Programming, algorithms, systems, data structures–these principles are hardwired into every computer science degree. But at Trinity University, CS majors also get a side of social responsibility in their course of learning. It was after taking Trinity’s Computing History and Ethics elective when Nitya Donthireddy ’28 and Iniya Gopal ’28 started thinking about researching the environmental impacts of artificial intelligence (AI).
“It started with seeing the rise in AI,” Gopal says, “and with every new technology there’s always a cost.” Specifically, she and Donthireddy were hearing about the large amounts of energy it takes to run an AI data center. In fact, Texas recently put a halt to building more of these computing warehouses in the state until audits can be done on their potential strain on the energy grid.
Last spring, Gopul and Donthireddy approached Johanna Jacob, Ph.D., assistant professor of computer science, with a Summer Undergraduate Research Fellowship idea. “AI isn’t going away any time soon, and so instead of trying to get rid of it, we should focus on regulating its use and developing future models to be more sustainable,” Donthireddy says.
Gopal and Donthireddy knew that green AI organizations and AI companies themselves use certain tools to measure how much energy AI computing workloads consume. But they wondered: How accurately do these tools capture real energy consumption? Turns out, not very.
In their Trinity lab, the team tested three popular tools. These tools are open-source, created by different research groups, and can be used by anyone to measure carbon emissions of machine-learning workloads. The Trinity researchers measured each tool against a physical electricity meter while running six open-source language and vision models on an NVIDIA graphics processing unit (GPU). A GPU is a chip designed to make many calculations at the same time, the way AI does. NVIDIA’s GPUs are widely used to train and run AI models.
To their surprise, none of the tools perfectly matched Gopal and Donthireddy’s real-world measurements: CodeCarbon was the most accurate, with an average error of about 14%, followed by CarbonTracker at 21%, and Eco2AI at 47%. The findings are significant because researchers increasingly rely on such tools to report AI’s environmental footprint.
It was clear that if there are measurement discrepancies on their small-scale testing, much larger commercial models at AI data centers might see even wider measurement gaps.
The next step is to see whether their results can be replicated using a more powerful GPU. With funding from a Clare Boothe Luce research grant, the team can purchase additional computing equipment during this academic year, or adapt equipment Trinity has on hand to continue their research.
“I hope that our findings will help researchers more accurately estimate the environmental impacts of generative AI,” Gopal says, “which is important in shaping policy and how we will use AI in the future.”
This type of undergraduate research is unique to liberal arts schools like Trinity, and learning how to innovate with a heart is, too. “I feel like us CS students aren’t always taught the effects of what our work can do,” Gopal says. She’s right. Researchers at Virginia Tech looked at computer science bachelor degree programs at 250 universities and found that only about half offer ethics courses like the one Gopal and Donthireddy took at Trinity.
Jacob says the liberal arts model at Trinity aims to help students go beyond traditional thinking, even in technical fields. “Students have an increased realization of what they build and who it benefits, who it harms, and their responsibilities to the computing society and to the broader community, in general,” she says.
The purpose of research by undergraduates at Trinity is to provide a foundation for conducting preliminary analyses, testing hypotheses, reproducing and validating existing work, identifying limitations or unexpected outcomes, and generating questions that can inform future investigation, says Jacob.
“While publication is certainly a valuable outcome when the work reaches that stage,” Jacob says, “I don’t believe it should be the sole factor of the success of an undergraduate summer research experience.”
And Jacob, whose academic focus is cybersecurity, found herself exploring AI models right along with Gopal and Donthireddy. She has been impressed by the students’ willingness to dig into the grind of research along the way. “They were very hard-working. They asked a lot of questions, developing a critical thinking mindset, and dove deep into the literature,” she says. “That really paid off. For them, this research is foundational to their next steps at Trinity and beyond.”