Volatile Memory Read Architecture for AI Model Bandwidth

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing memory devices struggle with slow data processing speeds, particularly in applications involving large generative AI models, due to inefficient read performance and data exchange between host devices and memory systems.

Innovation Solution

A volatile memory device with separate spaces for model data and normal data, utilizing a control logic circuit and clock generator to manage read operations with different clock frequencies, enabling sequential and non-sequential data access to improve read performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a single read mode is used for all data, then the memory device structure is simple, but the read performance for different data types cannot be optimized

Engineering Contradiction:
Improveread performanceVSAvoidmemory structure
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The memory cell array is divided into a first space for storing model data and a second space for storing normal data. This segmentation allows the memory device to apply different read modes to different data types, optimizing read performance for AI workloads while maintaining separate storage regions for different access patterns.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The memory device dynamically selects between sequential read mode and non-sequential read mode based on the type of data being accessed. When model data is identified, sequential read mode is applied; when normal data is identified, non-sequential read mode is applied. This dynamic adaptation resolves the contradiction by providing optimized performance without requiring entirely separate memory structures.

Inventive Principle:
Principle #15Dynamics

2Productivity

If sequential read mode is used for model data, then read bandwidth is improved for AI applications, but power consumption increases due to higher clock frequencies

Engineering Contradiction:
Improveread bandwidthVSAvoidpower consumption
Core Design Contradiction:
ProductivityVSUse of energy by moving object

Solution Approach 1:

Different clock frequencies are applied to different read operations based on data type. A first clock frequency is used for sequential reads of model data to maximize bandwidth, while a second clock frequency is used for non-sequential reads of normal data to minimize power consumption. This local differentiation resolves the contradiction by applying high performance only where needed.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The memory device changes the clock frequency parameter dynamically based on the read mode and data type. By adjusting this critical parameter, the system achieves high read bandwidth for AI model data when required, while consuming less power during normal data operations, thus resolving the trade-off between performance and energy usage.

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If non-sequential read mode is used for normal data, then data access flexibility is improved, but read performance for AI model data deteriorates

Engineering Contradiction:
Improvedata access flexibilityVSAvoidread performance
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The memory device segments read operations into two distinct modes: sequential read mode for model data and non-sequential read mode for normal data. This segmentation ensures that each data type receives the appropriate access pattern, preventing the degradation of AI workload performance while maintaining flexibility for general-purpose data access.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The memory device dynamically switches between sequential and non-sequential read modes based on the data type being accessed. This dynamic behavior allows the system to maintain high read performance for AI model data when sequential access is required, while providing flexible random access capabilities for normal data operations.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentUS20260050545A1Volatile memory device, memory controller and memory system
Publication Date: 2026.02.19 SAMSUNG ELECTRONICS CO LTD
  • US20260050545A1 patent drawing
  • US20260050545A1 patent drawing
  • US20260050545A1 patent drawing

AI summary

A volatile memory device, memory controller, and memory system are provided. A memory device comprises a memory cell array including a first space and a second space, a control logic circuit configured to control an operation of the memory cell array, wherein the control logic circuit is configured to, in response to receiving a first read control command targeting the first space, read, from first model data stored at a first start address in the first space to second model data stored at a first end address in the first space, in sequential order by address, and the control logic circuit is configured to, in response to receiving a second read control command targeting second data stored in the second space, read the second data in non-sequential order by address.