Stacked Crossbar Arrays for Neural Network Weight Storage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural network accelerators face limitations in memory bandwidth, energy consumption, and chip size, which restrict the scalability and efficiency of neural network computations, especially in mobile devices.
Innovation Solution
The proposed solution involves a stacked crossbar array configuration with memristor or memcapacitor crosspoint devices, which enables high-density storage of neural network weights on a single chip, reducing energy consumption and eliminating the need for off-chip parameter retrieval, allowing for larger and more complex neural network models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional neural network accelerators are used, then neural network computations can be performed, but memory bandwidth is limited and energy consumption is high
Solution Approach 1:
The patent merges the weight storage function and the computation function into a single integrated structure. The crossbar array simultaneously stores weight values in its conductance states and performs matrix multiplication computations, eliminating the separation between memory and processing units that causes the memory wall problem and associated energy consumption.
Solution Approach 2:
The patent introduces an analog conductance state as an intermediary representation for weight values. By encoding weights in the continuous conductance states of crossbar devices rather than digital memory cells, the system enables direct analog multiplication with input signals, achieving high-speed computation with minimal energy transfer between storage and processing components.
2Productivity
If conventional neural network accelerators are used, then computations can be performed, but chip size increases due to timesharing requirements
Solution Approach 1:
The patent transitions from a planar two-dimensional layout to a three-dimensional stacked architecture. Multiple crossbar arrays are vertically stacked and interconnected through through-silicon vias, enabling the system to accommodate millions or billions of operators in a compact footprint by utilizing the vertical dimension for inter-layer connectivity.
Solution Approach 2:
The patent implements a hierarchical nested structure where multiple crossbar arrays are stacked and interconnected. Each crossbar array layer is nested within the overall 3D stack, with lower layers providing weight storage and upper layers performing computations, creating a compact nested architecture that maximizes computational density.
3Quantity of substance
If weights are stored in crosspoint devices, then high-density storage is achieved, but precision may be limited by device variability
Solution Approach 1:
The patent uses 1-bit digital input signals that are applied excessively to multiple crossbar arrays simultaneously. By using simple digital inputs rather than precise analog signals, the system compensates for device variability through the sheer number of parallel operations and statistical convergence, achieving accurate results despite individual device imprecision.
Solution Approach 2:
The patent implements calibration and compensation mechanisms that use feedback from measured device characteristics to adjust weight values. By characterizing the actual conductance states of crossbar devices and applying correction factors during the weight programming phase, the system compensates for manufacturing variations and achieves the required precision for neural network computations.
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 configuration achieves an order of magnitude lower energy consumption and increased performance by reducing timesharing requirements, enabling larger neural networks with millions or billions of operators on a single chip, suitable for mobile devices with limited power and space.
Implementation Method 1
Each crossbar array includes a number of electronic operators that together define an output of the layer as a function of one or more inputs
Implementation Method 2
The crosspoint devices include a memristor device, and the electrical property that is tuned to the value is a conductance of the memristor device
Implementation Method 3
The crosspoint devices include a memcapacitor device, and the electrical property that is tuned to the value is a capacitance of the memcapacitor device
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A circuit for performing neural network computations for a neural network is described. The circuit includes plurality of neural network layers each including a crossbar arrays. The plurality of crossbar arrays are formed in a common substrate in a stacked configuration. Each crossbar array includes a set of crosspoint devices. A respective electrical property of each of the crosspoint devices is adjustable to represent a weight value that is stored for each respective crosspoint device. A processing unit is configured to adjust the respective electrical properties of each of the crosspoint devices by pre-loading each of the crosspoint devices with a tuning signal. A value of the turning signal for each crosspoint device is a function of the weight value represented by each respective crosspoint device.