Humanity is standing on the brink of a technological revolution with the potential to reshape our world: the advent of artificial general intelligence (AGI), which could endow machines with the ability to think and emote like humans. Yet, as we contemplate this profound leap in technology, we must ask a crucial question: Can AGI help address one of the most pressing issues of our era, climate change?
Today’s AI-driven models can indeed play a pivotal role in mitigating climate change by monitoring greenhouse gas emissions and predicting extreme weather events. For example, a collaborative effort between IBM and NASA has produced a sophisticated system for this purpose. However, as we harness the power of AI and machine learning, we must confront an environmental paradox: the ecological toll exacted by energy-intensive models. The carbon footprint of data centers that house these AI behemoths depends on factors like electricity consumption, water usage, and the frequency of equipment replacement.
According to a report by Climatiq, cloud computing is responsible for a significant share of global greenhouse gas emissions, contributing 2.5% to 3.7% of the total, surpassing even the emissions from commercial flights at 2.4%. It is important to note that these statistics are based on data that is now a few years old. Since then, the energy demands of artificial intelligence have only intensified, raising concerns about their sustainability.
AI models have a voracious appetite for energy, with data centers, like the Lincoln Laboratory Supercomputing Center (LLSC) at the Massachusetts Institute of Technology, witnessing a surge in the number of AI programs running on their servers. This spike in energy consumption prompted computer scientists to explore energy-efficient solutions.
Vijay Gadepally, a senior staff member at LLSC, emphasized the importance of energy-aware computing in addressing this issue. The team at LLSC adopted an innovative approach, capping the power intake of the energy-hungry graphics processing units (GPUs) that fuel AI models. For instance, OpenAI’s GPT-3, powered by NVIDIA GPUs, consumed an astounding 1,300 megawatt-hours of electricity, equivalent to the monthly consumption of a typical U.S. household. It’s estimated that OpenAI used around 10,000 GPUs to train GPT-3.
Capping the power intake allowed the researchers to reduce the energy consumption of AI models by 12-15%. However, this came at the cost of extended training times. In one experiment, the team limited a Google BERT language model’s GPU power to 150 watts, resulting in a two-hour increase in the training time (from 80 to 82 hours).
In addition to capping power, the team at LLSC developed software that enables data center operators to set energy limits for their entire system or on a per-job basis. As a result, the GPUs in their supercomputers ran 30 degrees Fahrenheit cooler, reducing strain on the cooling system and extending the reliability and lifespan of hardware equipment.
As we look ahead, the International Energy Agency (IEA) emphasized the importance of energy efficiency and zero-carbon electricity in limiting emissions growth. They underlined the necessity of strong climate policies to ensure that digital technologies, including AI, contribute to emissions reduction rather than exacerbation.
Furthermore, the researchers at MIT devised a novel approach to curbing energy consumption during AI model training. Training AI models requires vast amounts of data and the testing of thousands of configurations to determine the optimal parameters. This process is energy-intensive. To address this, the team developed a model that predicts the likely performance of different configurations, allowing early termination of underperforming models. This innovative approach led to an impressive 80% reduction in energy consumption during model training.
In conclusion, the promise of AGI and AI in addressing climate change is undeniable. However, the environmental impact of energy-intensive AI models cannot be ignored. Innovations like capping power intake, energy-efficient software, and early stopping of underperforming models offer solutions to this conundrum, ensuring that our journey towards artificial general intelligence is as sustainable as it is groundbreaking.
