Memory Controller Segmentation for AI Training Data Prioritization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current AI model training processes consume excessive computing and memory resources due to the use of all data sets, even when certain data sets have less impact, making it difficult to identify and prioritize higher priority data sets.

Innovation Solution

Implementing a memory controller in a pooled memory subsystem that associates specific data sets with AI models, allowing higher priority data sets to be used for training while omitting lower priority ones, thereby optimizing resource usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If all data sets are used for AI model training, then the model can potentially achieve better accuracy, but computing and memory resources are excessively consumed

Engineering Contradiction:
Improvemodel accuracyVSAvoidcomputing resource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent segments the training data into different priority levels (first priority, second priority, etc.) based on their impact on model training effectiveness. This segmentation allows the system to selectively process only the most important data segments, thereby reducing overall computing resource consumption while maintaining model accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by assigning different priority levels to different data sets based on their specific characteristics and impact on model training. High-priority data sets that significantly contribute to model accuracy are processed with higher resource allocation, while low-priority data sets are processed with reduced resources or omitted entirely.

Inventive Principle:
Principle #3Local quality

2Ease of operation

If all data sets are stored in memory for training, then data accessibility is improved, but memory consumption increases significantly

Engineering Contradiction:
Improvedata accessibilityVSAvoidmemory consumption
Core Design Contradiction:
Ease of operationVSQuantity of substance

Solution Approach 1:

The patent segments training data into priority-based groups and stores only the necessary portions in memory at any given time. High-priority data sets are loaded into memory for active training, while lower-priority data sets are stored externally and loaded only when needed, thereby reducing overall memory consumption while maintaining data accessibility.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements dynamic memory allocation where the amount of memory dedicated to training data adjusts based on the current training phase and priority requirements. Memory resources are dynamically allocated to load high-priority data sets during critical training stages and freed when no longer needed, optimizing the balance between accessibility and consumption.

Inventive Principle:
Principle #15Dynamics

3Productivity

If high-priority data sets are prioritized for training, then training efficiency is improved, but it becomes difficult to identify which data sets have higher priority

Engineering Contradiction:
Improvetraining efficiencyVSAvoiddata set priority identification
Core Design Contradiction:
ProductivityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent performs preliminary analysis and classification of data sets before the actual model training begins. During this preliminary phase, the system evaluates each data set's potential impact on model training and assigns priority levels in advance. This preliminary action eliminates the difficulty of identifying priority during training, as the classification is already established.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements a feedback mechanism where the system monitors the actual impact of different data sets on model training performance and adjusts priority assignments accordingly. This feedback loop helps accurately identify which data sets truly have higher priority based on their actual contribution to model improvement, resolving the difficulty of initial priority identification.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3506107B1Storing and retrieving training data for models in a data center
Publication Date: 2024.07.17 INTEL CORP
  • EP3506107B1 patent drawingFigure 1
  • EP3506107B1 patent drawingFigure 2
  • EP3506107B1 patent drawingFigure 3

AI summary

Technology for a memory controller is described. The memory controller can receive a request to store training data. The request can include a model identifier (ID) that identifies a model that is associated with the training data. The memory controller can send a write request to store the training data associated with the model ID in a memory region in a pooled memory that is allocated for the model ID. The training data that is stored in the memory region in the pooled memory can be addressable based on the model ID.