Resistive Processing Unit Weight Extraction via Analog Matrix-Vector Multiplication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing analog resistive processing systems for neuromorphic computing face challenges in accurately extracting weight values from resistive memory arrays due to hardware non-idealities such as offsets and noise.
Innovation Solution
The system employs a processor coupled with a resistive processing unit (RPU) that includes an array of cells with programmable resistive devices. A weight extraction process is performed by applying input vectors for analog matrix-vector multiplication, obtaining output vectors, and determining weight values using these vectors, while compensating for hardware non-idealities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If weight values are extracted directly from resistive memory array, then extraction process is simple, but measurement precision is poor due to hardware non-idealities
Solution Approach 1:
The patent introduces an intermediary weight extraction process that uses input vectors and output vectors as mediators between the resistive memory array and the final weight values. Instead of directly reading weights from the array, the system applies input vectors to the array, measures output vectors, and computationally derives weight values from these measurements, thereby achieving higher precision while managing complexity through structured intermediate steps
Solution Approach 2:
The patent implements feedback by using measured output vectors to computationally determine accurate weight values. The system continuously refines weight extraction by comparing expected outputs with actual measurements and adjusting the extracted weight values accordingly, creating a closed-loop feedback mechanism that improves measurement precision despite hardware non-idealities
2Productivity
If analog matrix-vector multiplication is performed, then computational speed is improved, but measurement precision deteriorates due to noise and offsets
Solution Approach 1:
The patent uses feedback to compensate for analog computation errors. By measuring actual output vectors from the analog matrix-vector multiplication and using these measurements to computationally determine weight values, the system creates a feedback loop that corrects for noise and offset errors accumulated during high-speed analog processing
Solution Approach 2:
The patent introduces output vectors as intermediaries between the analog computation process and the final weight values. These output vectors serve as measurable quantities that bridge the gap between fast analog computation and precise digital weight extraction, allowing the system to maintain high computational speed while achieving accurate weight determination through subsequent computational processing
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 enables accurate extraction of weight values from RPU arrays, effectively addressing hardware non-idealities and ensuring precise numerical computations in neuromorphic computing systems.
Implementation Method 1
The cells respectively comprise resistive devices, wherein at least a portion of the resistive devices are programmable to store weight values of a given matrix in the array of cells
Implementation Method 2
applying a set of input vectors to the resistive processing unit to perform analog matrix-vector multiplication operations on the stored matrix, obtaining a set of output vectors resulting from the analog matrix-vector multiplication operations
Data Source
AI summary
A system includes a processor, and a resistive processing resistive processing unit coupled to the processor. The resistive processing unit includes an array of cells, wherein the cells respectively include resistive devices, wherein at least a portion of the resistive devices are programmable to store weight values of a given matrix in the array of cells. When the given matrix is stored in the array of cells, the processor is configured to perform a weight extraction process. The weight extraction process applies a set of input vectors to the resistive processing unit to perform analog matrix-vector multiplication operations on the stored matrix, obtains a set of output vectors resulting from the analog matrix-vector multiplication operations, and determines weight values of the given matrix stored in the array of cells utilizing the set of input vectors and the set of output vectors.