Researchers are developing new approaches, such as eliminating matrix multiplication and using custom hardware, to power large language models with over 50 times greater energy efficiency than typical GPUs.
Open the full topic