MRAM Crosspoint Inference Engine With External Field Programming
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial neural networks, particularly Deep Neural Networks (DNNs), are computationally intensive and require significant data transfer for weight movement between memory and processing units, leading to high power consumption and inefficiency in inference and training operations.
Innovation Solution
The use of magnetoresistive random access memory (MRAM) with a crosspoint structure for storing weights, allowing in-array multiplication of inputs and weights, reducing power consumption through concurrent operations and eliminating the need for large MAC arrays, with external magnetic fields assisting in programming to achieve low resistance states with high thermal stability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If traditional von Neumann architecture is used with separate memory and processing units, then data transfer between memory and processing units occurs, but power consumption increases and computational efficiency decreases
Solution Approach 1:
The patent merges memory and processing functions into a single crosspoint memory array where weight values are stored in memory cells and multiplication operations are performed in-place using Ohm's law and Kirchhoff's current law, eliminating the need for separate memory and processing units and their associated data transfers
Solution Approach 2:
The patent introduces voltage-controlled magnetic anisotropy as an intermediary mechanism that enables weight programming through external magnetic fields, allowing the memory array to be configured for neural network operations without requiring complex on-chip programming circuitry
2Measurement precision
If large MAC arrays are used for neural network operations, then computational accuracy is maintained, but device complexity and manufacturing cost increase
Solution Approach 1:
The patent replaces complex MAC array hardware with a simplified crosspoint memory array that performs multiplication through fundamental electrical laws (Ohm's law and Kirchhoff's current law), where the product of input voltage and weight conductance naturally produces the correct result without requiring additional computational hardware
Solution Approach 2:
The crosspoint memory array serves multiple functions: it stores weight values, performs multiplication operations, and accumulates results all within the same structure, eliminating the need for separate MAC arrays and reducing overall device complexity
3Reliability
If MRAM cells are programmed to low resistance states for high thermal stability, then data retention improves, but programming difficulty increases without external magnetic field assistance
Solution Approach 1:
The patent introduces external magnetic fields as an intermediary that assists in programming MRAM cells to low resistance states, reducing the programming difficulty while maintaining the thermal stability and data retention benefits of low resistance programming
Solution Approach 2:
The patent changes the programming approach by using external magnetic fields to modify the magnetic anisotropy of MRAM cells, enabling easier programming to low resistance states without compromising the thermal stability required for reliable data retention
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 significantly reduces power consumption and increases efficiency in neural network operations by performing vector multiplication within the memory array, enabling ultralow power, low-cost inference engines without requiring special on-chip programming, while maintaining high data retention and stability.
Implementation Method 1
The MRAM memory cells can be individually programmed using a combination of input voltages and an external magnetic field. The external magnetic field is chosen so that a set of programming voltages reduces the anisotropy sufficiently to align a selected memory cell's state with the external field
Implementation Method 2
Vector multiplication is performed as an in-array multiplication between a vector of input voltages, corresponding to the input vector for a layer of a neural network, and a matrix of weight values encoded by the MRAM cell states
Implementation Method 3
the computations for each of the layers are performed on the device. To perform the in-array multiplication of inputs and weights for layers of a neural network
Data Source
AI summary
An MRAM-based vector multiplication device, such as can be used for inferencing in a neural network, is presented that is ultralow power, low cost, and does not require special on-chip programming. A crosspoint array has an MRAM cell at each crosspoint junction and periphery array circuitry capable of supplying independent input voltages to each word line and reading current on each bit line. Vector multiplication is performed as an in-array multiplication of a vector of input voltages with matrix weight values encoded by the MRAM cell states. The MRAM cells can be individually programmed using a combination of input voltages and an external magnetic field. The external magnetic field is chosen so that a write voltage of one polarity reduces the anisotropy sufficiently to align the cell state with the external field, but is insufficient to align the cell if only half of the write voltage is applied.