Author: Evolving AI - Bewertung: 218x - Views:7399
Extropic is trying to do something no GPU company ever dared to try: turn the “noise” inside a chip into the fuel for intelligence. Instead of fighting thermal jitter with more cooling and tighter tolerances, their thermodynamic computing chips use probabilistic bits—P-bits—that physically wobble between 0 and 1, then link thousands of them into thermodynamic sampling units that naturally relax into the most likely answer. In a small benchmark, their Denoising Thermodynamic Model used an estimated ten thousand times less energy than an Nvidia GPU running the same task, a number so extreme it sounds fake until you dig into the math and simulation code they’ve already published. This video breaks down how P-bits work, what makes thermodynamic sampling different from classic matrix-multiply AI, and why Extropic’s early X0 dev board and upcoming Z1 chip won’t just run today’s models faster—they’ll need completely new algorithms like DTMs that think in probability from the ground up. We’ll talk honestly about the catch: the 10,000× figure comes from a tiny toy problem, the Z1 is still in development, and you can’t just drag-and-drop a giant GPU transformer onto this hardware and expect it to run. But if this approach scales, it could rewrite the entire AI energy story, turning billion-dollar datacenters into something closer to a shoebox of ultra-efficient thermodynamic processors and making frontier-level AI possible on phones, cars, and even glasses. In short, this is the first serious candidate for a post-GPU future. And if you want the real story behind the world’s fastest-moving AI breakthroughs, make sure to like and subscribe to Evolving AI for daily coverage.
SOCIAL SHARE CARD GENERATOR