Current integration and charge sharing multi-bit convolution operation module with variable capacitance

Abstract

The invention relates to an analog operation module, in particular to an analog operation module related to convolution operation, and provides a group of analog multipliers and accumulators (MAC). The current integration in the capacitor is used for multiplication of the convolution process of two multi-bit binary numbers; change of multiplier or multiplicand bit weight in a multiplication stageis realized by utilizing capacity double decreasing coding of capacitors with the same clock period, an addition process is realized by charge sharing among the capacitors, and the module has higher speed and the same unit clock during convolution operation and can be used for exchange delay or acceleration in an area. This idea is applicable to a series of binary bit number adjustable multi-bit convolution that can be used to achieve general convolution with two or more inputs. In particular, a bias arithmetic unit array may be added. The operation module can be used as a neural network convolution operation unit or a memory or near memory operation unit realized by operation accelerator hardware.

Description

technical field [0001] The invention relates to an analog operation module, in particular to an analog operation module related to convolution operation, and also relates to an analog calculation method for convolution operation. Background technique [0002] For the quantization of low signal-to-noise ratio, analog operation has higher efficiency than traditional digital operation. Therefore, digital quantities are usually converted into analog quantities before operation. Especially for neural networks, compared to its computational energy consumption, in the medium and large-scale hardware implementation of neural networks, since traditional data is stored in disks, data needs to be extracted into memory when performing operations, and this process requires a lot of I / O Storage connected to traditional memory tends to consume more power. However, based on analog memory and near-memory computing, the computing process can be sent to the data for local execution, which gre...

AI Technical Summary

Problems solved by technology

In the implementation methods of these multi-bit operations, multi-bit analog multipliers and accumulators have always been controlled by very complex digital processing, but in terms of quantization with low signal-to-noise ratios, traditional digital operations consume a lot of power compared to analog operations, so these Multi-bit operations under the control of digital processing will generate a lot of computing energy consumption

[0007] The binarized convolution proposed by CN201910068644 performs the XOR operation stage by modulating the control bus in the SRAM to realize the potential change, but the technical solution and teaching given in this patent require complex digital processing control, and the control module The requirements are high and consume too much energy

Images