Neural Network MAC Array with Zero-Weight Skipping Modes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing electronic devices face inefficiencies in processing data using neural networks due to the need for mathematical operations between weights and input data, which are not optimized according to the device's situation, leading to suboptimal performance and power consumption.
Innovation Solution
The implementation of an electronic device with a switch to select modes and a MAC array that includes a zero weight skip unit or MAC activation control unit, allowing for the selection and forwarding of relevant weights to perform convolution operations efficiently, thereby optimizing data processing based on the device's conditions such as resolution, power state, or temperature.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If convolution operations are performed between input data and all weights, then processing completeness is improved, but power consumption increases
Solution Approach 1:
The patent extracts and processes only non-zero weights from the full weight set. The weight processing unit identifies and separates non-zero weights from zero weights, then performs convolution operations only with the extracted non-zero weights. This extraction principle directly reduces the number of mathematical operations required, thereby lowering power consumption while maintaining processing completeness for all meaningful data transformations.
Solution Approach 2:
The patent applies partial action by performing convolution operations only with a subset of weights (non-zero weights) rather than all weights. The mode selection unit can choose to process only essential non-zero weight convolutions under power-constrained conditions, providing a partial processing mode that balances completeness requirements with energy consumption limits.
2Measurement precision
If convolution operations are performed with all weights, then processing accuracy is improved, but processing speed decreases
Solution Approach 1:
The patent extracts only non-zero weights from the complete weight set and performs convolution operations exclusively with these extracted weights. This extraction eliminates redundant operations with zero weights, directly increasing processing speed while preserving all accuracy-critical computations that involve non-zero weight values.
Solution Approach 2:
The patent changes the operational parameter from processing all weights to processing only non-zero weights. The weight processing unit dynamically identifies which weights have non-zero values and adjusts the processing pipeline accordingly, transforming the fixed-parameter approach of universal weight processing into a variable-parameter approach that optimizes for both speed and accuracy.
3Use of energy by moving object
If mode selection is added to control weight processing, then power efficiency is improved, but device complexity increases
Solution Approach 1:
The mode selection unit serves multiple functions: it determines processing modes based on power conditions, controls the weight processing unit's behavior, and manages the flow of weight data to the MAC array. This multi-functional component consolidates control logic that could otherwise be distributed across multiple separate units, reducing overall device complexity while enabling power-efficient mode selection.
Solution Approach 2:
The weight processing unit acts as an intermediary between the mode selection unit and the MAC array. It receives mode selection signals, processes weights accordingly (either all weights or only non-zero weights), and outputs results to the MAC array. This intermediary structure provides a clean interface that manages complexity by centralizing the decision-making logic for weight processing in a single dedicated unit.
4Productivity
If zero weight skip unit is used to forward only non-zero weights, then processing speed is improved, but device complexity increases
Solution Approach 1:
The patent merges the functions of weight filtering and weight forwarding into a single weight processing unit. This unit simultaneously identifies non-zero weights, filters them from the zero weights, and forwards them to the MAC array in one integrated operation. By combining these functions rather than implementing them as separate sequential steps, the patent reduces the number of discrete components required, thereby limiting the increase in device complexity while achieving processing speed improvements.
Data Source
AI summary
An electronic device is provided. The electronic device includes a switch configured to select a mode of the electronic device, a multiply and accumulate (MAC) array configured to include a plurality of MAC units, and at least one processor configured to include a zero weight skip unit for confirming a first weight having a value of ‘0’ among weights related with input data, and for forwarding a second weight not having a value of ‘0’ among the weights, to the MAC array. The at least one processor is configured to acquire the input data, acquire the weights, select the mode of the electronic device by using the switch, in response to a first mode of the electronic device being selected, perform convolution operations between the input data and the second weight forwarded to the MAC array through the zero weight skip unit, and, in response to a second mode of the electronic device being selected, perform convolution operations between the input data and the weights forwarded to the MAC array.