Dynamic Memory Reconfiguration for Multi-Channel Page Interleaving

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional multi-channel memory systems utilize a fixed number of channels for data interleaving, which may not optimally accommodate varying data access patterns, leading to potential inefficiencies and increased latency.

Innovation Solution

Implementing a dynamic memory reconfiguration method that selectively allocates memory pages across configurable subsets of channels based on a mode of allocation, allowing for both unified and non-uniform memory architectures, thereby optimizing data access by partitioning memory usage according to specific data types.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a fixed number of channels are used for memory interleaving, then the memory system structure is simple and stable, but the system cannot adapt to varying data access patterns leading to increased latency and reduced efficiency

Engineering Contradiction:
Improveadaptability to varying data access patternsVSAvoidmemory reconfiguration complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements dynamic memory channel allocation where the number of channels used for interleaving is not fixed but can be adjusted based on workload characteristics. The system dynamically configures memory channels into different groups (first group and second group) and adjusts the interleaving depth and channel assignment according to real-time access patterns, transforming the static memory architecture into a dynamic one that adapts to varying computational demands.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent divides the memory channels into multiple groups (first group of channels and second group of channels) that can be independently configured and allocated to different memory pools. This segmentation allows the system to allocate specific channel groups to different processing contexts or memory regions, enabling flexible adaptation to various access patterns while maintaining manageable complexity through organized channel grouping.

Inventive Principle:
Principle #1Segmentation

2Loss of time

If memory pages are allocated across all channels, then bandwidth utilization is maximized, but latency increases due to interference between different memory pools

Engineering Contradiction:
Improvememory access latencyVSAvoiddata processing throughput
Core Design Contradiction:
Loss of timeVSProductivity

Solution Approach 1:

The patent segments memory channels into distinct groups that can be allocated to different memory pools, allowing simultaneous operations on different channels without interference. By dividing channels into first and second groups and assigning them to different memory pools based on access patterns, the system reduces cross-traffic interference and latency while maintaining overall throughput through parallel channel utilization.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different channel allocation strategies to different memory pools based on their specific access characteristics. Rather than uniformly allocating all channels to all memory pools, the system assigns specific channel groups to specific memory pools according to their workload requirements, optimizing local access performance for each pool while maintaining global productivity through coordinated multi-pool operation.

Inventive Principle:
Principle #3Local quality

3Ease of operation

If unified memory architecture is used for multiple processing units, then resource sharing is improved, but memory access conflicts and latency increase

Engineering Contradiction:
Improvememory sharing efficiencyVSAvoidmemory transfer time
Core Design Contradiction:
Ease of operationVSLoss of time

Solution Approach 1:

The patent divides memory channels into separate groups that can be allocated to different processing units or memory pools, creating a hybrid architecture that combines unified and distributed memory characteristics. This segmentation allows processing units to share access to the overall memory system while having dedicated channel groups that reduce contention and transfer times, balancing resource sharing efficiency with access speed.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4430478B1Dynamic memory reconfiguration
Publication Date: 2025.09.10 ADVANCED MICRO DEVICES INC
  • EP4430478B1 patent drawingFigure 1
  • EP4430478B1 patent drawingFigure 2
  • EP4430478B1 patent drawingFigure 3

AI summary

A processing system including a parallel processing unit selectively allocating pages of memory for interleaving across configurable subsets of channels based on a mode of allocation. In some embodiments, in a first mode, a page of memory is allocated to and interleaved across a plurality of channels, and in a second mode, a page of memory is allocated to and interleaved across a subset of the plurality of channels.