[24a-31A-1~9] FS.1 Focused Session "AI Electronics"

9:00 AM - 9:15 AM JST(12:00 AM - 12:15 AM UTC)

[24a-31A-1] Analog Photonics Simulation Based on Hardware Experiment:
Case Studies for Large-Scale AI-Model and LORA-Based Fine Tuning

〇Mitsumasa Nakajima¹, Kohei Ikeda², Satoshi Kawakami³, Toshikazu Hashimoto¹ (1.NTT Device Technology Labs., 2.NTT Basic Research Labs., 3.Kyusyu Univ.)

Keywords:

Optical computing,Analog computing,Neuromorphic

Photonic analog tensor processing has attracted much attention in recent years because of its high speed and low power operation by utilizing the spatial, wavelength, and temporal parallelism of light. On the other hand, such analog computation cannot avoid analog errors: e.g, noise of optical signals and control errors of physical hardware. As the most of the previous reports on optical computation machines have been limited to simple tasks, their applicability to large-scale models under the analog error existence has been unclear. In this report, we will examine the applicability of this method to large-scale AI models by using simulation model based our constructed optical computing hardware. We revealed that the optical computing can scale to practical models like transformer with the acceptable noise level.