Memory Training Method Using Machine Learning Eye Width Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated test equipment (ATE) for memory devices, especially DDR5, faces challenges in synchronizing data signals and data strobe signals, leading to increased training time when hardware blocks are not optimized, necessitating software-based compensation to find fail-to-pass points for each device.
Innovation Solution
A training method that calculates eye widths for signal synchronization at different operation speeds, performs machine learning to derive a model showing the relation between operation speed and eye width, and adjusts signal delays to determine valid window margins, reducing training time through the use of a test apparatus with a timing generator and controller.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If software-based compensation is used to find fail-to-pass points for each memory device, then synchronization accuracy is improved, but training time significantly increases
Solution Approach 1:
The patent applies preliminary action by pre-calculating eye widths at multiple operation speeds (first, second, and third operation speeds) and storing them in a lookup table before actual training operations. This pre-computation allows the system to quickly retrieve pre-determined synchronization parameters during training, eliminating the need for time-consuming software-based fail-to-pass point searches while maintaining synchronization accuracy.
Solution Approach 2:
The patent derives a machine learning model that predicts eye width based on operation speed in advance. This model is trained offline using measurement data from multiple operation speeds, enabling the system to instantly predict appropriate synchronization parameters for any given operation speed during training operations, significantly reducing training time while preserving synchronization precision.
2Productivity
If hardware blocks optimized for training operation are adopted, then training speed is improved, but device complexity increases
Solution Approach 1:
The patent implements a universal approach by using a single controller that performs multiple functions: it measures eye widths, derives machine learning models, stores lookup tables, and executes training operations. This multi-functional controller eliminates the need for dedicated hardware blocks optimized solely for training, achieving high training speed through software intelligence rather than specialized hardware, thereby avoiding increased device complexity.
Solution Approach 2:
The patent replaces potential hardware-based training optimization with a software-based machine learning approach. Instead of implementing complex hardware blocks for training optimization, the system uses algorithms and lookup tables stored in memory, substituting mechanical/hardware complexity with software intelligence that achieves the same training speed improvement without increasing device complexity.
Data Source
AI summary
Provided is a training method capable of reducing or minimizing a training time. The training method includes, for each of devices to be tested, calculating a first eye width at which a first signal and a second signal synchronize with each other at a first operation speed, and calculating a second eye width at which the first signal and the second signal synchronize with each other at a second operation speed different from the first operation speed; performing machine learning on the first eye width and the second eye width to derive a model showing a relation between operation speeds and eye widths; and calculating a third eye width corresponding to a third operation speed different from the first operation speed and the second operation speed, using the model.