Search

CN-121983103-A - AI-driven 3D NAND flash memory read voltage self-adaptive optimization method and system

CN121983103ACN 121983103 ACN121983103 ACN 121983103ACN-121983103-A

Abstract

The invention discloses an AI-driven 3D NAND flash memory read voltage self-adaptive optimization method and system, which relate to the technical field of data storage and comprise a characteristic sensing and prediction stage, wherein the characteristic sensing and prediction stage is used for collecting multidimensional characteristic vectors of a 3D NAND flash memory storage unit, inputting a lightweight AI prediction model, and outputting a prediction voltage interval, offset direction probability and prediction confidence; and a hierarchical knowledge construction and real-time fine tuning stage, which acquires multi-source voltage data and performs dynamic weighted fusion to obtain initial voltage, performs local voltage scanning by taking the initial voltage as a starting point to finish fine tuning, and starts a progressive safe rollback mechanism if the reading reliability gain is insufficient after fine tuning. The invention realizes extremely high calibration precision of the read voltage and reduces delay and power consumption.

Inventors

  • HE JIAOYANG
  • YANG WANYUN
  • Li Yufo
  • YUAN QINGQING
  • MA YI
  • XIONG WEI

Assignees

  • 芯盛智能科技(湖南)有限公司

Dates

Publication Date
20260505
Application Date
20260409

Claims (10)

