Dynamic-Precision Signal Coding for Low-Energy Digital MAC Operations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computing architectures for artificial neural networks face high energy consumption due to the restrictive number of multiply-accumulate (MAC) operations required for real-time operation, particularly when dealing with signals of varying dynamic ranges, leading to sub-optimal sizing of computing operators and increased energy usage.
Innovation Solution
A method and device for coding digital signals with dynamic precision, where signals are decomposed into maximum values and residual values, allowing for reduced bit size and optimized MAC operations, thereby minimizing energy consumption by adapting to the statistical distribution of signal values across different layers of the neural network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by stationary object
If fixed-bit quantization is used for all signal samples, then the computing operators can be sized consistently, but the energy consumption increases due to sub-optimal sizing when dynamic range varies
Solution Approach 1:
The patent applies dynamic quantization where the bit depth for coding signal samples is adjusted based on the actual dynamic range of the signal. Instead of using a fixed number of bits for all samples, the system determines the appropriate bit depth dynamically according to the signal's statistical properties, thereby optimizing the sizing of computing operators (particularly MAC operators) to match the actual signal requirements and reduce energy consumption.
2Use of energy by stationary object
If the number of MAC operations is reduced for real-time operation, then energy consumption decreases, but the precision of computations may be degraded
Solution Approach 1:
The patent changes the parameter of signal representation by decomposing samples into a sum of maximum values (equal to 2^Np-1) and residual values. This parameter transformation allows the system to perform MAC operations with reduced bit width (Np bits) while maintaining computational precision, as the decomposition preserves the essential information needed for accurate computation despite using fewer bits.
3Ease of manufacture
If quantization on a fixed number of bits is applied to all samples, then the coding process is simplified, but the dynamic range representation becomes sub-optimal
Solution Approach 1:
The patent segments each signal sample into multiple components: a sum of maximum values (each equal to 2^Np-1) and a residual value. This segmentation allows the system to represent the full dynamic range of the signal using a combination of standardized maximum value units and a smaller residual component, achieving both simplified coding (using repeated maximum values) and precise dynamic range representation (through the residual).
Data Source
AI summary
A computer-implemented method is provided for coding a digital signal quantized on a given number Nd of bits and intended to be processed by a digital computing system, the signal being coded on a predetermined number Np of bits which is strictly less than Nd, the method including the steps of: receiving a digital signal composed of a plurality of samples, decomposing each sample into a sum of k maximum values which are equal to 2NP−1 and a residual value, with k being a positive or zero integer, successively transmitting the values obtained after decomposition to an integration unit for carrying out a MAC operation between the sample and a weighting coefficient.