Crossbar Analog Quantizer With Intrinsic ReLU for Fewer Converters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional crossbar array circuits require additional Digital-to-Analog Converters (DACs) and Analog-to-Digital Converters (ADCs) for signal transmission, leading to increased power consumption and circuit complexity, especially in large-scale applications like AI computing and neural networks.
Innovation Solution
The implementation of an analog-to-analog quantizer with an intrinsic Rectified Linear Unit (ReLU) function, which uses a DAC at the input and an ADC at the output, along with buffers and crossbars, to enable direct analog signal transmission between crossbar array circuits without additional ADCs and DACs, reducing circuit complexity and overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If additional ADCs and DACs are used for signal transmission between crossbar array circuits, then signal conversion accuracy is improved, but circuit complexity and power consumption increase
Solution Approach 1:
The patent combines the ADC and DAC functions into a single integrated device that performs both analog-to-digital and digital-to-analog conversion. This merging eliminates the need for separate ADC and DAC components, thereby reducing circuit complexity and power consumption while maintaining signal conversion accuracy through the unified conversion mechanism.
Solution Approach 2:
The invention creates a universal conversion device that can perform multiple functions (ADC and DAC) within a single circuit architecture. This multi-functional device replaces the need for dedicated ADC and DAC circuits, reducing overall system complexity while providing accurate signal conversion in both directions through its dual-mode operation capability.
2Measurement precision
If additional ADCs and DACs are used for signal transmission between crossbar array circuits, then signal conversion accuracy is improved, but power consumption increases
Solution Approach 1:
The patent combines the ADC and DAC functions into a single integrated device that performs both analog-to-digital and digital-to-analog conversion. This merging eliminates the need for separate ADC and DAC components, thereby reducing circuit complexity and power consumption while maintaining signal conversion accuracy through the unified conversion mechanism.
Solution Approach 2:
The invention creates a universal conversion device that can perform multiple functions (ADC and DAC) within a single circuit architecture. This multi-functional device replaces the need for dedicated ADC and DAC circuits, reducing overall system complexity while providing accurate signal conversion in both directions through its dual-mode operation capability.
3Adaptability or versatility
If generic activation functions are provided for different types of neural network applications, then adaptability is improved, but training cost increases
Solution Approach 1:
The patent implements different activation function characteristics at different locations or layers of the neural network through the crossbar array configuration. By locally adjusting the circuit parameters and connection weights within the crossbar structure, the system can provide specialized activation behaviors for specific neural network layers without requiring generic activation functions, thereby reducing training costs while maintaining adaptability.
Data Source
AI summary
Technologies relating to analog-to-analog quantizers with an intrinsic Rectified Linear Unit (ReLU) function designed for in-memory computing are disclosed. An apparatus, in some implementations, includes: a DAC; a first crossbar connected to the DAC; a first analog quantizer connected to the first crossbar; a buffer connected to the first analog quantizer; a second crossbar connected to the buffer; and an ADC connected to the second crossbar.