University of Michigan software could cut AI training energy use by up to 50%

Researchers at the University of Michigan College of Engineering have developed software that can save around half the energy used to train AI without affecting speed or performance. They are now working with NVIDIA to integrate the approach into Megatron-LM, NVIDIA's open-source AI training platform.

• AI training energy use can be cut by up to 50% with no performance loss
• The approach relies on coordinated software tools
• The work was done at the University of Michigan College of Engineering
• NVIDIA is integrating the method into its open-source Megatron-LM platform