data centres. Choose applications known for their energy efficiency.
• Promote sustainable approaches: Let's call on companies to be transparent about their energy and water consumption, support regulations requiring the use of renewables, and engage in public debate about the environmental impact of AI.
• Influence with your decisions: Let's support politicians and political groups that are committed to sustainability and responsible technology development. Let us use platforms and tools that openly advocate sustainable operations.

Conclusion

These comprehensive approaches, both at an industry-wide level and in our individual actions, show that the industry is taking the environmental impacts of AI seriously and moving towards more responsible and sustainable technologies and approaches. Sustainability in AI is not just a technical task, it is a shared responsibility - and every thoughtful step counts.

What other innovative approaches would you see as key to reducing the ecological footprint of AI in the future?