As artificial intelligence software like ChatGPT continues to grow in popularity, environmental advocates have raised concerns over the high levels of energy consumed by AI. In a step toward addressing this issue, U-M alum Jie You, Rackham student Jae-Won Chung and Mosharaf Chowdhury, associate professor of electrical engineering and computer science, developed Zeus, an open-source framework that analyzes the energy efficiency of training AI. According to the study, implementation of this software could reduce AI energy consumption by up to 75%.
Zeus works by taking into account any time constraints for the AI training process and then figuring out the most energy-efficient way to complete the training within that time frame.
In an interview with The Michigan Daily, Chowdhury said increasing data consumption during training, which is necessary for AI to function, contributes to high levels of energy consumption. According to Chowdhury, reducing energy usage by even a small amount during the training process has a positive environmental effect, while maintaining minimal impact on its timeframe.
“What we found is that for different models, even if you slightly reduce the power they are drawing — instead of drawing 300 watts, they are drawing 250 watts — the speed of computation (doesn’t) get affected significantly, meaning (the training) will still finish roughly at the same time,” Chowdhury said.
Chung said the research team initially struggled to find a balance between energy use reduction and training speed.
“We found out that if you reduce energy too much, the training gets too long,” Chung said. “So we developed our problem formulation where we find a sweet spot between reducing energy and not inflating computation time too much.”
Rising LSA senior Jackson Leb is the founder of OpenMI, a student organization focused on AI and machine learning. Leb said it is difficult to decrease energy consumption by AI because of the sheer amount of data AI uses.
“These machine learning and artificial intelligence models that we see today, like ChatGPT and a lot of the more common ones, are built and function off these huge amounts of data that are pulled from data centers,” Leb said. “(It) requires a lot of computational power to get all that data, and a lot of computational power means that there’s a lot of energy consumption which is going to have a detrimental effect on the environment. There’s no … way to escape (the negative impacts) if you’re using that much energy.”
Chowdhury said Zeus allows the users to decide what trade-off they want to make — they can decide between saving more energy or having their AI trained quicker, depending on their specific needs.
“You can decide that I want to save some energy, and I’m fine with taking a little bit more time, or vice versa,” Chowdhury said. “Depending on what you want to optimize, Zeus allows someone to decide that maybe (they are) fine with (the training) taking one more day, but this will save maybe 20% more energy.”
Along with Chung, Rackham student Zhenning Yang and U-M alum Luoxi Meng developed a complementary software called Chase that aims to further reduce the carbon footprint of AI by looking not at just the amount of energy consumed, but also where the energy is coming from. This software increases the training speed when low-carbon energy is available and decreases it when there is higher-carbon energy available.
Chung said Chase works the same way Zeus does, speeding up when there is more clean energy available and slowing down when there is less.
“Carbon intensity changes throughout the day,” Chung said. “For example, when the sun is up, solar power is up, so electricity tends to get greener compared to night. (Chase uses) the same mechanism that Zeus developed, which is slowing down the GPU so that it consumes less energy but slightly slower. We do that while the job is training. When it’s daytime and we have greater energy, we speed up the GPU … but during night we don’t, so we slow down the GPU automatically.”
Chowdhury said the research team hopes the creation of Zeus will spark further conversation and action around the environmental effects of AI .
“Historically, what we found is that even though software consumes a lot of energy, there hasn’t been that much work (on how to make it more sustainable),” Chowdhury said. “It’s our small part that we are trying to play, but I feel like still (our contribution is) quite sizable (as) close to one percentage point of total energy consumption of the world can be saved.”
Summer News Editor Rebecca Lewis can be reached at rebeccl@umich.edu.