Analogue Arithmetic Unit for Edge Neural Network Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In edge computing, implementing large-scale arithmetic units for neural networks is challenging due to high arithmetic operation loads, which result in long processing times and increased power consumption when using digital arithmetic units for activation function processing.
Innovation Solution
An analogue arithmetic unit is developed, comprising input terminals, voltage/current conversion circuits, current addition circuits, current/voltage conversion circuits, and a division circuit, which converts arithmetic operations of activation functions into analogue processes in a physical device, reducing the need for sequential digital arithmetic operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If digital arithmetic units are used for activation function processing in neural networks, then processing can be performed with digital circuits, but the arithmetic operation load increases, resulting in long processing times and increased power consumption
Solution Approach 1:
The patent replaces digital arithmetic operations with an analogue physical system. Specifically, it uses a network of neurons with resistive connections where voltage potentials naturally evolve according to Kirchhoff's laws, effectively performing activation function computations through physical analogue processes rather than sequential digital arithmetic operations
Solution Approach 2:
The patent changes the computational paradigm from digital discrete values to analogue continuous voltage potentials. By representing neural network states as voltage levels and computations as voltage evolution through resistive networks, the system achieves parallel analogue processing that eliminates the sequential bottleneck of digital arithmetic
2Reliability
If digital arithmetic units are used for activation function processing in neural networks, then processing can be performed with digital circuits, but the arithmetic operation load increases, resulting in increased power consumption
Solution Approach 1:
The patent replaces power-hungry digital arithmetic operations with a passive resistive network that computes activation functions through natural voltage distribution. The analogue neural network uses ohmic resistors and voltage sources to perform computations without active switching, dramatically reducing power consumption compared to digital arithmetic units
Solution Approach 2:
The analogue neural network performs computations autonomously through the natural physical behavior of electrical circuits. The voltage potentials automatically evolve according to circuit laws, eliminating the need for clocked operation and control logic that consume power in digital systems
3Productivity
If large-scale arithmetic units are implemented in edge terminals for neural network processing, then processing capability is improved, but device complexity and implementation difficulty increase
Solution Approach 1:
The patent replaces complex digital arithmetic logic with simple resistive connections and voltage sources. The analogue neural network achieves large-scale processing capability through parallel resistive elements rather than sequential digital logic gates, dramatically simplifying the physical implementation while maintaining computational power
Solution Approach 2:
The patent divides the neural network into modular neuronal units, each implemented as a simple circuit with resistive connections to other neurons. This segmentation allows scalable implementation where processing capability increases with the number of modular units without proportionally increasing individual unit complexity
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 solution reduces the arithmetic operation load and processing time while minimizing power consumption by performing analogue processing of activation functions in parallel, thereby enhancing the efficiency of neural network operations in edge terminals.
Implementation Method 1
a voltage/current conversion circuit configured to exponentially convert each input voltage applied to each of the plurality of input terminals and output the converted input voltage as a current
Data Source
AI summary
An analogue arithmetic unit according to an embodiment includes a plurality of input terminals, a voltage/current conversion circuit, a current addition circuit, a current/voltage conversion circuit, and a division circuit. The voltage/current conversion circuit exponentially converts each input voltage applied to each of the plurality of input terminals and outputs the converted input voltage as a current. The current addition circuit obtains a sum of the currents converted by the voltage/current conversion circuit. The current/voltage conversion circuit converts each of the currents and the sum of the currents into voltages. The division circuit calculates a ratio of the voltage obtained by converting each of the currents to a total voltage obtained by converting the sum of the currents.