[20a-A303-1~8] FS.1 Focused Session "AI Electronics"

11:15 AM - 11:30 AM JST(2:15 AM - 2:30 AM UTC)Highlighted Presentation

[20a-A303-8] 32×100 WDM-SDM Defined Photonic Vector-Matrix Multiplication

〇Mitsumasa Nakajima¹, Kenji Tanaka¹, Katsuma inoue², Kohei Nakajima², Hashimoto Toshikazu¹ (1.NTT Device Technology Labs., 2.Tokyo Univ.)

Keywords:

Photonic computing,Neuromorphic,Machine learning

Photonic analog computing is raising interests as it promises massive parallelized computation with low energy consumption for tensor processing in machine learning. One of the key advantages of photonic tensor processing is ultrawide bandwidth of light, which is accessible by utilizing wavelength division multiplexing (WDM) technique. Recent demonstration revealed that photonic engines can execute the tensor processing with competitive computation speed (tera-scale operation per second, OPS). However, their performance is still limited by the scalability of the optical filters for weighting. Here, we experimentally demonstrate 32×100λ multiplexed photonic tensor processor based on spatial and planar optical circuit technique, which is considered to be capable of sub-peta scale computation speed.