A New Chip Trick Could Cut AI's Power Bill by Two-Thirds

Artificial intelligence runs on chips that do a lot of arithmetic, very fast. Today there are two main styles of that arithmetic. One is the standard approach used by nearly all AI today. The other, called "spiking" neural networks, mimics how brain cells fire — mostly staying quiet and only working when there's something worth signaling. Spiking chips can be far more energy-efficient, but they aren't equally good at every task. Until now, chips had to pick one style for a whole block of work, which meant half the hardware sat idle while the other half ran.

NeuroFlex takes a different approach: it decides, for every single output number, whether standard math or brain-style math is the better deal. The team proved the two methods produce identical results at that fine-grained level, so nothing is lost in translation. A scheduling program then scores each number by its energy-versus-speed trade-off and packs the work neatly across the chip's processors. That pushes hardware utilization from roughly 40-45% up to 97-99% — the difference between a half-empty restaurant and a packed one.

The numbers are striking. Compared with a strong standard-only chip, NeuroFlex cut combined energy and time costs by 57-67%. Against a spiking-only chip, it was up to 2.5 times faster. And its smart scheduler beat random assignment by 16-19% across vision, language, and transformer workloads — the kinds of AI behind image recognition and chatbots.

The catch: this is a research paper, not a product. The results come from simulations of a chip design, not a physical chip you can buy. Building real silicon takes years and millions of dollars, and the gains may shrink in a commercial setting. Still, the direction is clear — and the first beneficiaries would be data centers paying enormous electricity bills, which eventually shapes what you pay for AI features.