Variable Resistance Memory Segmentation for AI Weight Coefficient Writing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural network computing circuits face inefficiencies in both initial setting of connection weight coefficients at product shipment and frequent updates post-shipment, as they either require precise conductance writing or conductance change-based training, leading to inaccurate or inefficient operations.
Innovation Solution
An artificial intelligence processing device utilizing two variable-resistance nonvolatile storage elements with the same structure but under different driving conditions, where one element's conductance changes discontinuously with a single voltage pulse and the other changes continuously with successive pulses, enabling accurate initial setting and efficient training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conductance is written to nonvolatile memory based on conductance itself, then initial setting accuracy is improved, but training efficiency deteriorates
Solution Approach 1:
The patent divides the storage system into two separate nonvolatile memories: one dedicated to storing initial conductance values with high precision, and another dedicated to storing conductance changes during training. This segmentation allows each memory to be optimized for its specific function, resolving the contradiction between initial setting accuracy and training efficiency.
Solution Approach 2:
The patent performs preliminary writing of conductance values to the first nonvolatile memory before training begins. This preliminary action stores the baseline conductance values that ensure accurate initial setting, while subsequent training operations only need to write conductance changes to the second memory, improving training efficiency.
2Productivity
If conductance change-based writing is used, then training efficiency is improved, but initial setting accuracy deteriorates
Solution Approach 1:
The patent separates the storage of absolute conductance values from the storage of conductance changes by using two different nonvolatile memories. The first memory stores precise initial conductance values, while the second memory stores changes during training, allowing each to use the most efficient writing method for its purpose.
Solution Approach 2:
The patent changes the writing parameter from absolute conductance values to conductance changes (differences) for the training phase. This parameter change allows faster, more efficient training updates without compromising the accuracy of the initial setting, which is stored separately with high precision.
3Speed
If single voltage pulse writing is used, then writing speed is improved, but manufacturing precision deteriorates
Solution Approach 1:
The patent assigns different writing methods to different storage scenarios by segmenting the system into two memories. The first memory uses single voltage pulse writing for fast initial setting, while the second memory uses multi-pulse writing for precise conductance change storage during training, balancing speed and precision for each function.
4Manufacturing precision
If multi-pulse voltage writing is used, then manufacturing precision is improved, but writing speed deteriorates
Solution Approach 1:
The patent performs the time-consuming multi-pulse writing operations only once during the preliminary initial setting phase to establish accurate baseline conductance values. After this preliminary action, subsequent training operations use faster single-pulse writing for conductance changes, minimizing the impact of slow writing speed on overall training efficiency.
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
The device achieves high-accuracy initial setting and high-efficiency training of connection weight coefficients, enhancing the operational efficiency of AI processing devices by simplifying manufacturing and reducing chip size.
Implementation Method 1
variable-resistance nonvolatile storage elements each having a resistance that varies according to a given electrical signal
Data Source
AI summary
An artificial intelligence processing device includes: a substrate; an operation circuit such as a multiply-accumulate operation circuit, which includes a first variable-resistance nonvolatile storage element and a second variable-resistance nonvolatile storage element that are provided on the substrate and have a same structure and each of which holds conductance; and a write circuit that rewrites the conductance of the first variable-resistance nonvolatile storage element by applying a first voltage pulse having a first voltage to the first variable-resistance nonvolatile storage element, and rewrites the conductance of the second variable-resistance nonvolatile storage element by applying a second voltage pulse having a second voltage to the second variable-resistance nonvolatile storage element, the second voltage being different from the first voltage.