SCIENCE CHINA Information Sciences, Volume 63 , Issue 6 : 160403(2020) https://doi.org/10.1007/s11432-020-2872-3

Towards silicon photonic neural networks for artificial intelligence

More info
  • ReceivedFeb 15, 2020
  • AcceptedApr 13, 2020
  • PublishedMay 9, 2020



This work was supported by National Natural Science Foundation of China (Grant Nos. 61635001, 61822508), Beijing Municipal Science Technology Commission (Grant No. Z19110004819006), and National Key RD Program of China (Grant No. 2018YFB2201704).

  • Figure 1

    (Color online) (a) Schematic of a neuron. (b) An artificial neuron with simple nonlinear model: showing the input ($x_{1},x_{2},~\ldots~,~x_{n}$), their relevant weights ($w_{1},w_{2},~\ldots~,w_{n}$), bias $b$ and the non-linear activation function $f(x)$ applied to the weighted sum of input signals. The output is connected to other neurons through synapses (connecting links), forming a neural network.

  • Figure 2

    (Color online) Multilayer artificial neural network scheme composed of an input layer, multiple hidden layers and an output layer. The circles are neurons and each neuron is connected to all neurons in the next layer.

  • Figure 3

    (Color online) (a) An individual programmable Mach-Zehnder interferometer with two thermo-optic phase shifters [21]@Copyright 2017 Springer Nature. (b) Schematic illustration of a programmable MZI, which comprises a phase shifter ($\theta~$) between two 50:50 evanescent directional couplers, followed by another phase shifter ($\varphi$).

  • Figure 4

    (Color online) (a) Top-view SEM image of a silicon all-pass microring resonator [36]@Copyright 2010 IEEE. (b) A symmetric add-drop microring resonator on SOI with O/E conversion and amplification. (c) Output of the balanced photodiode (blue triangle curve) as a function of the detuning $\varnothing$. The orange and green lines are the transmissions of drop $T_{\textup{drop}}$ and through $T_{\textup{thru}}$ ports, respectively.

  • Figure 5

    (Color online) (a) Optical micrograph image of the silicon photonic ANN using MZI arrays with 4 input ports and 4 output ports. (b) General PNN architecture and an individual layer consisting of optical interference and nonlinearity units. Reprinted from [21]@Copyright 2017 Springer Nature.

  • Figure 6

    (Color online) (a) A micrograph image of the fabricated silicon photonic neuron. (b) Equivalent circuit diagram of the silicon MRR modulator neuron. Two photodetectors are connected to the neuron, resulting in an O/E/O nonlinear transfer function. (c) The normalized relationship between the input power and output power under different bias current $I_{b}$. Reproduced from [12]@Copyright 2019 American Physical Society.

  • Figure 7

    (Color online) (a) Optical micrograph of one complete neuron (D1) with a zoomed in ring resonator to implement activation function. (b) Schematic of a single-layer neurosynaptic system consisting of four neurons with 15 synapses each. (c) The output spike intensity of the four trained patterns (four letters A, B, C and D) illustrated on the right side. Reproduced from [27]@Copyright 2019 Springer Nature.

  • Figure 8

    (Color online) Comparison of digital electronic architectures with photonic platforms for multiply-accumulate computations (MACs) which takes the form ${b}'\leftarrow~b+w\times~x$. Here $b$ is accumulator, $w$ is multiplier and $x$ is input. Photonic neural networks have the potential to outperform digital hardwares. Taken from [26]@Copyright 2017 IEEE.

  • Figure 9

    (Color online) Block diagram of a silicon photonic AI processor. On-chip WDM technology makes full use of the advantages of large bandwidth and parallel processing of light. The control unit allows for training the PNN.

  • Table 1  

    Table 1Comparison of electronic AI chips with photonic neural networks. Modified from [29].

    Architecture Energy efficiency/MAC Vector size Latency$^{\rm~a)}$
    Google TPU [50] 0.43 pJ 256 2 $\mu$s
    Flash (analog) [51] 7 fJ 100 15 ns
    Hybrid laser neural networks [52] 0.22 pJ 56 $<$100 ps
    Integrated silicon PNN [12] 2.7 fJ 148 $<$100 ps
    Sub-$\lambda$ nanophotonics (prediction) 30.6 aJ 300 $<$50 ps

    a) Latency is the required time for completing a single MAC at the given vector size.