Artificial Neural Network Precision Weights Using eFuse Programming
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural networks face challenges due to variable characteristics in PCM devices and chip-to-chip variations in manufacturing, limiting their performance and adoption for AI applications, particularly in low-power devices with limited computational resources.
Innovation Solution
The implementation of an artificial neural network with precision weights using a combination of precision metal resistors, eFuses, and programmable transistors, which are programmed to provide precise synaptic weights and are less susceptible to manufacturing variations and voltage drops, enabling improved energy efficiency and reduced computation and memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If PCM devices are used for artificial neural networks, then non-volatile memory and low-power operation are achieved, but variable characteristics and manufacturing variations limit performance and precision
Solution Approach 1:
The patent changes the physical parameter of weight representation from continuous resistance values in PCM devices to discrete precision weight levels achieved through eFuse-based transistor programming. This allows the system to maintain non-volatile low-power operation while achieving manufacturing-tolerant precision weights with large separations between states, resolving the contradiction between power efficiency and weight precision.
Solution Approach 2:
The patent creates a composite neural network architecture that combines eFuse-based precision weight storage with PCM devices for non-volatile memory functionality. This hybrid approach leverages the precision and programmability of eFuses while maintaining the energy efficiency and non-volatility of PCM, thereby resolving the contradiction between manufacturing precision and power consumption.
2Adaptability or versatility
If conventional neural network weights are used, then computational flexibility is maintained, but susceptibility to manufacturing variations and voltage drops reduces reliability
Solution Approach 1:
The patent implements beforehand cushioning by designing precision weights with large separations between possible weight values. This pre-built margin of error protects against manufacturing variations and voltage drops, ensuring that small perturbations cannot push the weight into an incorrect operational state, thereby improving reliability while maintaining computational flexibility through programmable eFuses.
Solution Approach 2:
The patent substitutes the continuous analog resistance mechanism with a digital-like eFuse-based selection mechanism. Instead of relying on analog resistance values that are susceptible to variations, the system uses discrete transistor switching controlled by eFuses, replacing the vulnerable mechanical/electrical continuous system with a more robust digital control approach.
3Manufacturing precision
If precision weights with large separations are implemented, then susceptibility to manufacturing variations is reduced, but device complexity increases due to additional components
Solution Approach 1:
The patent applies universality by designing the eFuse-based precision weight mechanism to serve multiple functions: weight storage, weight programming, and weight selection. The same eFuse array and transistor structure are used throughout the neural network, providing a universal building block that reduces overall system complexity despite the added precision components, as the same methodology can be applied to all neurons.
4Use of energy by moving object
If eFuse-based precision weights are used, then energy efficiency is improved, but one-time programmability limits adaptability for training applications
Solution Approach 1:
The patent applies preliminary action by programming the eFuses during manufacturing or device initialization before the device is deployed for its intended application. This preliminary programming establishes the precision weights that will be used throughout the device's operational life, allowing the device to achieve high energy efficiency during inference while the adaptability requirement is satisfied during the initial configuration phase through the one-time programmable nature of eFuses.
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 approach results in an artificial neural network that provides precise weights for each neuron, enhancing energy efficiency, reducing computational and memory requirements, and making it suitable for low-power applications across various platforms, including mobile devices.
Implementation Method 1
the first selection component coupled with the fuse may be virtually grounded... the voltage source may be further configured to activate the fuse for programming the resistor
Data Source
AI summary
Methods, systems, and devices for an artificial neural network are described. In one example, an artificial neuron in an artificial neural network may include a resistor coupled with an input line and configured to indicate a synaptic weight and a fuse coupled with the resistor. The artificial neuron may also include a selection component coupled with the fuse and configured to activate the fuse for programming the resistor, and a second selection component coupled with the resistor and an output line, the second selection component configured to select the resistor for a read operation.