The enormous amounts of power AI needs may become the biggest obstacle to its growth

By Laio

The rapid growth of artificial intelligence (AI) has brought with it a spike in carbon emissions, as the creation of chatbots and image generators demands vast amounts of electricity. Asking Chat-GPT ten million questions equals the energy to power 5,000 homes daily, Mark Papermaster (cto at AMD) said at the ITF World conference.

Major players like Microsoft, Google, and OpenAI use cloud computing in massive data centres globally to train AI algorithms, known as models. The success of ChatGPT has spurred a race among companies to develop rival AI systems, chatbots, and products that use large AI models for various users. As AI continues to revolutionise industries such as retail and medicine, its environmental impact is becoming a growing concern, with greater transparency on emissions likely to lead to increased scrutiny.

ITF World puts imec at the center of the global semiconductor industry

To train over a thousand trillion of its parameters, GPT-4 needed almost ten thousand MW/h. To stay up to date and exploit AI’s capabilities, producing more – and more powerful – chips is key. 

Measuring the Carbon Footprint of AI Technologies

Current industry standards and best practices for measuring and reporting the carbon footprint of AI and chatbot technologies are still in their infancy. However, efforts are being made to establish a universal standard. AI startup Hugging Face, for example, has developed a method to estimate the broader carbon footprint of large language models (LLMs) throughout their entire life cycle, instead of focusing solely on the training phase. Hugging Face’s approach takes into account energy consumption during training, manufacturing and maintaining supercomputer hardware, computing infrastructure, and post-deployment energy usage.

Although Hugging Face‘s work has not yet been peer-reviewed, it is considered the most thorough, honest, and knowledgeable analysis of the carbon footprint of a large ML model to date, according to Emma Strubell, an assistant professor at Carnegie Mellon University. The company’s paper provides much-needed clarity on the carbon footprint of LLMs, but more research is needed to better understand the real-world environmental impact of AI systems, including knock-on emissions from AI-driven recommendation and advertising algorithms.

Reducing AI’s Environmental Impact

As awareness of AI’s carbon footprint grows, companies are exploring ways to reduce their technologies’ environmental impact. Hugging Face, for instance, trained its LLM, BLOOM, on a French nuclear-powered supercomputer, resulting in lower CO2 emissions compared to models trained in regions relying on fossil fuels. BLOOM’s training phase produced 25 metric tons of CO2 emissions, which doubled when accounting for manufacturing, infrastructure, and operations. In comparison, OpenAI’s GPT-3 and Meta’s OPT were estimated to emit over 500 and 75 metric tons of CO2 during training, respectively.

Experts believe that the findings from Hugging Face’s paper could encourage a shift towards more energy-efficient AI research, such as fine-tuning existing models instead of creating larger ones. The increasing awareness of AI’s carbon footprint might also prompt large tech companies to focus on energy efficiency and explore alternative energy sources for their data centres. For instance, Microsoft has committed to becoming carbon negative by 2030, while Google has set a goal to run its data centres and offices on carbon-free energy 24/7.

Revolutionizing the chip industry: The power of chiplets

Chiplets are transforming the computer chip industry by offering a radical new approach to microprocessor design.

The Consequences of Unchecked AI Emissions

As AI and chatbot technologies continue to grow in usage and popularity, failing to address their carbon footprint could have significant consequences. Atsuyoshi Koike, CEO of Japanese chip manufacturer Rapidus, predicted that in two years, AI data centres would require more than ten per cent of global power generation. If left unchecked, the increasing energy demand from AI technologies could hinder efforts to achieve climate change goals and exacerbate global emissions.

Moreover, unchecked AI emissions could also lead to regulatory measures and increased scrutiny of the tech industry, impacting companies’ reputations and bottom lines. A proactive approach to managing the carbon footprint of AI technologies is not only environmentally responsible but also strategically beneficial for companies looking to stay ahead in an increasingly eco-conscious world.