New AI Method Slashes Energy Use One Hundred Times

Researchers replaced dense matrix multiplications with sparse activation patterns and low precision arithmetic on neuromorphic hardware. Their approach cut energy consumption from 100 joules per inference to under one joule. Accuracy on ImageNet rose from 76.2 percent to 78.1 percent.

This teaches that efficiency gains do not require sacrificing performance when computation is restructured at the hardware level. Users should evaluate models by energy cost per correct prediction rather than raw accuracy alone. The workflow changes from scaling up compute to designing lightweight inference pipelines.

A team at MIT published results in Nature showing their neuromorphic chip achieved 100 times lower energy draw on vision tasks. They open sourced the training code at https://github.com/mit-neuro/ai-energy-efficient.