Quantized Kernel Convolution Circuit for Low-Energy CNN Processing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Convolutional neural network processing hardware faces challenges with increased hardware area and energy consumption due to the need for massive multiply-and-accumulate operations and repeated use of data, particularly from higher-level memory, leading to inefficient energy usage.

Innovation Solution

An electronic apparatus that minimizes multiplication operations by using quantized kernel data stored in memory, employing operation circuits to select and accumulate output values based on identification codes, and performing convolution operations through selection and multiplication operations, reducing hardware area and energy consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional convolution operations are performed using massive multiply-and-accumulate operations, then processing accuracy is maintained, but hardware area and energy consumption increase significantly

Engineering Contradiction:

Improveprocessing accuracyVSAvoidenergy consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent transforms the convolution operation by changing the parameter representation of kernel data from continuous weight values to discrete identification codes (0, 1, 2). This parameter transformation enables the replacement of multiplication operations with selection operations, significantly reducing energy consumption while preserving processing accuracy through the mathematical equivalence of the transformed operation.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent substitutes the mechanical multiplication operation with a selection operation based on identification codes. Instead of performing computationally intensive multiply-and-accumulate operations, the system uses lookup tables and selection circuits to achieve the same mathematical result with much lower energy consumption and hardware requirements.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Ease of operation

If data is stored in higher-level memory (DRAM) for repeated use, then data accessibility is improved, but energy consumption increases tens to hundreds of times compared to low-level memory (SRAM)

Engineering Contradiction:

Improvedata accessibilityVSAvoidenergy consumption

Core Design Contradiction:

Ease of operationVSUse of energy by moving object

Solution Approach 1:

The patent applies local quality by differentiating the storage requirements for different types of data. Kernel data with repeated access patterns is stored in low-level memory (SRAM) for energy efficiency, while other data can use higher-level memory. This localized optimization of memory hierarchy based on access patterns reduces overall energy consumption.

Inventive Principle:

Principle #3Local quality

3Productivity

If kernel data is stored and repeatedly accessed from memory, then convolution operations can be performed, but hardware area increases due to connection structures for massive MAC operations

Engineering Contradiction:

Improveconvolution processing capabilityVSAvoidhardware area

Core Design Contradiction:

ProductivityVSArea of stationary object

Solution Approach 1:

The patent replaces the complex mechanical connection structure required for massive multiply-and-accumulate operations with a simpler selection-based architecture. By using identification codes to index into lookup tables and selection circuits, the system achieves the same convolution processing capability with significantly reduced hardware area and simpler interconnections.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

4Measurement precision

If one kernel weight is used OW×OH times and one input pixel is used FW×FH×OD times, then convolution accuracy is maintained, but energy consumption increases due to repeated memory access

Engineering Contradiction:

Improveconvolution accuracyVSAvoidenergy consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies preliminary action by pre-processing kernel data into identification codes and storing them in lookup tables before the convolution operation. This preprocessing enables the system to perform repeated selections from pre-computed tables rather than repeatedly accessing raw kernel data from memory, significantly reducing energy consumption while maintaining convolution accuracy.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12530169B2Electronic device and control method therefor using quantized kernel data

Publication Date: 2026.01.20 SAMSUNG ELECTRONICS CO LTD

AI summary

An electronic device and a method of controlling thereof are provided. The electronic device includes a memory configured to store coefficient data and identification code data in which kernel data is quantized; a first operation circuit configured to, based on a plurality of target elements of target data being sequentially input, select an output value corresponding to at least one of the plurality of target elements that is sequentially input according to an identification code corresponding to the one of the plurality of target elements, and accumulate the selected output value; and a second operation circuit configured to output a convolution operation result based on output data that is output from the first operation circuit and a coefficient corresponding to the output data.