Dynamic Precision Signal Coding for Sparse MAC Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing calculation architectures for artificial neural networks face high energy consumption due to the large number of MAC operations required for real-time processing, particularly when dealing with signals that have a high density of zero values, which leads to unnecessary operations if not accounted for in signal coding methods.
Innovation Solution
A dynamic precision coding method that optimizes signal encoding by decomposing quantized numbers into binary words based on the statistical distribution of zero values, allowing for the omission of zero binary words and adaptive sizing of encoded data to minimize MAC operations and energy consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional signal coding methods are used without considering signal characteristics, then the implementation is simple, but the number of MAC operations increases and energy consumption rises
Solution Approach 1:
The patent applies dynamic precision coding where the number of bits used to represent signal values is adjusted based on the actual signal characteristics. The system dynamically determines the precision level required for each signal based on its statistical properties, particularly the density of zero values, thereby adapting the coding scheme to minimize MAC operations while maintaining necessary computation accuracy.
Solution Approach 2:
The invention changes the parameter of signal representation precision based on signal characteristics. By analyzing the statistical distribution of signal values (especially zero density), the system modifies the number of bits allocated for coding different signal components, transforming fixed-precision coding into variable-precision coding to reduce unnecessary computations.
2Use of energy by moving object
If dynamic precision coding is implemented to optimize for signals with high zero density, then energy consumption is reduced, but the coding complexity increases
Solution Approach 1:
The patent performs preliminary analysis of signal characteristics (specifically zero density statistics) before encoding the signal. By pre-characterizing the signal distribution properties, the system can determine the optimal precision level and coding strategy in advance, avoiding the need for complex real-time adjustments during the encoding and MAC operations themselves.
3Use of energy by moving object
If the number of MAC operations is reduced through optimized coding, then energy consumption decreases, but calculation precision may be degraded
Solution Approach 1:
The patent applies different precision levels to different parts of the signal based on their local characteristics. Specifically, signal components with high zero density are encoded with lower precision (fewer bits), while other components maintain higher precision. This localized adaptation of precision ensures that energy is saved where possible without compromising the overall calculation accuracy required for the neural network operation.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
Method, implemented by computer, for coding a digital signal intended to be processed by a digital computing system, the method comprising the steps of: - receiving (101) a sample of the digital signal quantified on a number Nd of bits; - decomposing (102) the sample into a plurality of binary words of parameterizable size Np bits; - coding (103) the sample with a plurality of pairs of values, each pair comprising one of the binary words and an address corresponding to the position of the binary word in the sample; - transmitting (104) the pairs of values to an integration unit in order to perform a MAC operation between the sample and a weighting coefficient.