Q: What trends do you see in the utilize of generative artificial intelligence in computing?
AND: Generative AI uses machine learning (ML) to create novel content, such as images and text, based on data fed into the ML system. At LLSC, we design and build some of the largest academic computing platforms in the world, and over the last few years we have seen an explosion in the number of projects requiring access to high-performance computing for generative artificial intelligence. We’re also seeing generative AI transforming all kinds of fields and domains – ChatGPT, for example, is already impacting the classroom and workplace faster than regulations seem to be able to keep up.
We can imagine all kinds of applications of generative AI over the next decade, such as powering highly capable virtual assistants, developing novel drugs and materials, and even improving our understanding of basic science. We can’t predict everything that generative AI will be used for, but I can certainly say that with increasingly elaborate algorithms, their computational power, energy and climate impact will continue to grow very rapidly.
Q: What strategies does LLSC utilize to mitigate these climate impacts?
AND: We are always looking for ways to do this more efficient data processingbecause it helps our data center make full utilize of its resources and allows our research colleagues to advance their fields in the most productive way possible.
For example, we reduce the amount of energy our equipment uses by making plain changes such as dimming or turning off the lights when we leave a room. In one experiment, we reduced the power consumption of a group of GPUs by 20-30 percent, with minimal impact on their performance, by forcing power cap. This technique also lowered the operating temperature of the hardware, making GPUs easier to frosty and lasting longer.
Another strategy is to change our behavior to be more climate conscious. At home, some of us may choose to utilize renewable energy sources or sharp planning. At LLSC, we utilize similar techniques, such as training artificial intelligence models when temperatures are lower or when demand for power from the local grid is low.
We also realized that a lot of the energy used on a computer is often wasted, for example a water leak adds to the bill but without any benefit to the home. We have developed several novel techniques that allow us to monitor running compute workloads and then terminate those that are unlikely to produce good results. Surprisingly, v a number of cases we found that most calculations can be completed earlier without compromising the final effect.
Q: What is an example of a project you have completed that reduces energy production in a generative artificial intelligence program?
AND: We recently built a climate-aware computer vision tool. Computer vision is a field that focuses on applying artificial intelligence to images; i.e. distinguishing cats and dogs in an image, correctly labeling objects in an image or looking for compelling elements in an image.
We have included real-time carbon telemetry in our tool, which generates information about the amount of carbon emitted by our local network while the model is running. Depending on this information, our system will automatically switch to a more energy-efficient version of the model, which typically has fewer parameters during periods of high carbon intensity, or a much higher fidelity version of the model during periods of low carbon intensity.
By doing this, we almost saw Reduction of greenhouse gas emissions by 80 percent within one to two days. We recently expanded this idea for other generative AI tasks, such as text summarization, and achieved the same results. Interestingly, performance sometimes improved after using our technique!
Q: What can we do as consumers of generative AI to aid mitigate its impact on the climate?
AND: As consumers, we can ask our AI providers to be more crystal clear. For example, in Google Flights I see different options that indicate the carbon footprint of a specific flight. We should get similar measurements from generative AI tools so that we can make an informed decision about which product or platform to utilize based on our priorities.
We can also make efforts to learn more about AI’s generative emissions in general. Many of us are familiar with vehicle emissions, and it may be helpful to discuss AI’s generative emissions in comparative terms. People may be surprised to learn, for example, that one of the tasks involved in image generation is roughly equivalent to drive four miles in a gas car, or that charging an electric car requires as much energy as generating approximately 1,500 text summaries.
There are many cases where customers would be willing to make a trade-off if they knew the impact.
Q: What do you see for the future?
AND: Mitigating the climate impact of generative AI is one of those problems that people around the world are working on with a similar goal. We do a lot of work here at Lincoln Laboratory, but this is just scratching the surface. In the long term, data centers, AI developers and energy grids will need to work together to provide “energy audits” and discover other unique ways to improve computing performance. To move forward, we need more partnerships and cooperation.
.