Matrix Multiplication Device Using Element Classification for Power Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Matrix multiplication operations in electronic circuits, such as neural network models, consume excessive power due to repeated multiplication operations on similar element values.
Innovation Solution
A matrix multiplication device that classifies element values into categories and adjusts operation modes between low power and normal modes based on classification thresholds, reducing the number of multiplication operations by transmitting product values through a routing circuit to corresponding calculation circuits.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If traditional matrix multiplication operations are performed on all element values, then calculation completeness is maintained, but power consumption increases due to repeated multiplication operations on similar values
Solution Approach 1:
The patent merges multiple calculation circuits into groups where circuits within the same group share common multiplication resources. When multiple calculation circuits need to multiply by the same second matrix element, they are grouped together and share a single multiplication circuit, reducing redundant operations and power consumption while maintaining complete calculation coverage.
Solution Approach 2:
The patent dynamically adjusts the operation mode of calculation circuits based on the distribution characteristics of first matrix elements. The control circuit determines whether to operate in low power mode (when elements are concentrated in few categories) or normal mode (when elements are widely distributed), optimizing power consumption while ensuring calculation completeness adapts to different data scenarios.
2Loss of energy
If calculation circuits operate in low power mode with element classification, then power consumption is reduced, but device complexity increases due to control and routing circuits
Solution Approach 1:
The control circuit performs preliminary classification of first matrix elements before multiplication operations begin. By categorizing elements in advance and determining the appropriate operation mode beforehand, the system avoids complex real-time decisions during multiplication, reducing control complexity while enabling power optimization.
Solution Approach 2:
The routing circuit acts as an intermediary between calculation circuits and multiplication circuits. It intelligently directs product values to appropriate calculation circuits based on their group assignments, simplifying the control logic by separating the routing function from the multiplication control, thereby reducing overall device complexity.
3Loss of energy
If element values are classified into categories before multiplication, then repeated operations are reduced and power consumption decreases, but processing time increases due to classification step
Solution Approach 1:
The classification of first matrix elements is performed as a preliminary action before the multiplication operations begin. By completing the classification and determining operation modes in advance, the system enables subsequent low-power multiplication operations without incurring time penalties during the actual computation phase, as the classification overhead is incurred only once.
Solution Approach 2:
The system employs periodic classification and mode determination for different columns of the first matrix. By classifying elements column-by-column and switching between low power and normal modes periodically based on classification results, the system minimizes the time spent in high-power states while ensuring complete processing, thereby reducing overall processing time compared to continuous high-power operation.
Data Source
AI summary
A matrix multiplication device and an operation method thereof are provided. The matrix multiplication device includes calculation circuits, a control circuit, a multiplication circuit, and a routing circuit. The calculation circuits produce multiply-accumulate values. The control circuit receives a plurality of first element values of a first matrix. The control circuit classifies the first element values into at least one classification value. The multiplication circuit multiplies the classification value by a second element value of a second matrix in a low power mode to obtain at least one product value. The routing circuit transmits each of the product values to at least one corresponding calculation circuit in the calculation circuits in the low power mode.