Hybrid Analog-Digital Neural Computing for Accuracy and Energy Balance
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 memory barriers and parallel processing limitations, 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, and eliminating the need for DRAM and external storage.
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 two distinct processing units: a digital processing unit for precise computations and an analog processing unit for energy-efficient operations. This segmentation allows each unit to operate in its optimal mode, with the digital unit handling tasks requiring high precision and the analog unit handling tasks where energy efficiency is paramount, thereby resolving the contradiction between accuracy and energy consumption.
Solution Approach 2:
Different computing tasks are assigned to different processing units based on their specific requirements. Digital MAC computations are performed in the digital processing unit where precision is critical, while analog MAC computations are performed in the analog processing unit where energy efficiency is prioritized. This local quality assignment ensures that each computation is performed in the most appropriate environment, balancing accuracy and energy consumption.
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:
The device is divided into digital and analog processing units, allowing the analog unit to handle computations that do not require extreme precision. This segmentation reduces the overall complexity and manufacturing cost compared to using entirely digital computing units, while maintaining sufficient accuracy for the analog-handled tasks through appropriate algorithm design and error correction mechanisms.
Solution Approach 2:
By assigning specific computation types to specific units based on their characteristics, the system optimizes manufacturing costs. The analog processing unit, which is cheaper to manufacture than digital units, handles computations where moderate precision is acceptable, thereby reducing overall device cost while maintaining adequate accuracy for those specific tasks.
3Use of energy by moving object
If analog computing is used for deep neural networks, then energy consumption is reduced and manufacturing cost decreases, but computing precision and reproducibility deteriorate
Solution Approach 1:
The computing workload is segmented between analog and digital processing units. The analog unit handles energy-efficient computations where moderate precision is acceptable, while the digital unit handles computations requiring high precision. This segmentation allows the system to leverage the energy efficiency of analog computing without sacrificing overall computing precision, as critical tasks are offloaded to the digital unit.
Solution Approach 2:
Different computation types are assigned to different units based on their precision requirements. Analog MAC computations are performed in the analog processing unit for energy efficiency, while digital MAC computations are performed in the digital processing unit for high precision. This local quality assignment ensures that each computation is performed in the most appropriate environment, balancing energy consumption and precision.
4Use of energy by moving object
If analog computing is used for deep neural networks, then energy consumption is reduced, but device complexity and programming difficulty increase
Solution Approach 1:
The system is segmented into digital and analog processing units, each with clearly defined functions. The digital unit handles tasks requiring high precision and complex control, while the analog unit handles energy-efficient computations with simpler control requirements. This segmentation manages device complexity by isolating the complex digital control functions from the simpler analog computation functions, making the overall system more manageable despite the presence of analog components.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach reduces power consumption and manufacturing costs, enabling efficient and cost-effective AI applications by leveraging parallel processing and minimizing energy usage.
Implementation Method 1
said input unit may include a digital-to-analog converter (DAC) that converts a digital signal into an analog signal
Implementation Method 2
said output unit may include an analog-to-digital converter (ADC)
Implementation Method 3
said output unit may include an analog-to-digital converter (ADC) and a sense amplifier (SA)
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An analog-digital hybrid deep neural network computing device according to the invention includes a control unit (4000), an analog processing unit (1000), a digital processing unit (2000), and a multi-channel bus (3000), 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.