Vijay GadepallyElder employee MIT Lincoln Laboratory, conducts a number of projects in Lincoln Supercomputing Center laboratory (LLSC) In order to create computing platforms and artificial intelligence systems that work on them, more efficient. Here, Gadeply discusses the growing use of generative artificial intelligence in everyday tools, hidden impact on the environment and some of the ways in which Lincoln laboratory and a larger AI community can reduce emissions for a more green future.
Q: What trends do you see to the extent, how is generative AI used in calculating?
AND: Generative AI uses machine learning (ML) to create new content, such as images and text, based on data entered into the ML system. In LLSC, we design and build one of the largest academic computer platforms in the world, and in the last few years we have seen an explosion of the number of projects requiring access to high -performance calculations for generative artificial intelligence. We also see how generative artificial intelligence changes all kinds of fields and domains – for example, chatgpt already affects the class, and the workplace may seem to keep up faster than the regulations.
We can imagine all kinds of applications to generative artificial intelligence in the next decade, such as supplying highly talented virtual assistants, developing new drugs and materials, and even improving our understanding of basic science. We cannot predict everything that generative artificial intelligence will be used for, but I can certainly say that thanks to the increasingly complex algorithms, their calculations, energy and climate impact will continue to grow very quickly.
Q: What strategies does LLSC use to alleviate this impact on the climate?
AND: We are always looking for ways to do Calculation of more efficientBecause it helps our data center fully use its resources and allows our scientific colleagues to push their fields in such an efficient way.
As one example, we reduce the amount of power supply that our equipment uses by introducing simple changes, as well as darkening or turning off the lights after leaving the room. In one experiment, we reduced the energy consumption of a group of graphics processing units by 20 to 30 percent, with minimal impact on their performance, by enforcing capacity. This technique has also reduced the working temperatures of the equipment, making GPU easier to cool and longer.
Another strategy is to change our behavior so that it is more aware of the climate. At home, some of us can choose renewable energy sources or an intelligent schedule. We use similar techniques in LLSC – such as AI training models when the temperatures are cooler or when the local demand for network energy is low.
We also realized that a lot of energy spent on calculations is often wasted, such as water leakage increases the bill, but without any benefits for your home. We have developed several new techniques that allow us to monitor the computing loads during work, and then finish those that are unlikely to bring good results. Surprising, in Many cases We found that most of the calculations may be completed earlier no harm.
Q: What is the example of a completed project that reduces the energy efficiency of the AI ​​generative program?
AND: We have recently built a tool for a computer vision of climate awareness. A computer vision is a domain that focuses on the use of artificial intelligence to images; Therefore, differentiation between cats and dogs in the picture, correctly marking objects in the image or looking for interesting elements in the image.
In our tool we included real -time coal telemetry, which produces information on how much coal is emitted by our local network as a model. Depending on this information, our system will automatically switch to a more energy -saving version of the model, which usually has fewer parameters, in times of high coal intensity or a much larger version of the model at times of low coal intensity.
In this way we saw almost 80 percent reduction of coal emissions For one to two days. Recently expanded this idea For other AI generative tasks, such as a text summary and found the same results. Interestingly, performance sometimes improved after using our technique!
Q: What can we do as consumers of generative artificial intelligence to alleviate its impact on the climate?
AND: As consumers, we can ask our AI suppliers to offer greater transparency. For example, in Google Flight I see various options that indicate a carbon trail of a specific flight. We should get similar types of measurements from AI generative tools so that we can make a conscious decision from which product or platform we use based on our priorities.
We can also try to develop more generative AI emissions in general. Many of us know vehicle emissions and can help in a conversation about generative AI in comparative terms. People may be surprised, for example, that one task of generating an image is Roughly equivalent For driving four miles by a gas car or that it requires the same energy to charge an electric car as generating about 1,500 text summaries.
There are many cases in which customers would like to compromise if they knew the influence of the compromise.
Q: What do you see in the future?
AND: Righting the influence of the artificial intelligence generative climate is one of the problems that people around the world are working on and has a similar goal. We do a lot of work here in Lincoln laboratory, but it's just scratching on the surface. In the long run, data centers, AI programmers and energy nets will have to cooperate to ensure “energy audits” to discover other unique ways to improve computing performance. We need more partnerships and greater cooperation to prepare forward.
If you want to learn more or cooperate with Lincoln Laboratory on these efforts, please contact Vijay Gadepally.
