They estimated teaching the neural super-network in a Microsoft data center using Nvidia GPUs required roughly 190,000 kWh, which using the average carbon intensity of America would have produced 85,000 kg of CO2 equivalents, the same amount produced by a new car in Europe driving 700,000 km, or 435,000 miles, which is about twice the distance between Earth and the Moon, some 480,000 miles. The Register
Introduction
Generative AI, a subfield of Artificial Intelligence where sophisticated neural networks create images with a simple text description or text responses to your questions. It can generate something new from initial input and has been revolutionizing various sectors, from creating pieces of “art” to making significant breakthroughs in the medical field. Its potential seems boundless, promising advancements that once seemed like science fiction.
On another front, we are facing a global climate change crisis defined by long-term shifts in temperatures and weather patterns. This crisis is largely caused by human activities, specifically, the emission of greenhouse gases, including those emitted by using AI tools. The evidence is irrefutable - rising sea levels, more frequent and severe weather events, the decline in biodiversity, etc. Our planet is in peril and we need immediate solutions and action because people want change.
Interestingly, Generative AI and climate change have a paradoxical relationship. The computing requirements, and thus energy, required to train and use these models are staggering. AI could also be our way out of this mess. AI applications in energy, transportation, and agriculture could reduce global greenhouse gas emissions by up to 4% by 2030. AI is now a potential tool in the toolbox of climate change mitigation efforts, aiding in everything from predicting extreme weather events to optimizing renewable energy production. It can help us harness data and computational power for a more sustainable world.
The Role and Impact of AI in Climate Change
Energy Consumption and Computational Resources:
Generative AI, characterized by its ability to produce original looking content and sound knowledgeable, has become a driving force behind numerous applications across industries. Whether it's crafting art, generating human-like text, or simulating realistic environments, these AI models demand substantial computational power during their training phases. This, in turn, translates into high energy consumption, contributing to the carbon footprint of AI technologies.
The training process for generative models requires immense computational resources that draw electricity from power grids. In regions where fossil fuels still dominate the energy mix, the environmental cost becomes pronounced. The emissions generated from this energy consumption, primarily in the form of carbon dioxide and other greenhouse gases, exacerbate our current climate challenges.
Data Centers and Environmental Degradation:
Supporting the image and text generation are data centers housing racks upon racks of servers. These data centers are necessary for processing the huge amount of data required for training AI models; however, they also contribute to environmental degradation through the constant operation of these facilities generates substantial heat, necessitating energy-intensive cooling systems. This energy-hungry cooling of the energy-hungry servers themselves, places an additional burden on the grid and the environment.
To add insult to injury, the construction and maintenance of data centers can have extensive ecological effects, often requiring significant land and water use, leading to habitat disruption and deforestation. The infrastructure needed to support these data centers, such as roads and power lines, can impact ecosystems and disrupt local environments. Essentially, the environmental impact of using ChatGPT to write clam poems is not zero.
E-Waste and Environmental Pollution: Beyond the training and operation of generative AI models, the entire lifecycle of AI hardware presents environmental concerns. From the production of computer components to their eventual disposal, Generative AI not only generates images and text, but also generates electronic waste. Improper disposal of e-waste can lead to the release of hazardous materials into the environment, including heavy metals and toxic chemicals, which can contaminate soil, water sources, and air. We need to address the e-waste problem as the AI industry continues to expand, with the potential to worsen environmental pollution. With this wide array of climate problems created by generative AI, it is evident that the technology's transformative potential comes with a substantial environmental cost if we’re not smart about it.
Case Study
The Carbon Footprint of Meta’s Llama 2 Models:
Meta, a leader in AI, shares the energy requirements for training their latest set of open-source large language models (LLM). Once again, the open-source community paves the way in transparency on carbon emissions (versus back-of-envelope calculations for closed-source models).
The 539 tCO2eq is roughly equivalent to about the yearly emissions of 120 cars in the US. That’s not a lot, but Meta is one of many players in the now-crowded Generative AI space. And these emissions only take into account training the models, not accounting for actually using the trained models. Usage costs will vary greatly depending on the model, hardware, and local energy costs wherever generation is taking place, but each usage incurs the cost of electricity.
