In-Memory DAC Structure for Accurate Low-Movement MAC Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional computing structures are inefficient in terms of energy when performing multiply-accumulate (MAC) operations in convolution layers of artificial neural networks due to the large amount of data movement required.
Innovation Solution
A computing-in-memory device is developed, which includes a memory cell array with an analog multiplication unit that performs MAC operations on pre-stored weights and multi-bit input data converted into an analog voltage using a digital-to-analog converter (DAC) integrated within the memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If conventional computer structure with separate memory and operator is used, then data movement between memory and operator occurs, but energy consumption becomes very large
Solution Approach 1:
The patent merges the memory function and computing function into a single integrated structure. Memory cells store weights while peripheral circuits perform analog-to-digital conversion and multiplication operations, eliminating the need for separate memory and operator components and reducing energy-consuming data movement between them.
Solution Approach 2:
The patent implements a nested structure where analog multiplication units are embedded within memory blocks. The peripheral circuits containing ADCs and multipliers are integrated within the memory cell array structure, allowing computing operations to be performed directly where data is stored.
2Measurement precision
If multi-bit input data is converted to analog voltage using DAC inside memory, then accuracy of MAC operation increases, but device complexity increases
Solution Approach 1:
The patent makes the peripheral circuit universal by enabling it to perform multiple functions: storing weight data in memory cells, converting multi-bit digital input signals to analog voltages via ADCs, and performing multiplication operations. This multi-functional design achieves high MAC operation accuracy without requiring separate dedicated circuits for each function.
Solution Approach 2:
The patent changes the parameter of input data precision by supporting multi-bit (e.g., 4-bit, 6-bit) input data conversion to analog voltage, thereby improving MAC operation accuracy. The ADC structure is designed to handle multiple bits simultaneously, transforming digital precision requirements into analog voltage precision.
Data Source
AI summary
Disclosed is a computing-in-memory device, which includes a memory cell array including an analog multiplication unit that performs a multiply-accumulate (MAC) operation on a pre-stored weight and a first analog voltage corresponding to multi-bit input data, and a driver that applies the multi-bit input data to the analog multiplication unit, and the analog multiplication unit includes a digital-to-analog converter that converts the multi-bit input data into the first analog voltage.