Neural Network Inverter Circuit With Threshold Voltage Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As the complexity and size of artificial neural networks increase, so does power consumption and heat generation, affecting circuit components and requiring efficient solutions to manage variations in characteristics and temperature sensitivity.
Innovation Solution
A semiconductor device with a hierarchical artificial neural network structure, incorporating transistors and capacitors to manage signal processing and temperature effects, including a switching circuit to adjust signal output times based on potential changes, and a correction unit to handle threshold voltage variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the number of circuits included in a chip is increased to implement a larger artificial neural network, then the processing capability and complexity of the neural network is improved, but the power consumption increases
Solution Approach 1:
The patent divides the artificial neural network into multiple layers (first layer, second layer, third layer, etc.) with each layer containing multiple neurons. This segmentation allows the system to achieve high processing capability through modular architecture while managing power consumption by enabling selective activation of specific layers and neurons based on computational requirements.
2Productivity
If the number of circuits included in a chip is increased to implement a larger artificial neural network, then the processing capability is improved, but the amount of heat generated increases
Solution Approach 1:
By segmenting the neural network into multiple layers and neurons, the patent enables distributed heat generation across the chip rather than concentration in a single area. This segmentation, combined with selective activation of computational units, helps manage thermal load while maintaining high processing capability.
3Productivity
If the scale of the artificial neural network is increased with more layers and neurons, then the computational power is improved, but the amount of arithmetic operation becomes enormous
Solution Approach 1:
The patent implements weight initialization circuits that pre-calculate and store weight coefficients in memory units before neural network operations begin. This preliminary action eliminates the need for repeated arithmetic operations during inference, significantly reducing computation time while maintaining high computational power for complex neural network architectures.
4Ease of manufacture
If circuit components are formed in the fabrication process, then the device functionality is achieved, but variations in characteristics of the circuit components occur
Solution Approach 1:
The patent implements correction circuits that measure actual threshold voltages of transistors and adjust weight coefficients accordingly. This feedback mechanism compensates for manufacturing variations in circuit components, ensuring reliable and consistent neural network performance despite process variations in transistor characteristics.
5Measurement precision
If the number of circuits is increased to handle more parameters, then the neural network accuracy is improved, but the temperature sensitivity of circuit components increases
Solution Approach 1:
The patent implements temperature sensing circuits and correction mechanisms that monitor temperature variations and adjust circuit operations accordingly. This feedback approach maintains high neural network accuracy by compensating for temperature-induced variations in circuit component characteristics, enabling reliable operation even with large numbers of circuits.
Data Source
AI summary
In a semiconductor device capable of product-sum operation, variations in transistor characteristics are reduced. The semiconductor device includes a first circuit including a driver unit, a correction unit, and a holding unit, and an inverter circuit. The first circuit has a function of generating an inverted signal of a signal input to an input terminal of the first circuit and outputting the inverted signal to an output terminal of the first circuit. The driver unit includes a p-channel first transistor and an n-channel second transistor having a back gate. The correction unit has a function of correcting the threshold voltage of one or both of the first transistor and the second transistor. The holding unit has a function of holding the potential of the back gate of the second transistor. The output terminal of the first circuit is electrically connected to an input terminal of the inverter circuit. The time from the input of a signal to the input terminal of the first circuit to the output of a signal from an output terminal of the inverter circuit depends on the potential of the back gate of the second transistor.