Fixed-point quantitative convolutional neural network accelerator calculation circuit

Abstract

The invention discloses a fixed-point quantitative convolutional neural network accelerator calculation circuit, and belongs to the technical field of integrated circuits. The method comprises the following steps of: processing input data of N input channels by using N input channel processing units, correspondingly multiplying a plurality of input feature maps and a plurality of weights in the input data of each input channel, and quickly adding to obtain convolution results of the N input channels; a partial accumulation unit accumulates all convolution results of the N input channels and outputs the convolution results to a quantization activation unit; and the quantization activation unit sequentially performs bias accumulation, approximate multiplier multiplication, right shift, function activation, zero point data addition and output digit amplitude limiting to obtain an output result of the convolutional neural network accelerator calculation circuit. According to the method, the calculation speed of the fixed-point quantized convolutional neural network is increased on the premise of not generating obvious precision loss, and the method has the characteristics of low powerconsumption and relatively small circuit area and is suitable for a convolutional neural network system needing to use shaping quantization.

Description

technical field [0001] The invention belongs to the field of integrated circuits and relates to a computing circuit of a fixed-point quantized convolutional neural network accelerator. Background technique [0002] Convolutional Neural Networks (CNNs) have achieved great success in the field of image recognition due to their excellent predictive performance. Nonetheless, modern CNNs with high inference accuracy usually have large model sizes and high computational complexity, which complicates their deployment in data centers or edge devices, especially for applications that require low resource consumption or low response latency. application scenarios. To facilitate the application of complex CNNs, an emerging field of model compression research focuses on reducing the model size and execution time of CNNs with minimal loss of accuracy. [0003] The accuracy of network parameters can be quantized as low as 1 bit, and XNOR-net and related network variants can achieve a 32...

AI Technical Summary

Problems solved by technology

[0004] Although fixed-point quantization has reduced the complexity of convolution calculations, for deep neural networks, there are still a lot of calculations

Images