Computer engineers pioneer approaches to energy-efficient supercomputing

https://techxplore.com/news/2024-09-approaches-energy-efficient-supercomputing.html

"trade-off between time and energy efficiency... maximum GPU frequency achieving fastest throughput not necessarily most energy-efficient frequency... pull back 5 to 10% might get 20 or 25% energy savings by adapting per application, maybe even within single run... use of mixed-precision arithmetic: 64 bit full-precision floating point arithmetic nearly as fast as single-precision arithmetic at 32, and 16 or fewer accurate enough for training neural networks/ data-science... algorithms selectively reducing data movement"