Integrated Circuit Analog Neural Network Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current technologies face inefficiencies in processing large amounts of image data from image sensors, particularly in intensive computations like image segmentation and object recognition, as they require transmission to general-purpose microprocessors, which can be resource-intensive.
Innovation Solution
An integrated circuit device with an image sensing pixel array and a memory cell array performs inference computations using hybrid bonding to directly connect the image sensor chip and memory chip to a logic wafer, enabling efficient multiplication and accumulation operations within the device.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If image data is transmitted to general-purpose microprocessors for processing, then computational flexibility is maintained, but processing efficiency and energy consumption deteriorate
Solution Approach 1:
The patent merges image sensing functionality with neural network computation capabilities in a single integrated circuit device. The image sensor array and memory cell array are integrated to perform computation directly at the pixel level, eliminating the need to transmit data to external microprocessors and thereby reducing energy consumption while improving processing efficiency.
Solution Approach 2:
The patent replaces the conventional mechanical data transmission approach (moving data from sensor to processor) with an in-situ computation approach where the image sensor array directly performs neural network computations through its integrated memory cell array, substituting the traditional data movement mechanism with a compute-embedded sensing system.
2Productivity
If data transmission to microprocessors is used for image processing, then computational capability is maintained, but data transmission requirements and system complexity increase
Solution Approach 1:
The patent combines the image sensor array with a memory cell array that can store and process data locally. This integration allows the system to perform neural network computations directly within the sensor array without requiring complex data transmission infrastructure to external microprocessors, thereby reducing system complexity while maintaining computational capability.
Solution Approach 2:
The image sensor array is designed to be self-sufficient by incorporating its own memory cell array for computation. The sensor can perform neural network inference operations independently without requiring constant data transmission to external processing units, reducing the complexity of system interconnections and data communication protocols.
3Productivity
If specialized circuits like MAC units are used for multiplication and accumulation, then computation performance is improved, but device complexity and manufacturing difficulty increase
Solution Approach 1:
The patent utilizes the threshold voltage parameter of memory cells to encode weight values for neural network computations. By programming the threshold voltages of memory cells in the array to represent different weight values, the system achieves multiplication and accumulation operations without requiring complex specialized hardware circuits, thereby improving computation performance while maintaining manufacturing simplicity.
Data Source
AI summary
A method of artificial neural network computations, including: receiving image data having pixel values; generating, from the pixel values, a column of inputs to a set of artificial neurons; identifying a region of memory cells of the integrated circuit device having threshold voltages programmed to represent a weight matrix for the set of artificial neurons; instructing voltage drivers in the integrated circuit device to apply voltages to the region of memory cells according to the column of inputs; obtaining, based on the region of memory cells responsive to the applied voltages, a first column of data from an operation of multiplication and accumulation applied on the weight matrix and the column of inputs; and applying activation functions of the set of artificial neurons to the first column of data to generate a second column of data representative of outputs of the set of artificial neuron.