Acta Photonica Sinica, Volume. 50, Issue 10, 1020001(2021)
Progress and Prospects of Photonic Neuromorphic Computing（Invited）
Figures & Tables(9)
Table 1. Key performance of optical spiking neurons
View tableView in ArticleTable 1. Key performance of optical spiking neurons
Optical piking neurons Power Speed Cascadability Footprints Injection scheme Pump VCSEL-SA mW sub-ns Yes Big Electrical Electrical DFB Laser mW sub-ns Yes Big Electrical Electrical VCSEL mW sub-ns Yes Big Coherent optical Optical Quantum-dot laser mW sub-ns Yes Big Coherent optical Electrical Micropillar laser mW sub-ns Yes Big Electrical Electrical PCM mW ns No Small Photonic laser pulses /
Table 2. Comparison of optical matrix calculation methods
View tableView in ArticleTable 2. Comparison of optical matrix calculation methods
Method Integration Refresh rate/GHz Computing range Computing method MRRs Integrated >10 Real field Explicit computing MZIs Integrated >10 Complex field Implicit computing
Table 3. Key process of optical neural networks
View tableView in ArticleTable 3. Key process of optical neural networks
Year First author/Journal Type Network scale Performance 2014 VANDOORNE K，Nature Communications［81］ The integrated passive silicon photonics reservoir chip on a silicon platform A 16-node square mesh reservoir.The chip contained waveguides，splitters，and combiners Performed arbitrary Boolean logic operations with memory，5-bit header recognition up to 12.5 Gbit/s，and classification of spoken digits 2017 SHEN Yichen，Nature Photonics［70］ Feedforward fully connected optical neural network based on MZI 4×4 weight matrix based on56 MZIsThe neuron was simulated by a nonlinear activation function in the electrical domain The accuracy of vowel recognition is 76.7%，which is more than two orders of magnitude faster than the latest electronic chip at that time，but the energy used is less than one thousandth 2017 TAIT A N，Scientific Reports［59］ Recurrent silicon photonic neural network based on MRR 4×4 weight matrix based on 16 MRRs24 optical neurons based on EOM The network had a 294-fold acceleration against a conventional benchmark in performing a differential system emulation task 2018 LIN Xing，Science［82］ 3D-printed diffractive deep neural network Classification（Imaging）network：5-layers，200×200（300×300）neurons of a layer In the task of handwritten digit classification，the accuracy was 91.75% for a five-layer design，and 93.39% for a seven-layer design 2019 FELDMANN J，Nature［80］ Optical spiking neural network based on PCMs and MRRs 4 spiking neurons based on MRRs with PCMs60 optical synapses based on PCMs and integrate waveguide The network can implement supervised and unsupervised learning and was able to successfully classify the four 15-pixel images 2019 ZUO Ying，Optica［83］ All-optical neural network（AONN）with linear operations and nonlinear activationfunctions Two-layer AONN：16×4×2 A two-layer AONN can classify the phases of a prototypical Ising model，and successfully capture the essential features that distinguish the order and disorder phases 2019 BERNSTEIN L，Physical Review X［84］ A new type of photonic accelerator based on coherent detection The number of neurons can be extended to ≥106 by the massive spatial multiplexing enabled by standard free-space optical components The standard quantum limit can be as low as 50 zJ/MAC when neural networks are trained on the MNIST dataset 2020 RAFAYALYAN M，Physical Review X［85］ Reservoir network based on the spatial light modulator and scattering medium Up to 50000 optical nodes The network successfully predicted on large spatiotemporal chaotic datasets 2020 SHI Bin，IEEE JSTQE［86］ Feedforward neural network based on SOAs 8×8 InP on-chip weighted circuitsThe non-linear function of neuron was implemented via software The prediction accuracy of the Iris flower classification problem by the 3-layer photonic deep neural network was 85.8% 2020 BANGARI V，IEEE JSTQE［62］ Digital electronics and analog photonics for convolutional neural networks（DEAP-CNNs） There are up to 1200 MRRs in the weight bank array theoretically DEAP-CNN was 2.8 to 14 times faster while using almost 25% less energy than current state-of-the-art graphical processing units. Overall accuracy was 97.6% for a test set of 500 images in the MNIST task 2021 FELDMANN J，Nature［87］ A specific integrated photonic hardware accelerator based on photonic tensor core 16×16 PCM integrated array to realize the matrix vector multiplication The accelerator operated at the speed of 1012 MAC operations per second 2021 XU Xingyuan，Nature［88］ Optical vector convolutional accelerator Ten 3×3 convolutional kernels.Optical frequency combs provide 90 optical signals with different wavelengthsUsing wavelength division multiplexing，time division multiplexing and space division multiplexing The computing speed of a single processor exceeded 10 TOPS，and the accuracy of handwritten digit images recognition was 88%
Table 4. Photonic AI chip company
View tableView in ArticleTable 4. Photonic AI chip company
Company Date of
establishment
Founder/Team/Place Achievements Lightelligence 2017 MIT，
Dr. SHEN Yichen
In April 2019，Lightelligence released the prototype of its optical AI computer，the first of its kind in the world
In 2021，the world's first commercial optical chip will be available soon
Lightmatter 2017 MIT In 2020，the chip Mars for AI inference acceleration was presented. It is planned to launch its first optical AI chip Envise by the end of 2021 Optalysys 2013 University of Cambridge In 2015，an optical computing prototype was created，with a processing speed of about 320Gflops and very low energy efficiency. On March 7，2019，FT：X 2000 which is the world’s first optical co-processor system for AI computing was announced Fathom Computing 2014 Britain Photon prototype computer was the first time that machine learning software used laser pulse circuits instead of power for training. In 2014，the accuracy of handwritten digits recognition was only about 30%，and by 2018，it had exceeded 90% Ayar Labs 2015 MIT Ayar Labs demonstrated the industry's first terabit optical link for co-packaged optics and chip-to-chip connectivity，which provided optical communication with high bandwidth，low delay，and low power consumption LightOn 2016 France A coprocessor Aurora has been fabricated，embedded with a very efficient optical core Nitro Photoncounts 2017 Beijing Jiaotong University，
Dr. BAI Bing
The field programmable photonic gate arrays chip has been developed，and a server-oriented photoelectric hybrid AI accelerated computing card has been built together with Beijing universities Luminous Computing 2018 Princeton University The scheme of Broadcast and Weight based on Micro-ring optical filters was used. The prototype at that time saved three orders of magnitude more energy than other most advanced AI chips
Tools
Get Citation
Copy Citation Text
Shuiying XIANG, Ziwei SONG, Shuang GAO, Yanan HAN, Yahui ZHANG, Xingxing GUO, Yue HAO. Progress and Prospects of Photonic Neuromorphic Computing（Invited）[J]. Acta Photonica Sinica, 2021, 50(10): 1020001
Paper Information
Category:
Received: Jul. 21, 2021
Accepted: Aug. 31, 2021
Published Online: Nov. 3, 2021
The Author Email: Shuiying XIANG (syxiang@xidian.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20