Compute has scaled by six orders of magnitude since the 1990s; electrical interconnect by only three. As frontier AI workloads become interconnect-bound in both training and inference, that gap is now the defining constraint on supercomputer design. I will present Lightmatter's 3D silicon photonics platform — bidirectional DWDM co-packaged optics, a photonic interposer integrating compute ASICs directly on silicon, and an integrated laser architecture — with measured results up to >100 Tbps per package at sub-3 pJ/bit, and the system-level consequences for collapsing fragmented GPU domains into a single coherent compute fabric.
Speaker's Bio
Dr. Darius Bunandar, Co-founder and Chief Scientist at Lightmatter, spearheads the company's pioneering research and development in novel hardware solutions, with a significant focus on advancing 3D photonic architectures for next-generation AI infrastructure. Darius holds over 50 patents and has written over 40 multidisciplinary journal articles. He earned his PhD in Physics from MIT with Prof. Dirk Englund in 2019.