Students are actively working to enhance the environmental sustainability of AI technology. At the Lincoln Laboratory Supercomputing Center (LLSC) within the Massachusetts Institute of Technology (MIT), the increased utilization of AI programs has spurred computer scientists to redirect their efforts towards optimizing energy efficiency.
(Photo : FABRICE COFFRINI/AFP via Getty Images)
Teachers are seen behind a laptop during a workshop on ChatGpt bot organised for by the School Media Service (SEM) of the Public education of the Swiss canton of Geneva, on February 1, 2023.
Making AI Greener
Vijay Gadepally, a senior staff member at LLSC leading energy-aware research efforts, notes that energy-aware computing is a relatively unexplored area, but it's gaining momentum as the need for sustainable AI practices becomes increasingly evident.
Interesting Engineering reported that the research team imposed restrictions on the power consumption of graphics processing units (GPUs), the essential components fueling energy-intensive AI models.
By imposing these power limits, the researchers managed to curtail the energy consumption of AI models by an impressive 12-15%. However, there was a trade-off involved: the models now required more time to complete their training.
Reducing Strain
The team also developed specialized software enabling data center operators to set power limits either system-wide or on a per-job basis. The impact of this innovation has been noticeable at LLSC, where the GPUs within their supercomputers are now operating at temperatures 30 degrees Fahrenheit lower.
This not only reduces strain on the cooling system but also has the potential to enhance the reliability and longevity of the hardware. According to a report from the International Energy Agency (IEA), the coming decade is expected to witness a significant surge in demand for digital technologies and services.
Curbing emissions growth will hinge on advancements in energy efficiency, including research and development into next-generation technologies, widespread adoption of zero-carbon electricity sources, and the decarbonization of supply chains.
Also Read: Can AI Improve Civility and Quality of Online Discussions? New Study Puts It to the Test
The MIT research team also devised an alternative approach to significantly reduce energy consumption. Training AI models typically entails processing vast quantities of data and experimenting with thousands of parameter configurations to optimize performance, which consumes substantial energy resources.
To mitigate this, the team created a predictive model that assesses the expected performance of various configurations. Models that exhibit underperformance are identified and halted at an early stage. Remarkably, this early stopping technique resulted in an impressive 80% reduction in energy consumption during the model training process.
AI & Climate Change
We may be on the verge of a significant technological milestone: the development of artificial general intelligence (AGI), where machine learning models can exhibit conscious thinking and emotions akin to humans. However, the question remains: can this cutting-edge technology effectively tackle the most pressing concern of our era - climate change?
The carbon footprint of data centers relies on factors such as electricity and water usage, as well as the frequency of equipment replacement. As per findings from Climatiq, cloud computing has been identified as a source responsible for generating 2.5% to 3.7% of global greenhouse gas emissions.
This surpasses the environmental impact of commercial flights, which account for 2.4% of emissions. It's important to note that these statistics are based on data from a few years ago, and since then, the energy demands have surged with the progress in artificial intelligence technologies.
Related Article: New Study Delves Into World of Artificial Neural Networks, Unraveling How AI Brains Learn
