Crossbar Array Matrix-Vector Multiplication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The Von Neumann architecture is inefficient for cognitive computing due to its separation of computing and memory units, which hampers high-speed data transfer and processing, particularly in matrix-vector multiplications that are fundamental to linear algebra operations.
Innovation Solution
A crossbar array with programmable resistive elements and access elements is used to perform matrix-vector multiplications, allowing for simultaneous programming and reading of conductance values and current measurements, enabling efficient and scalable operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If Von Neumann architecture is used with separated computing and memory units, then device complexity is reduced and ease of manufacture is improved, but data transfer speed and processing efficiency deteriorate
Solution Approach 1:
The patent merges computing and memory functions into a unified crossbar array structure where memory elements serve dual purposes as both storage and computation units. The crossbar array integrates weight storage, input vector application, and multiplication operations into a single compact structure, eliminating the need for separate data transfer pathways between distinct memory and processing units.
Solution Approach 2:
The memory elements in the crossbar array perform multiple functions simultaneously: they store weight values, receive input vectors, perform analog multiplication through conductance modulation, and generate output currents. This multi-functionality allows the same physical structure to handle both memory and computation tasks that would traditionally require separate units in Von Neumann architecture.
2Productivity
If crossbar array with programmable resistive elements is used, then matrix-vector multiplication efficiency is improved, but manufacturing precision requirements increase
Solution Approach 1:
The patent utilizes the continuous parameter space of resistive element conductance to represent weight values. By programming conductance values to specific ranges rather than requiring exact discrete values, the system tolerates manufacturing variations. The analog nature of conductance modulation allows for gradual parameter adjustments that compensate for fabrication imperfections while maintaining computational functionality.
Solution Approach 2:
The system employs iterative programming and verification processes where initially imprecise conductance values are progressively refined. Through multiple programming cycles with feedback verification, the system recovers from manufacturing imprecision by adjusting conductance values to achieve the desired weight representations, effectively discarding initial inaccurate values and recovering accurate ones through iteration.
3Adaptability or versatility
If bidirectional charge counting is implemented to support negative values, then adaptability is improved, but device complexity increases
Solution Approach 1:
The readout circuit is segmented into separate charge counting paths for positive and negative current directions. By dividing the measurement function into distinct segments that handle positive and negative values independently, the system achieves bidirectional capability while keeping each segment relatively simple. The segmentation allows parallel processing of positive and negative contributions without requiring a single complex circuit to handle all cases.
Solution Approach 2:
The patent introduces an intermediary charge integration stage that converts current directions into charge accumulations on capacitors. This intermediary mechanism mediates between the analog current outputs and the digital readout, translating directional current flow into separable charge quantities that can be independently measured. The intermediary charge storage step simplifies the final measurement process by converting complex bidirectional current signals into separate, measurable charge values.
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 facilitates fast, low-power, and scalable matrix-vector multiplications, supporting negative values and reducing power consumption compared to conventional Von Neumann methods, while maintaining high precision for algorithms that do not require exact results.
Implementation Method 1
Each junction comprises a programmable resistive element and an access element for accessing the programmable resistive element
Implementation Method 2
The readout circuit is configured to apply positive read voltages having a positive voltage sign and negative read voltages having a negative voltage sign to the plurality of row lines of the crossbar array
Data Source
AI summary
A device performs a matrix-vector multiplication of a matrix with a vector. The device includes a crossbar array having row lines, column lines and junctions arranged between the row lines and the column lines. Each junction includes a programmable resistive element and an access element for accessing the programmable resistive element. The device further includes a signal generator configured to apply programming signals to the resistive elements to program conductance values for the matrix-vector multiplication. The device further includes a readout circuit and control circuitry configured to control the signal generator and the readout circuit. The readout circuit is configured to apply read voltages having a positive voltage sign and negative read voltages having a negative voltage sign to the row lines of the crossbar array. The readout circuit is further configured to read out column currents of the plurality of column lines of the crossbar array.