Pop Count Neural Network Multiply Accumulator

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Deep learning neural networks with low bit-precision suffer from inefficiencies in power consumption and processing speed due to repetitive multiplication operations with the same result values, leading to high energy and area costs in hardware.

Innovation Solution

A computation method that generates one-hot encoding codes for multiplication results, performs pop-counts, and accumulates results using constant multipliers or shifters, reducing repetitive operations by classifying and processing multiplications based on repetition probability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If low bit-precision computation is used, then power consumption and hardware area are reduced, but processing speed deteriorates due to repetitive multiplication operations

Engineering Contradiction:

Improvepower consumptionVSAvoidprocessing speed

Core Design Contradiction:

Use of energy by moving objectVSProductivity

Solution Approach 1:

The patent performs preliminary classification of multiplication operations by result value types before execution. By pre-grouping operations with identical results and generating one-hot encoding codes in advance, the system avoids redundant computation during the actual processing phase, thereby improving processing speed while maintaining low bit-precision efficiency

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent merges multiple identical multiplication operations into a single operation by grouping them according to their result values. Operations with the same result are combined and executed once with a count multiplier, reducing the total number of operations and improving processing speed without increasing power consumption

Inventive Principle:

Principle #5Merging (Combining)

2Area of stationary object

If low bit-precision computation is used, then hardware area is reduced, but processing speed deteriorates due to repetitive multiplication operations

Engineering Contradiction:

Improvehardware areaVSAvoidprocessing speed

Core Design Contradiction:

Area of stationary objectVSProductivity

Solution Approach 1:

The system performs preliminary classification and grouping of multiplication operations before execution. By pre-organizing operations into groups based on result values and generating one-hot encoding codes in advance, the hardware avoids redundant computation paths during processing, improving speed without requiring additional hardware area

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent segments the computation process into distinct phases: classification phase (grouping operations by result value), encoding phase (generating one-hot codes), and execution phase (performing consolidated multiplications). This segmentation allows efficient resource utilization and improves processing speed within the same hardware area

Inventive Principle:

Principle #1Segmentation

3Measurement precision

If conventional repetitive multiplication operations are performed, then computation accuracy is maintained, but power consumption increases due to lack of optimization for repeated result values

Engineering Contradiction:

Improvecomputation accuracyVSAvoidpower consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent changes the computational parameters by introducing result value type classification and one-hot encoding. Instead of performing each multiplication independently, the system transforms the computation into a parameterized form where identical result operations share the same computational path, reducing power consumption while preserving accuracy through proper accumulation

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent merges operations that produce identical results into a single consolidated operation. By grouping multiplications with the same result value and executing them once with an accumulated count, the system reduces redundant computational activity and power consumption while maintaining the same computational accuracy through proper summation

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentUS12260188B2Pop count-based deep learning neural network computation method, multiply accumulator and device

Publication Date: 2025.03.25 KOREA ELECTRONICS TECH INST

AI summary

The present invention relates to a pop count-based deep learning neural network computation method, a multiply accumulator, and a device thereof. The computation method according to an exemplary embodiment of the present invention is a computation method for a deep learning neural network, including a step of generating one-hot encoding codes according to the type of first multiplication result values for a multiplication (first multiplication) of weights (W) and input values (A); a step of performing a pop-count for each generated code; and a step of accumulating result values for a constant multiplication (second multiplication) between each type of the first multiplication result value and each count value of the pop-count which are different constant values.