Digital signal modulation method of photon artificial intelligence computing chip

Abstract

The invention provides a digital signal modulation method of a photon artificial intelligence computing chip. The method comprises the following steps: modulating one or more digital electric signal groups into optical signals, wherein the digital electric signal group comprises a plurality of time sequence signals which are sequentially output in a fixed time period, each time sequence signal hasthe same basic clock and signal time length, a single time sequence signal for transmitting N-bit digital information is provided with 2N-1 basic clocks, the number value of the basic clocks occupiedby a high-level signal or a digital signal '1' in the time sequence signal is the signal value of the time sequence signal and is equal to the value of the transmitted N-bit digital information, andthe time sequence signal is a modulation signal for converting an electric signal into an optical signal. Compared with an existing calculation scheme based on digital-to-analog conversion and analogsignal modulation, a digital-to-analog conversion converter high in cost and power consumption is not needed, the method can be directly connected with digital signals of an electronic chip, and quantization errors generated when the digital signals are subjected to digital-to-analog conversion are avoided.

Description

technical field [0001] The invention relates to the technical field of optical computing, in particular to a digital signal modulation method of a photonic artificial intelligence computing chip. Background technique [0002] Deep learning (Deep Learning, DL) is one of the technology and research fields of machine learning. By establishing artificial neural networks (Artificial Neural Networks, ANNs) with a hierarchical structure, artificial intelligence is realized in computing systems (see Lecun, Y., Bengio, Y., Hinton, G., et al. (2015) Deep Learning. Nature, 521, 436.). Deep learning proposes a method for the computer to automatically learn pattern features, and integrates feature learning into the process of building models, thereby reducing the incompleteness caused by artificially designed features. In recent years, thanks to the development of big data, the substantial improvement of computer computing power and the breakthrough of the algorithm itself, deep learnin...

AI Technical Summary

Problems solved by technology

However, this chip uses analog signals to modulate and calculate light (see B, Hengameh, et al. "On-Chip Optical Convolutional Neural Networks" arXiv:1808.03303(2018).), which has low precision and anti-interference ability Poor problem, and higher requirements on the dynamic range of electro-optic modulators and photodetectors

At the same time, AD / DA conversion is required when communicating with digital chips such as FPGAs, which additionally brings quantization errors and increases system costs

Images