Version 3 2024-05-23, 16:00Version 3 2024-05-23, 16:00
Version 2 2024-03-15, 16:00Version 2 2024-03-15, 16:00
Version 1 2024-03-09, 17:00Version 1 2024-03-09, 17:00
preprint
posted on 2024-03-15, 16:00authored byZhaoang Deng, Jie Liu, Chuyao Bian, Jiaqing Li, Ranfeng Gan, Zhenhua Li, Zihao Chen, Kaixuan Chen, Changjian Guo, Liu Liu, Siyuan Yu
Advancements in artificial intelligence (AI) and neuromorphic computing increasingly rely on the integration of photonics to achieve breakthroughs in processing capabilities. Our pioneering work introduces a photonic linear processing unit (LPU) that utilizes a cascading modulator structure incorporating micro-rings. This device, elegantly designed with only two thin-film lithium niobate (TFLN) modulators coupled with a micro-ring, stands as a paradigm of innovation that merges low-power consumption with formidable computational throughput, achieving 36.7 billion operations per second (GOPs).The crux of our design lies in the ring-modulator's flexible architecture, engineered as a compact and singular unit, which markedly streamlines system complexity while bolstering energy efficiency. It adeptly facilitates large-scale dot-product operations, supporting vector dimensions up to 5832, an impressive feat by current standards. Furthermore, this ring-modulator-based LPU exhibits proficiency in image classification, processing 28*28-pixel resolution imagery post hardware training. As data volume demands surge, our architecture offers a scalable solution through parallel matrix multiplications, which hinge solely on increased modulation rates. This development paves the way for a new class of photonic processors that promise to handle escalating data workloads with unprecedented flexibility and efficiency.
History
Disclaimer
This arXiv metadata record was not reviewed or approved by, nor does it necessarily express or reflect the policies or opinions of, arXiv.