MTJ Stochastic Memory for SNN Weight Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing stochastic memory circuits for Spiking Neural Networks (SNNs) rely on standard CMOS hardware, which results in large area and power overhead due to the need for stochasticity in neurons' membrane potentials and input synapse weights.
Innovation Solution
The implementation of a stochastic memory array using magnetic tunnel junctions (MTJs) with switching devices that produce a switching output based on stochastic switching probabilities, allowing for efficient updating of weight values in SNNs, where bit cells include MTJs coupled with NMOS transistors and controlled by bit and source lines to switch between resistance states, enabling stochastic switching probabilities for weight updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If standard CMOS hardware is used to enable stochasticity in SNNs, then stochastic learning can be achieved, but area overhead increases significantly
Solution Approach 1:
The patent replaces standard CMOS hardware with magnetic tunnel junction (MTJ) devices to generate stochastic switching behavior. The MTJ's inherent magnetic switching stochasticity naturally provides the required randomness for SNN learning without needing complex CMOS circuitry, thereby reducing area overhead while maintaining stochastic learning capability
Solution Approach 2:
The patent changes the physical state parameters of the memory device by using MTJ resistance states (high and low resistance) to represent stochastic outcomes. By controlling the switching probability through current pulse parameters, the system achieves stochastic behavior with simpler hardware, reducing the area required compared to CMOS-based stochastic generators
2Reliability
If standard CMOS hardware is used to enable stochasticity in SNNs, then stochastic learning can be achieved, but power overhead increases significantly
Solution Approach 1:
The patent substitutes CMOS-based stochastic generation with MTJ-based stochastic switching. The MTJ device generates stochastic outcomes through its magnetic switching characteristics, which consume less power than equivalent CMOS circuitry, thereby reducing power overhead while maintaining stochastic learning functionality
Solution Approach 2:
The MTJ device inherently provides stochastic switching behavior through its magnetic properties without requiring external power-intensive control circuits. The device self-generates the stochasticity needed for SNN learning, eliminating the need for additional power-consuming CMOS components dedicated to stochastic generation
3Area of stationary object
If MTJ-based stochastic memory is used, then area overhead is reduced, but device complexity increases
Solution Approach 1:
The patent makes the MTJ device perform multiple functions: it serves as both the memory storage element and the stochastic number generator for SNN learning. This multi-functionality eliminates the need for separate CMOS circuits for stochastic generation, reducing overall area while the added complexity is confined to the MTJ device itself rather than the entire memory system
4Use of energy by stationary object
If MTJ-based stochastic memory is used, then power overhead is reduced, but manufacturing precision requirements increase
Solution Approach 1:
The patent leverages the inherent stochasticity of the MTJ magnetic switching process, which is sensitive to manufacturing variations. Rather than requiring tight precision control, the design accepts and utilizes these variations to generate the desired stochastic behavior, transforming a potential disadvantage into a functional advantage for SNN learning
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 reduces the area and power overhead in SNNs by enabling efficient stochastic learning and weight updates, improving learning accuracy and classification performance in SNNs and Synaptic Sampling Machines.
Implementation Method 1
individual bit cells of the plurality of bit cells correspond to a respective weight value and include a switch device that has a magnetic tunnel junction (MTJ)
Data Source
AI summary
Embodiments include apparatuses, systems, and methods including a memory apparatus including a plurality of bit cells, wherein each of the plurality of bit cells correspond to a respective weight value and include a switch device that has a magnetic tunnel junction (MTJ) or other suitable resistive memory element to produce stochastic switching. In embodiments, the switch device may produce a switching output according to a stochastic switching probability of the switch device. In embodiments, a bit line or a source line passes a current across the MTJ for a switching time associated with the stochastic switching probability to produce the switching output which enables a determination of whether the respective weight value is to be updated. Other embodiments may also be described and claimed.