Aiding Climate Change Research
One of AI’s superpowers is processing large amounts of data to identify complex, underlying patterns – weather prediction is one area where research is beginning to bear fruit. DeepMind’s GraphCast is providing accurate predictions at unprecedented speed through training on four decades of historical weather data observations.
“GraphCast predicts weather conditions up to 10 days in advance more accurately and much faster than the industry gold-standard weather simulation system – the High Resolution Forecast (HRES), produced by the European Centre for Medium-Range Weather Forecasts (ECMWF).
GraphCast can also offer earlier warnings of extreme weather events. It can predict the tracks of cyclones with great accuracy further into the future, identifies atmospheric rivers associated with flood risk, and predicts the onset of extreme temperatures. This ability has the potential to save lives through greater preparedness.”
We can make more precise predictions about future weather patterns by using AI to analyze historical climate data, helping communities prepare for extreme events like hurricanes and droughts. AI is also helping us track climate change. We can track deforestation, measure changes in ice caps, and monitor the health of oceans and ecosystems by analyzing satellite images with advanced machine learning algorithms. AI can also play an important role in optimizing energy consumption. Smart grids powered by AI help with the efficient distribution of electricity, reducing waste and integrating renewable energy sources. And with Google’s immense map and traffic data, we are finding ways to optimize traffic flows for reduced emissions.
Possible Solutions and Future Prospects
As we confront the climate impact of generative AI, we need to address this complex issue with a multi-pronged strategy that combines technological solutions, policy interventions, and industry-wide cooperation. A few areas to focus:
Energy-Efficient Algorithms and Hardware Development: Even a small increase in efficiency can greatly reduce the energy required to use Generative AI. Even with current AI architectures, we can achieve greater efficiency through smaller experts (small models trained to do something well, improving efficiency), pruning (removing less important info from models to make them more efficient), and quantization (how compressed the model is, smaller might not be as accurate but uses less energy). And there are advanced architectures like MAMBA that are more efficient than current models.
Hardware Development: Hardware advancements can make running these models require less energy. Introducing AI-accelerators whose sole job is doing the specific math that AIs require can greatly improve the speed and efficiency of using Generative AI. The US Department of Energy found that NVIDIA’s A100 processors are 5x more energy efficient than traditional servers. These energy savings directly translate to reducing Generative AI’s carbon footprint.
Renewable Energy Integration: Data centers should be using renewable energy sources, including solar, wind, and hydroelectric - tech giants like Microsoft, T-Mobile, and Amazon have joined climate pledges to reduce carbon emissions to zero. This transition will help lessen the environmental impact of generating infinite images or chatting with characters and hopefully encourage others to follow. Running on renewable energy means working on potential solutions doesn’t come at the expense on the very problem it’s trying to solve.
Green AI Initiatives: Use AI to help solve the problem by using AI to develop better solar cells, storage, and energy management systems. New “thinking” on how to approach known problems and spur innovation while also anticipating future problems can help mitigate potential effects down the road.
Policy and Regulation: Governments and international organizations need to encourage sustainable AI practices. We need to establish standards for energy efficiency and incentivize AI companies to reduce emissions and sustainably recycle or properly dispose of hazardous hardware materials. Encourage innovation via grants to apply AI to help solve climate change.
Conclusion
Generative AI is more than a toy and is here to stay, if not for more and more use cases – but we can do it more efficiently with less negative impact on the environment. Combating climate change requires a comprehensive plan, effective rewards and punishments, and innovation. Turning AI on itself to develop more energy-efficient AI architectures, improve use of renewable energy, and optimize the power grid can hopefully accelerate our efforts to slow/reverse climate change and not just contribute to the problem. The future requires a delicate balancing act of technological innovation and energy efficiency brought by tech companies, governments, and users of the technology. Generative AI’s promise of infinite possibilities includes the possibility of turning things around on climate change.
