University of Michigan Software Achieves Up to 50 Percent Savings in AI Training Energy
Engineers at the University of Michigan have developed two new software tools that cut the energy consumed in training large-scale artificial intelligence models by half.
Two software tools named Kareus and Perseus, developed by researchers at the University of Michigan College of Engineering, succeed in reducing the energy required for training large-scale artificial intelligence models by up to 50 percent.
Working Principle of the Software Tools
The two software tools, named Kareus and Perseus, work in tandem to lower the energy expended during artificial intelligence training. This system optimizes the way artificial intelligence training is coordinated across hundreds to thousands of computer processors.
Contributions of Perseus and Kareus
Perseus reduces artificial intelligence training energy by up to 30 percent by managing the speed of processors that run unnecessarily fast. Kareus, introduced in July, provides an additional 15 to 30 percent savings in training by finding the most efficient scheduling and resource sharing.
Integration Process with NVIDIA
NVIDIA is currently evaluating how these energy-saving tools can be integrated into its open-source platform, Megatron-LM. Megatron-LM is used by companies either directly or as a foundation to build custom platforms.
Future Expectations
The research team aims to enable artificial intelligence to perform more beneficial work with less energy and cost through this newly developed approach. It is stated that total energy savings could approach 40 to 50 percent in large-scale models.