World Engineering Day: how can we make AI as sustainable as possible?

The day was proclaimed by UNESCO at its 40th General Conference in November 2019, and it has been celebrated worldwide each year since 2020. According to UNESCO, “World Engineering Day is an opportunity to celebrate engineering and the contribution of the world’s engineers for a better, sustainable world.”

Engineering does indeed have a lot to celebrate. The amazing work of engineers combined with the leap in AI capability is delivering essential solutions to aid the fight against climate change. The United Nations, for example, is harnessing the technology to help communities facing climate risks in Burundi, Chad, and Sudan. The UN project uses AI “to help predict weather patterns, so communities and authorities can better plan how to adapt to climate change and mitigate its impact”. It’s an illustration of how engineers are using AI to directly reduce the impact of climate change on societies.

And this is just the start – expect a raft of further innovation in using AI to help solve climate challenges in the year ahead.

AI’s own sustainability issue

Although it is wonderful that AI innovation is being used to reduce climate change and its impact, this progress is counterproductive if, as an engineering community, we don’t address AI’s own environmental issues – we need to make sure that AI isn’t also contributing to climate change.

This is a tall ask. Data centres, which are used in both training and deploying AI, are already consuming a significant amount of energy: although somewhat of an outlier, in 2022 Ireland’s datacentres were consuming 18% of the country's electricity – the same amount as every urban household in Ireland combined. And this is on track to worsen. As reported in the New Scientist, datacentres globally “could double their electricity consumption in just two years” (equivalent to as much energy as the whole of Japan consumes today).

Moreover, when looking at the power consumption of AI models, it is estimated that ChatGPT consumes more than 100MWh per month (based on a report on Business Insider that ChatGPT costs OpenAI over $700,000 a day to operate and $200 per KWh per month).

If surging power consumption continues on this path, the environmental impact, especially on carbon, water, and energy, will be substantial. This will reduce the ability to roll-out and advance AI, something which is needed to help tackle the climate crisis. Therefore, the pressure is building on governments, companies, and universities to reduce the environmental impact of AI processing.

Current approaches to improving energy efficiency

The IT industry is a major adopter of renewable energy. But until there is greater availability of renewable energy, datacentres need to find other ways of creating a step-change in energy efficiency. As such, several solutions are in motion to build more sustainable data centres.

Evaporative cooling, for example, is starting to be phased out due to its drain on water resources, with some countries using their natural climate to enable ‘free cooling’ instead. Other data centres are introducing innovative ways of capturing and reusing the heat produced for other purposes, such as heating nearby buildings. Then, tools like asset performance management software are enabling engineers to optimise efficiency by adjusting heating and cooling levels in real time.

However, even with these methods of improving energy efficiency, the growth and complexity of AI requires a technical performance solution too.

The power of light

It is time for engineers to be open minded and to seek new, innovative technologies to solve the challenge of creating more sustainable AI. One of these innovations is using light instead of solely using electrons to compute – and, in the coming years, this is likely to become a lead technology in datacentres. Why so?

The majority of data centres still rely on traditional electronics. But the silicon-based GPU AI accelerators used consume vast amounts of energy and therefore give off vast amounts of heat. This means they are limited in their ability to produce the amount of compute power needed to meet rapidly increasing AI data centre performance demand.

A 3D optical AI processor, on the other hand, computes with photons instead of electrons. This means it can significantly surpass the limitations of existing transistor-based digital electronics. Most importantly, the processor is capable of highly energy-efficient and ultra-fast parallel processing – many times faster than traditional computing – that can help unlock the potential of next generation AI. And not only does it exhibit impressive power efficiency, but it subsequently significantly reduces the operating costs of running AI processors.

In this light, the power of light can create a sustainable solution for generations of AI to come. 

Dealing with the challenge, and what’s ahead for AI

Discussing sustainability and AI, the UN says: “In all countries, there is also a great deal to be done – to deal with the impacts of climate change, environmental issues, our growing cities and the challenges of emerging technologies including artificial intelligence.”

There is a lot to deal with, and deals are being made: the Climate Neutral Data Centre Pact has over 100 data centre signatories and, as part of its ‘climate pledge’, has set a goal for all datacentres to solely use renewable energy by 2030. But if this pledge is to become meaningful action, engineers need to embrace new approaches and technologies they may not have even considered yet.

So, what are our predictions for the future of AI?

What is very clear is that we are very far from reaching the peak of AI processing demand. Therefore, to meet this demand there will be a rapid shift in the way that AI processing is performed, with engineers and the industry embracing new technologies to meet the needs of AI. We also predict a move to reduce the amount of processing and memory required for AI inference through quantisation – 8-bit number formats are already extensively used, so expect this to drop to 4-bit or even 1-bit in the future.

At the heart of these shifts is a move away from traditional electronics to 3D optical AI computing. This has the potential to transform the energy efficiency of datacentres and meet the compute demand for advanced AI, which itself can then help tackle the climate crisis head-on.

The AI industry needs a major paradigm shift – and what better time to drive it forward then on World Engineering Day for Sustainable Development.