NVM Neuromorphic Circuit Refresh via Randomized Synapse Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In Non-Volatile Memory (NVM)-based neuromorphic circuits using 2-Phase Change Memories, once conductance reaches its maximum, no further valid update is possible, limiting accuracy, and existing refresh operations consume excessive power and require significant circuitry.
Innovation Solution
A method and circuit design for refreshing cells in NVM-based neuromorphic circuits that involves randomly selecting neurons, reading conductance, resetting and reconfiguring Gp and Gm cells to recover effective weights, using a weight refresh circuit with comparator circuits to manage conductance, reducing the number of refresh operations and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional refresh operations are performed to reset conductance when maximum is reached, then learning accuracy is maintained, but power consumption increases and circuit complexity increases
Solution Approach 1:
The patent applies partial action by performing refresh operations only on a subset of synapses rather than all synapses. Specifically, refresh is triggered only when conductance reaches maximum thresholds, and only for selected synapses based on random neuron selection and conductance comparison, rather than systematically refreshing all synapses in the array.
Solution Approach 2:
The patent changes the parameter of refresh operation frequency from a fixed systematic schedule to a dynamic threshold-based trigger. The refresh operation is activated only when conductance Gp or Gm reaches its maximum value, transforming the refresh parameter from time-based to state-based control.
2Measurement precision
If conventional refresh operations are performed to reset conductance when maximum is reached, then learning accuracy is maintained, but device complexity increases
Solution Approach 1:
The patent extracts the refresh control logic from a centralized systematic approach and distributes it to individual synapse-level comparators. Each synapse has its own comparator circuit that independently monitors conductance and triggers refresh only when needed, removing the need for complex centralized control circuitry.
Solution Approach 2:
The synapse refresh system is designed to be self-service through autonomous comparator circuits at each synapse that automatically detect when conductance reaches maximum and trigger refresh operations without external control. The system serves itself by using local conductance information to make refresh decisions independently.
3Reliability
If systematic refresh of all synapses is performed, then conductance saturation is prevented, but the number of refresh operations increases excessively
Solution Approach 1:
The patent applies partial action by performing refresh operations only on a subset of synapses rather than all synapses. Specifically, refresh is triggered only when conductance Gp or Gm reaches maximum thresholds, and only for selected synapses based on random neuron selection and conductance comparison, rather than systematically refreshing all synapses in the array.
Solution Approach 2:
The patent implements feedback through comparator circuits that continuously monitor conductance levels and provide feedback signals when maximum conductance is reached. This feedback mechanism enables the system to respond dynamically to actual conductance states, triggering refresh operations only when and where needed based on real-time conductance information.
Data Source
AI summary
A computer-implemented method is provided for refreshing cells in a Non-Volatile Memory (NVM)-based neuromorphic circuit wherein synapses are each composed of a respective cell pair formed from a respective Gp cell and a respective Gm cell of the cells. The method includes randomly selecting multiple neurons and reading a conductance of any of the synapses connected to the multiple neurons. The method further includes selecting any of the synapses connected to the selected multiple neurons for which the Gm conductance has reached a maximum conductance. The method also includes resetting the Gp cell and Gm cell of the selected synapses, and setting, at most, one of the Gp cell and Gm cell of each of the selected synapses to recover an effective total weight of each of the selected synapses.