Dynamic Neural Network Quantization for Memory Bandwidth

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

If the processor of an electronic device fails to identify the available bandwidth of a memory in real time, it may not accurately calculate the memory bandwidth required to process a neural network model, leading to unsatisfactory quality of service in the processing result.

Innovation Solution

The processor identifies the available bandwidth of the memory in real time and dynamically applies a bit depth according to the memory's available bandwidth during the quantization of the neural network model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If the processor uses a fixed bit depth for quantizing the neural network model, then the quantization process is simple and fast, but the processing accuracy and quality of service cannot be satisfied when memory bandwidth is insufficient

Engineering Contradiction:
Improveprocessing accuracyVSAvoidquantization complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent implements dynamic bit depth adjustment during quantization based on real-time memory bandwidth availability. The processor monitors memory bandwidth and dynamically changes the bit depth parameter to optimize between processing accuracy and resource constraints, transforming a static quantization process into an adaptive one that responds to runtime conditions.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the bit depth parameter dynamically based on memory bandwidth conditions. When memory bandwidth is sufficient, higher bit depths are used for better accuracy; when bandwidth is limited, lower bit depths are applied to reduce memory traffic. This parameter adaptation resolves the contradiction between accuracy and system constraints.

Inventive Principle:
Principle #35Parameter changes

2Manufacturing precision

If the processor increases the bit depth to improve processing accuracy, then the quality of service is satisfied, but the memory bandwidth consumption increases

Engineering Contradiction:
Improveprocessing accuracyVSAvoidmemory bandwidth consumption
Core Design Contradiction:
Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The system dynamically adjusts bit depth based on real-time memory bandwidth monitoring. When bandwidth is abundant, higher bit depths are used to maximize accuracy; when bandwidth is constrained, the system reduces bit depth to prevent bandwidth exhaustion, thereby adaptively balancing accuracy requirements with available resources.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The bit depth parameter is changed dynamically in response to memory bandwidth conditions. The processor monitors bandwidth availability and adjusts the quantization precision parameter accordingly, allowing the system to operate at optimal accuracy levels without exceeding memory bandwidth capacity.

Inventive Principle:
Principle #35Parameter changes

3Quantity of substance

If the processor decreases the bit depth to reduce memory bandwidth consumption, then the memory bandwidth is sufficient, but the processing accuracy and quality of service deteriorate

Engineering Contradiction:
Improvememory bandwidthVSAvoidprocessing accuracy
Core Design Contradiction:
Quantity of substanceVSManufacturing precision

Solution Approach 1:

The system employs dynamic bit depth adjustment to prevent premature reduction of precision. By monitoring memory bandwidth in real-time, the processor maintains higher bit depths when bandwidth is available, ensuring accuracy is not compromised unless absolutely necessary due to resource constraints.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The quantization parameter (bit depth) is adaptively changed based on memory bandwidth conditions. The system avoids fixed low-bit-depth quantization and instead modulates the precision parameter to match available resources, preventing accuracy deterioration when bandwidth is sufficient.

Inventive Principle:
Principle #35Parameter changes

4Manufacturing precision

If the processor performs real-time memory bandwidth identification and dynamic bit depth adjustment, then the quality of service is satisfied, but the processing complexity and computation overhead increase

Engineering Contradiction:
Improvequality of serviceVSAvoidprocessing complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The processor performs self-monitoring of memory bandwidth and self-adjusts the quantization bit depth without external intervention. This autonomous adaptation allows the system to maintain optimal performance by automatically responding to resource conditions, reducing the need for complex external control mechanisms.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system implements a feedback loop where the processor monitors memory bandwidth usage and uses this information to adjust bit depth in subsequent quantization operations. This closed-loop control enables the system to adapt to changing conditions and maintain quality of service while managing resource consumption effectively.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4296904B1Electronic apparatus for processing neural network model and operating method therefor
Publication Date: 2025.09.17 SAMSUNG ELECTRONICS CO LTD
  • EP4296904B1 patent drawingFigure 1
  • EP4296904B1 patent drawingFigure 2
  • EP4296904B1 patent drawingFigure 3

AI summary

An electronic apparatus is provided. The electronic apparatus comprises a memory, and a processor including a resource management unit and a neural processing unit, wherein the processor may be configured to: obtain a request to execute a specific function operating on the basis of a specific neural network model; identify, by using the resource management unit, an available bandwidth of the memory; and quantize, by using the neural processing unit, the specific neural network model on the basis of the available bandwidth of the memory.