Author: Evolving AI - Bewertung: 122x - Views:3747
Engineers in the U.S. just pulled off something chip companies have been talking about for years but never actually shipped: a fully monolithic 3D computer chip built inside a real commercial fab—and it already beats today’s best 2D designs. In this video, we break down how researchers working with SkyWater Technology stacked logic and memory directly on top of each other, connecting them with record-setting ultra-fine vertical interconnects to smash straight through the “memory wall” that limits Nvidia, AMD, and today’s AI accelerators. Early silicon is already showing around 4× better performance than comparable 2D chips, with simulations of taller stacks hitting up to 12× gains on real AI workloads derived from Meta’s LLaMA models. We’ll walk through how low-temperature processing lets them build multiple layers without destroying what’s underneath, why technologies like RRAM memory and future carbon nanotube transistors make vertical integration even more powerful, and how this approach could unlock 100–1000× improvements in energy–delay for AI systems that are currently choking on power and bandwidth limits. We also look at why it matters that this was done in a U.S. foundry with backing from DARPA, the Department of Energy, NSF, Samsung, and the DoD’s AI Hardware Hub, and what this means for the next decade of AI hardware as monolithic 3D starts to compete with flat GPUs in data centers. Years from now, when AI feels instantly responsive and far less power-hungry, this may be the quiet breakthrough everyone points back to. 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