Hybrid Neural MAC Architecture for Accuracy-Energy Tradeoffs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital computing units for deep neural networks face high energy consumption, large size, and high cost due to limitations in parallel processing and memory barriers, while analog computing offers low cost and reduced energy but struggles with precision and reproducibility issues.
Innovation Solution
An analog-digital hybrid deep neural network computing device that combines analog and digital processing units, performing MAC computations in an analog manner for some layers and digital manner for others, utilizing non-volatile memories and SRAM without DRAM, and includes converters and a multi-channel bus for efficient information transfer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If digital computing units are used for deep neural networks, then high accuracy is achieved, but energy consumption increases and device size becomes large
Solution Approach 1:
The computing device is segmented into multiple computing units, each capable of independent MAC operations. This segmentation enables parallel processing across multiple units, reducing the energy burden on individual units while maintaining overall computational accuracy through coordinated operation of the plurality of computing units.
Solution Approach 2:
A weight memory unit is introduced as an intermediary component that stores synaptic weights and provides them to multiple computing units. This intermediary approach eliminates the need for each computing unit to have separate weight storage, reducing total memory energy consumption while maintaining access to all necessary weights for accurate computation.
2Measurement precision
If digital computing units are used for deep neural networks, then high accuracy is achieved, but device size and manufacturing cost increase
Solution Approach 1:
Each computing unit is designed as a universal module capable of performing the same MAC operation with identical structural components. This universality allows multiple units to be fabricated using the same design blueprint, reducing overall device complexity while enabling parallel processing that maintains computational accuracy across the entire system.
Solution Approach 2:
The patent employs multiple copies of the same computing unit structure, where each unit is an identical replica designed for parallel operation. This copying approach simplifies manufacturing by using repeated patterns, reducing device complexity while maintaining computational accuracy through the collective output of multiple identical units performing the same function.
3Use of energy by moving object
If analog computing is used for deep neural networks, then energy consumption is reduced and cost decreases, but precision and reproducibility deteriorate
Solution Approach 1:
The patent replaces analog mechanical computing systems with digital electronic computing units that perform MAC operations using digital logic. This substitution maintains the low energy consumption advantage while achieving high computational precision through digital arithmetic operations, eliminating the precision and reproducibility issues inherent in analog systems.
Solution Approach 2:
The computing units utilize configurable parameters including selectable activation functions and adjustable weight values stored in the weight memory unit. These parameter changes enable the system to adapt to different computational requirements while maintaining digital precision, allowing the same hardware to achieve both energy efficiency and high computational accuracy across various neural network configurations.
Data Source
AI summary
An analog-digital hybrid deep neural network computing device according to one embodiment includes a control unit, an analog processing unit, a digital processing unit, and a multi-channel bus, and in which the algorithm includes a plurality of layers for computation, and an analog MAC computation in an analog computing manner is performed for a first group of layers including one or more layers among the plurality of layers in the analog processing unit and a digital MAC computation in a digital computing manner is performed for a second group of layers, which is the remaining layers except for the first group of layers, in the digital processing unit.