New algorithm slashes AI energy consumption by two orders of magnitude while raising accuracy

Researchers replaced dense matrix multiplications with sparse, event-driven updates that fire only when activation thresholds are crossed. On standard language-model benchmarks the method cut energy per inference from 3.2 joules to 0.03 joules and lifted accuracy from 78.4 percent to 81.1 percent.

You learn that energy cost is not an immutable tax on intelligence but a tunable variable. Re-examining the arithmetic primitives inside your own pipelines can turn an expensive model into one that runs on edge devices.

The SparseCompute group at MIT CSAIL released open-source kernels that now power a 7-billion-parameter chatbot serving 12,000 daily queries on a single Raspberry Pi 5 with a measured 94-watt-hour daily budget.