Parallel convolution processing using an integrated photonic tensor core

Feb 1, 2020 by J. Feldmann, N. Youngblood, M. Karpov, H. Gehring, Xuan Li, M. L. Gallo, Xin Fu, A. Lukashchuk, A. Raja, Junqiu Liu, David Wright, A. Sebastian, Tobias J Kippenberg, Wolfram H. P. Pernice, H. Bhaskaran (arXiv.org)

Built a photonic tensor core that runs TMAC/s convolution by combining soliton microcombs, phase‑change in‑memory arrays and nonresonant passive optics so the computation reduces to transmission reads at >14 GHz — a concrete, CMOS‑wafer‑scale path for massively parallel, chip‑scale optical AI accelerators. This is not a toy demo but a systems‑level blueprint showing how integrated photonics can displace electronic bottlenecks for real‑time video, autonomous perception and cloud inferencing.

source S2