  1. 1. An AI-driven 3D NAND flash memory read voltage adaptive optimization method, comprising: in the characteristic sensing and predicting stage, collecting multidimensional characteristic vectors of a 3D NAND flash memory storage unit, inputting a lightweight AI prediction model, and outputting a predicted voltage interval, an offset direction probability and a prediction confidence coefficient, wherein the multidimensional characteristic vectors comprise aging characteristics, environmental characteristics, interference characteristics, topological characteristics, historical characteristics and process characteristics; The self-adaptive decision and optimizing stage selects a corresponding searching strategy according to the comparison result of the prediction confidence coefficient and a preset threshold value, and scans the prediction voltage area or the periphery of the reference voltage to determine the optimal reading voltage; And in the hierarchical knowledge construction and real-time trimming stage, multi-source voltage data are obtained, initial voltage is obtained through dynamic weighted fusion, local voltage scanning is performed by taking the initial voltage as a starting point to finish trimming, final reading voltage is obtained, a progressive safe rollback mechanism is started if the reading reliability gain is insufficient after trimming, and the multi-source voltage data comprise knowledge base reference voltage, spatial adjacent unit voltage average value, time sequence prediction voltage and AI prediction voltage.
  2. 2. The AI-driven 3D NAND flash memory read voltage adaptive optimization method of claim 1, wherein the multi-dimensional feature vector is obtained by stitching the aging feature, the environmental feature, the interference feature, the topology feature, the history feature, and the process feature, expressed as , wherein, The aging characteristic is the number of effective program/erase cycles that are temperature and rate compensated, The environmental characteristics comprise the mean value and variance of the current temperature and the historical temperature statistics; The interference features include read interference times, program interference intensity, and data retention time; The topological feature is the three-dimensional normalized coordinate of the storage unit in the chip; The history characteristic is a history optimal voltage sequence of the storage unit; the process features are inherent bias parameters for manufacturing lot calibration.
  3. 3. The method for adaptively optimizing read voltage of an AI-driven 3D NAND flash memory as recited in claim 1, wherein said lightweight AI prediction model is a multi-task learning neural network comprising a feature embedding layer, a CNN/GRU multi-branch feature extraction layer and an attention weight modulation layer, and is input as said multi-dimensional feature vector and output as a predicted voltage interval Probability of offset direction And predicting the confidence coefficient C, wherein the value range is 0-1.
  4. 4. The AI-driven 3D NAND flash memory read voltage adaptive optimization method as recited in claim 1, wherein the search strategy comprises the specific contents of a preset high threshold value And a low threshold The confidence level C of the output is matched with the high threshold value And a low threshold Comparing if Selecting a high confidence accuracy strategy in the predicted voltage interval Fine sweeping with small step length if Selecting a message balance strategy to predict a voltage interval For the central expansion range, adopting a variable step-length mode of dense centers and sparse edges for scanning, if Selecting a low-confidence robust strategy, backing to a knowledge base reference voltage, performing wide-range conservative scanning around the knowledge base reference voltage, and finally determining the original error rate through multi-stage scanning searching Minimum optimum voltage 。
  5. 5. The method for adaptively optimizing the read voltage of the AI-driven 3D NAND flash memory of claim 1, wherein said acquiring the multi-source voltage data comprises the steps of: analyzing the target physical address as a multi-level index key, and querying the hierarchical voltage knowledge base in parallel to obtain voltage values of all levels Confidence weight Obtaining the reference voltage of the knowledge base through the following weighted fusion calculation ; Generating a target unit peripheral address list according to a predefined proximity strategy, searching the latest voltage value corresponding to the address in the list in batches, removing abnormal values, and calculating arithmetic average on the residual voltage values to obtain a spatial proximity unit voltage average value ; Reading the most recent N times of historical optimal voltage sequences of a target unit Obtaining time sequence prediction voltage through exponential smoothing model prediction Smoothing factor Dynamically adjusted according to the fluctuation of the sequence, Is in a model state; Collecting 5-7 key features to form a simplified feature vector, inputting a special neural network or decision tree model with the parameter number of less than 1KB, and completing reasoning and outputting AI predicted voltage within 1 microsecond 。
  6. 6. The method for adaptive optimization of AI-driven 3D NAND flash memory read voltage as recited in claim 5, wherein the hierarchical voltage knowledge base is a multi-level architecture including chip level, die level, block/level, page/word line level, each level of voltage data being associated with a corresponding confidence weight.
  7. 7. The method for adaptively optimizing the read voltage of the AI-driven 3D NAND flash memory of claim 5, wherein said initial voltage is obtained by dynamic weighted fusion of said multi-source voltage data, expressed as: ; Wherein the method comprises the steps of And the dynamic weight coefficients are distributed in real time according to the long-term historical accuracy of each voltage source and the feature similarity with the current scene.
  8. 8. The method for adaptively optimizing the read voltage of the AI-driven 3D NAND flash memory of claim 1, wherein the step of starting a progressive safe rollback mechanism if the post-trimming read reliability gain is insufficient comprises the steps of: Comparing the original bit error rate RBER_new of the final read voltage V_final obtained after fine adjustment with the estimated bit error rate RBER_base of the reference voltage V_base, calculating a relative gain G= (RBER_base-RBER_new)/RBER_base, and if the relative gain G is lower than a preset threshold value or the original bit error rate RBER_new exceeds an error correction tolerance, judging that the gain is insufficient, and triggering a progressive safety rollback mechanism; the progressive secure rollback mechanism includes: A step of rescanning an enlarged window, in which a search window is enlarged by 3-5 times by taking a final read voltage V_final as a center, and rescanning is performed by using step sizes of 2-3 voltage gears to find a better voltage; a step of resetting the reference voltage of a higher level, wherein if the scanning and reading of the enlarged window resetting step fails, the step of resetting the reference voltage of the upper level is inquired and attempted to be directly read by using the statistical reference voltage of the upper level; And (3) simplifying the global scanning step, namely if the reference higher-level reference voltage rescanning step still fails, carrying out non-uniform step scanning in a full voltage range [ V_min, V_max ], setting a smaller step of 1 voltage gear for fine searching in a voltage distribution high probability area counted by historical data, setting a larger step of 4-5 voltage gears for rapid crossing searching in a distributed low probability edge area, and finding available voltage.
  9. 9. The method for adaptively optimizing the read voltage of the AI-driven 3D NAND flash memory of claim 1, further comprising a model tuning and evolution stage, specifically comprising: continuously collecting all-link data of the feature perception and prediction stage, the self-adaptive decision and optimizing stage and the hierarchical knowledge construction and real-time fine adjustment stage; fine-tuning the lightweight AI prediction model parameters in real time according to the collected data to perform online incremental learning; retraining the lightweight AI predictive model with the collected big data periodically, optimizing parameters for offline depth optimization; And according to the performance statistics result, optimizing a strategy threshold value to adaptively modify and evolve the search strategy.
  10. 10. An AI-driven 3D NAND flash memory read voltage adaptive optimization system, characterized by comprising: The characteristic sensing and predicting module is used for collecting multidimensional characteristic vectors of the 3D NAND flash memory storage unit, inputting a lightweight AI predicting model, and outputting a predicted voltage interval, an offset direction probability and a prediction confidence coefficient, wherein the multidimensional characteristic vectors comprise aging characteristics, environmental characteristics, interference characteristics, topological characteristics, historical characteristics and process characteristics; the self-adaptive decision and optimization module is used for selecting a corresponding search strategy according to a comparison result of the prediction confidence coefficient and a preset threshold value, and executing scanning in the prediction voltage area or around the reference voltage to determine the optimal reading voltage; The hierarchical knowledge construction and real-time fine adjustment module is used for acquiring multi-source voltage data and performing dynamic weighted fusion to obtain initial voltage, performing local voltage scanning to finish fine adjustment by taking the initial voltage as a starting point, and starting a progressive safe rollback mechanism if the reading reliability gain is insufficient after fine adjustment, wherein the multi-source voltage data comprises a knowledge base reference voltage, a spatial adjacent unit voltage average value, a time sequence prediction voltage and an AI prediction voltage.

