Customizable AI Chip With Near-Memory CNN Layout for Low Power
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional CNN-based systems for image processing and object recognition require significant power consumption, making them unsuitable for deployment on small devices without constant power sources.
Innovation Solution
Implementing a CNN on a programmable logic device (PLD) or ASIC, utilizing a near-memory architecture and parallel computation to reduce power consumption by up to a factor of 100 compared to CPU or GPU implementations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If CNN is implemented on CPU or GPU, then computational capability is provided, but power consumption is excessive
Solution Approach 1:
The patent segments the CNN architecture into multiple parallel processing units, each dedicated to specific convolutional layers. This segmentation allows independent processing of different computational tasks simultaneously, reducing the power consumption of individual units while maintaining overall computational capability through parallel execution across multiple units.
Solution Approach 2:
The patent transitions from sequential processing (traditional CPU/GPU approach) to parallel processing by adding a spatial dimension of computation. Multiple convolutional layers are processed simultaneously across different processing units, effectively changing the computational dimension from time-sequential to space-parallel execution, thereby reducing power consumption per operation.
2Ease of operation
If CNN is implemented on small devices, then portability is improved, but power consumption becomes excessive
Solution Approach 1:
The patent implements a streamlined CNN architecture that processes only the essential computational steps required for image classification. By using a simplified parallel processing structure with dedicated units for each convolutional layer, the system performs partial computations that are sufficient for the application while consuming significantly less power than full CNN implementations on mobile devices.
3Adaptability or versatility
If ambient power sources are used, then deployability on small devices is improved, but power availability is limited
Solution Approach 1:
The patent adjusts the computational parameters of the CNN system to match the power availability constraints of ambient power sources. By optimizing the number of parallel processing units, the depth of convolutional layers, and the precision of computational operations, the system achieves acceptable performance while operating within the limited power budget provided by ambient energy sources.
Data Source
AI summary
In one embodiment, a computing device includes an input sensor providing an input data; a programmable logic device (PLD) implementing a convolutional neural network (CNN), wherein: each compute block of the PLD corresponds to one of a multiple of convolutional layers of the CNN, each compute block of the PLD is placed in proximity to at least two memory blocks, a first one of the memory blocks serves as a buffer for the corresponding layer of the CNN, and a second one of the memory blocks stores model-specific parameters for the corresponding layer of the CNN.