Researchers at the University of Michigan have created an open source optimization framework called Zeus that addresses the problem of power consumption in deep learning models. As the trend to use larger models with more parameters grows, the power demand to train these models also increases. Zeus seeks to solve this problem by identifying the optimal balance between power consumption and training speed during the training process without the need for hardware changes or new infrastructure.
Zeus achieves this by using two software knobs: the GPU power limit and the deep learning model batch size parameter. The GPU power cap controls the amount of power consumed by the GPU, and the batch size parameter controls how many samples are processed before the model representation of the data relationships is updated. By adjusting these parameters in real time, Zeus seeks to minimize energy use and have the least possible impact on training time.
Zeus is designed to work with a variety of GPUs and machine learning tasks and can be used without changes to hardware or infrastructure. In addition, the research team has also developed complementary software called Chase, which can reduce the carbon footprint of DNN formation by prioritizing speed when low-carbon power is available and efficiency during peak hours.
The research team aims to develop solutions that are realistic and reduce the carbon footprint of DNN training without running afoul of constraints such as large data sets or data regulations. While deferring training jobs to greener time frames may not always be an option due to the need to use the most up-to-date data, Zeus and Chase can still provide significant energy savings without sacrificing accuracy.
The development of Zeus and companion software like Chase is a crucial step in addressing the power consumption issue of deep learning models. By reducing the power demand of deep learning models, researchers can help mitigate AI’s impact on the environment and promote sustainable practices in the field. Optimizing deep learning models through Zeus doesn’t come at the cost of accuracy, as the research team has shown significant power savings without impacting training time.
In short, Zeus is an open source optimization framework that aims to reduce the power consumption of deep learning models by identifying the optimal balance between power consumption and training speed. By adjusting the GPU power limit and batch size parameter, Zeus minimizes power usage without affecting accuracy. Zeus can be used with a variety of GPU and machine learning tasks, and Chase companion software can reduce the carbon footprint of DNN training. The development of Zeus and Chase promotes sustainable practices in the field of artificial intelligence and mitigates its impact on the environment.
review the study 1, studio 2and Github. Don’t forget to join our 19k+ ML SubReddit, discord channel, and electronic newsletter, where we share the latest AI research news, exciting AI projects, and more. If you have any questions about the article above or if we missed anything, feel free to email us at [email protected]
🚀 Check out 100 AI tools at AI Tools Club
Niharika is a technical consulting intern at Marktechpost. She is a third year student, currently pursuing her B.Tech from the Indian Institute of Technology (IIT), Kharagpur. She is a very enthusiastic individual with a strong interest in machine learning, data science, and artificial intelligence and an avid reader of the latest developments in these fields.
