Fig. 1. Conceptual architecture diagram of SPPU on chip. (a) Basic process of SPPU inference for neural networks with different structures. (b) Schematic illustration of SPPU chip. This chip includes eight independent computational channels, with each channel consisting of an edge coupler (EC), two phase modulators (PMs), two microring resonators (MRRs), a balanced photodetector (BPD), and two grating couplers (GCs).

Fig. 2. Optical micrograph of SPPU. (a) Microscopic image of the SPPU chip. (b) Image of SPPU unit after photoelectric hybrid package. Microscopic images of (c) phase modulator, (d) add-drop microring resonator, and (e) photodetector.

Fig. 3. Characterization of key photonic components on chip. (a) Measured bandwidth of the on-chip silicon phase modulator. (b) Normalized weights and the voltage applied to the on-ring heater for one microring resonator. (c) Scatter plots of the measured weights versus calculated expectations. (d) Measured bandwidth of the on-chip photodetector. (e) Responsivity of the on-chip photodetector measured at −1  V bias. (f) Dark current measured by the on-chip photodetector at different bias voltages.

Fig. 4. Working flow of one SPPU channel. (a) Schematic of an SPPU channel performing optical neural network computation. (b) Sequence calculation results when the weight is one and the input data vector is [1, 0.75, 0.5, 0.25, 0, −0.25, −0.5, −0.75, −1], which is used for SPPU channel calibration. (c) Ideal (orange line) and experimental (blue curve) convolution output waveform with 4-bit coding.

Fig. 5. SPPU for a CNN. (a) Structure of the CNN for handwritten digits classification, which consists of two convolutional layers and a fully connected layer. Here, each convolutional layer can be subdivided into convolutional operations and pooling operations. (b) Variation in calculation accuracy and loss of validation dataset during the training. (c) Confusion matrices of digit classification. The experimental accuracy (97.9%) is in good agreement with the calculated results (98.3%).

Fig. 6. SPPU for a residual CNN. (a) Structure of the residual CNN for signal modulation format classification. The CNN contains a residual block and two fully connected layers, where the residual block can be subdivided into three convolutional layers. (b) Variation in calculation accuracy and loss of validation dataset during the training. (c) Confusion matrices of modulation format classification.

Fig. 7. SPPU for a simple RNN. (a) Structure of the simple RNN for emotion recognition of reviews. It contains an embedding layer, a recurrent layer, and a fully connected layer. Variation in calculation accuracy and loss for both (b) Amazon Fine Food Reviews dataset and (c) IMDBMovie Reviews dataset, during the training. Confusion matrices of emotion recognition for the (d) Amazon Fine Food Reviews dataset and (e) IMDB Movie Reviews dataset.