Description

AI-driven 3D NAND flash memory read voltage self-adaptive optimization method and system Technical Field The invention relates to the technical field of data storage, in particular to an AI-driven 3D NAND flash memory read voltage self-adaptive optimization method and system. Background With the increasing number of stacked layers of the 3D NAND flash memory, the storage density is significantly improved, but at the same time, more serious data reliability challenges are brought. The threshold voltage (Vth) distribution of a memory cell can drift and broaden due to process fluctuations, program disturb, read disturb, decay in data retention characteristics, and cyclic wear. To ensure proper reading of data, accurate read voltages must be applied to distinguish between different memory states. The traditional read voltage management scheme mainly comprises two types, namely, a fixed and factory preset read voltage is used, but the traditional fixed voltage scheme adopts a factory preset static voltage, so that the traditional fixed voltage scheme cannot adapt to the dynamic drift of threshold voltage caused by factors such as abrasion and temperature in use of a chip, and the fundamental contradiction is mismatching of static setting and dynamic change. This necessarily causes that the read voltage gradually deviates from the optimal position, which causes that the original error rate continuously rises, and further causes a series of linkage problems such as increased error correction burden, read delay, rising power consumption and the like, and seriously damages the long-term data reliability and the system performance. And secondly, performing real-time voltage calibration (such as read voltage scanning), wherein the real-time full-scan calibration searches for an optimal value by performing full-range voltage traversal on each read operation. However, this method requires performing several tens to hundreds of test reads and voltage switches, introducing a huge time delay and additional power consumption, and the calibration process itself becomes a bottleneck for performance and energy efficiency. Particularly in multilevel units such as QLC, the problem is more acute, so that although the method can ensure the precision, the high overhead makes the method difficult to be applied to a high-performance and low-power-consumption storage system. There are some approaches in the prior art that attempt to reduce calibration overhead, such as voltage tracking based on error correction code feedback or sampling calibration of local blocks. However, these methods either have slow response speed or fail to fully utilize the characteristic correlation existing inside the 3D NAND flash memory (e.g., different layers, different blocks), and the accuracy and efficiency of calibration remain insufficient. In particular, in products such as high-density QLC (four-layer cell), the number of required read voltages is large, the voltage window is narrow, and higher requirements are put on the calibration technology. Therefore, a technical solution capable of intelligently, rapidly and precisely adaptively optimizing the read voltage is needed to minimize the read delay and the system overhead on the premise of ensuring the data reliability. Disclosure of Invention The invention aims to overcome the defects of the prior art and provide an AI-driven 3D NAND flash memory read voltage self-adaptive optimization method and system, which realize extremely high calibration precision and reduce delay and power consumption. The aim of the invention is realized by the following technical scheme: in a first aspect, an AI-driven 3D NAND flash memory read voltage adaptive optimization method includes: in the characteristic sensing and predicting stage, collecting multidimensional characteristic vectors of a 3D NAND flash memory storage unit, inputting a lightweight AI prediction model, and outputting a predicted voltage interval, an offset direction probability and a prediction confidence coefficient, wherein the multidimensional characteristic vectors comprise aging characteristics, environmental characteristics, interference characteristics, topological characteristics, historical characteristics and process characteristics; The self-adaptive decision and optimizing stage selects a corresponding searching strategy according to the comparison result of the prediction confidence coefficient and a preset threshold value, and scans the prediction voltage area or the periphery of the reference voltage to determine the optimal reading voltage; And in the hierarchical knowledge construction and real-time trimming stage, multi-source voltage data are obtained, initial voltage is obtained through dynamic weighted fusion, local voltage scanning is performed by taking the initial voltage as a starting point to finish trimming, and if the reading reliability gain is insufficient after trimming, a progressive safe rollback mechanism is started