Researchers slash AI power draw by two orders of magnitude while lifting accuracy

A team replaced standard matrix multiplications with a sparse, event-driven algorithm that activates only 1 percent of weights per inference. On ImageNet they measured 100 times lower joules per image and a 0.8-point accuracy gain over the dense baseline. The method was validated on both an NVIDIA A100 and an edge TPU.

You stop treating model size as destiny and start measuring joules per token. Track energy alongside accuracy in every benchmark so you can pick the cheapest model that still meets your quality bar.

The Sparse Inference Lab at ETH Zurich reports 94 percent energy reduction on BERT-base with no loss in GLUE score; their open implementation is already used by two Swiss fintech startups for on-premise document classification.