A Chip With No Clock Could Make AI Decisions on Far Less Power

Imagine a factory where every worker is doing their own thing instead of all moving to the beat of one drum. That's roughly the idea behind this chip. Most computer chips have a built-in metronome, called a clock, that tells every part when to act. This team, Eric Oliveira-Gomes and Damien Rontani, built one that has no clock at all. Instead, each of its thousands of tiny logic elements flips on and off in its own chaotic rhythm — a bit like thousands of coins being flipped at once.

That randomness is exactly what an AI that learns by trial and error needs. This type of AI, called reinforcement learning, figures things out by trying options and remembering what worked — the same way you learn a video game by playing badly at first. To explore options well, it needs a steady supply of randomness. The problem is that randomness usually takes energy and space to produce. Here, the randomness comes free from the chip's own chaotic behavior, so it's cheap and comes in enormous parallel amounts.

The researchers tested their chip on a classic decision challenge: pick the best of 1,024 options, like a slot machine with 1,024 levers. That's a much bigger problem than earlier hardware could handle. They also pushed the chip to 5,120 parallel random-signal channels, producing 2.14 trillion samples per second. Crucially, they did it on a commercial, reprogrammable chip — the kind you can buy and reprogram — rather than something exotic and custom-made.

So what? For you, this is early-stage research, not a product. But it points in a direction that matters: AI that decides quickly without draining batteries or spinning up giant data centers. Think smarter hearing aids, robots, and sensors that run for months on a coin cell. The honest catch is that chaotic hardware is hard to program and harder to trust — engineers still need to prove it behaves reliably outside the lab.