Fig. 1. Advantages of the silicon photonic computing

Fig. 2. Take deep convolutional neural network (CNN) as an example. CNN implementation contains complicated matrix productions and convolution operations, which can be accelerated by light and puts forward high request to hardware. (a) A simple CNN. The scale and architecture of actual neural networks are more complex; (b) convolution calculation. Matrix productions between convolutional kernels and features are used for numerical feature extraction; (c) linear interlayer transmission of fully connected

Fig. 3. Silicon photonic system for matrix operation. (a) A matrix can be decomposed into the product of two unitary eigenvector matrices and one eigenvalue diagonal matrix; (b) an MZI basic block can be employed to achieve matrix product with correspondingly configured phase modulator ^[18]; (c) the matrix multiplication can be implemented by a cascaded photonic network, which can be decomposed into two unitary MZI network and eigenvalue active device ar

Fig. 4. Derived computing concept of artificial neural network based on computing on silicon photonic platform. (a) Cascaded MZI matrix computation architecture optimization proposed by Fang et al.^[23]; (b) photoelectron CNN operation proposed by Bagherian et al.^[24]; (c) tunable real-time nonlinear activation function proposed by Williamson et al.^[

Fig. 5. Silicon photonic system for solving NP-complete problems. (a) One-dimensional Ising problem. Solve the minimal energy state of the array with given Hamiltonian energy; (b) MAX-CUT problem in a graph is equivalent to the Ising problem with similar Hamiltonian energy^[26]; (c) heuristic recurrent photonic Ising problem solver proposed by Roques-Carmes et al.^[27], composed of linear cascaded MZ

Fig. 6. Silicon photonic system for facilitating reservoir computing. (a) Fixed-connection recurrent reservoir can map an input spatiotemporal signal into a high dimensional space to facilitate signal classification and processing; (b) take a simple reservoir as an example. Spatiotemporal signal processing is achieved by a pre-trained recurrently-connected reservoir, which can be equivalently implemented by well-designed photonic waveguide delay lines. Through this software and hardware integrated desig

Fig. 7. Silicon photonic system for implementing discrete Fourier transform. (a) Fourier transform for incident beam can be simply implemented by an ideal thin lens, realizing reciprocity transformation between spatial distribution of electromagnetic field and wave vector space; (b) in OFDM systems, fast Fourier transform is the fundamental operation for demultiplexing of subcarriers. With the increase of OFDM subcarriers and the broadening of frequency band, a high speed and low latency DFT accelerator

Fig. 8. Silicon photonic devices for analog computing. Schematics of (a) integrator and (b) differentiator circuits in analog circuits. And they are the basic components of analog circuits; (c) all-optical temporal integrator with ~500 GHz bandwidth and 8 ps temporal resolution can be realized by using high-Q micro-ring resonator on silicon optoelectronic chips proposed by Ferrera et al.^[42];(d) Estakhri et al.^{[Download full size}

Fig. 9. Examples of silicon photonic quantum processor. (a) A programmable bipartite entangled quantum system is realized by Wang et al.^[44] with more than 550 photonic components on a silicon photonic chip, including entangled photon source; (b) a programmable two-qubit quantum processor based a silicon photonic chip demonstrated by Qiang et al.^[45] is fabricated for 98 kinds of two-qubit q

Fig. 10. Examples of silicon photonic neuromorphic processor. (a) Biological neuron models communicated with spikes, i.e., nerve impulses; (b) time course of leaky-integrate-and-fire (LIF) neuron with regards to external excitement. When the threshold is reached, the neuron will fire a new spike^[46]; (c) all-optical neuron with GST materials and MRRs proposed by Feldmann el al.^[50] on a silicon phot

Fig. 11. Proposed computing on silicon photonic platform. Collaborated with common microelectronic units, e.g. ALU, CU, registers, and memories, the OECU manipulating multidimensional information of light, such as wavelength, modes, phase, amplitude, and polarizations, to achieve efficient optoelectronic computing and break through performance limits of the current microprocessor. In addition, the generated massive data need to be handled by optical I/O for interconnects