Image Sensor ACiM Architecture for In-Sensor Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current electronic devices face increased power and delay overhead in processing large-capacity data due to memory data access and transfer, particularly in AI applications like deep learning, where existing solutions do not effectively integrate analog computing-in-memory (ACiM) technology with image sensors.
Innovation Solution
An image sensor design incorporating analog computing-in-memory (ACiM) technology, featuring a pixel array and memory cell array on separate dies, with control signals generated to process pixel and convolution signals, allowing for efficient data processing by converting analog signals to digital and vice versa using analog-to-digital converters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If traditional digital signal processing is used in image sensors, then processing accuracy is maintained, but power consumption and delay increase significantly due to memory data access and transfer overhead
Solution Approach 1:
The patent merges the pixel array and memory cell array into a single integrated circuit structure where pixels and memory cells share common readout circuits and column lines. This integration enables analog computing-in-memory operations directly within the sensor array, eliminating the need for separate memory access and transfer operations, thereby reducing power consumption and processing delay while maintaining computational accuracy.
Solution Approach 2:
The patent replaces traditional digital signal processing mechanisms with analog computing operations performed directly in the memory array. Instead of transferring data to external processors for digital computation, the system performs matrix multiplication and convolution operations analogously within the memory cells themselves, substituting the mechanical data transfer and digital processing sequence with a unified analog computation process that reduces overhead.
2Productivity
If separate pixel array and memory array are used, then data processing capability is improved, but occupied area and device complexity increase
Solution Approach 1:
The patent combines the pixel array and memory cell array into a single integrated structure where both functional elements coexist on the same substrate and share common readout circuits and column lines. This merging approach enables the system to maintain enhanced data processing capability through analog computing operations while significantly reducing the total occupied area compared to having separate arrays requiring independent readout paths and processing circuits.
Solution Approach 2:
The patent designs the integrated array where each element serves multiple functions: pixels perform both light detection and participate in analog computing operations, while memory cells store data and perform convolution operations. The shared readout circuits and column lines serve both pixel readout and memory cell access functions, enabling multi-functionality that reduces overall device complexity and occupied area while maintaining high data processing capability.
3Use of energy by moving object
If analog computing-in-memory is implemented, then power consumption is reduced, but manufacturing precision requirements increase
Solution Approach 1:
The patent implements self-service mechanisms where the analog computing operations are performed automatically within the memory array without requiring external precision control systems. The shared readout circuits and column lines inherently provide the necessary signal conditioning and timing synchronization, allowing the system to maintain analog signal precision through its own internal mechanisms rather than requiring external precision manufacturing and control infrastructure.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This design improves the performance-to-power ratio for data processing in image sensors by enabling analog matrix multiplication, reducing occupied area and enhancing efficiency in deep learning operations.
Implementation Method 1
Image sensors are devices for capturing images using the property of a semiconductor which reacts to light
Data Source
AI summary
Disclosed is an image sensor including a plurality of column lines, a plurality of pixels coupled to the plurality of column lines, and configured to output a plurality of pixel signals to the plurality of column lines in response to first control signals, and a plurality of memory cells coupled to the plurality of column lines, and configured to output a plurality of convolution signals, in which a plurality of data signals are reflected in the plurality of pixel signals, to the plurality of column lines in response to second control signals.