Core Concepts
Partially-precise computing paradigm improves hardware efficiency for application-specific embedded systems.
Abstract
The article introduces the concept of partially-precise computing to enhance the implementation of application-specific embedded systems. It discusses the conflict between conventional precise computational blocks and custom applications, proposing a novel computational paradigm inspired by neuroscience. The content covers the design flow, implementation processes, and benefits of partially-precise computing in Gaussian denoising filters, image blending, and face recognition neural networks.
-
Introduction
- Emerging embedded systems in various domains.
- Limitations of conventional precise digital computational blocks.
-
Partially-Precise Computing Paradigm
- Introduction to partially-precise computing.
- Inspiration from brain information reduction hypothesis.
-
Design Flow and Implementation
- Development process for customized partially precise computational blocks.
- Experimental results on Gaussian denoising filters, image blending, and face recognition neural networks.
-
Gaussian Denoising Filter Implementation
- Utilizing natural sparsity for improved physical properties without accuracy degradation.
-
Image Blending Implementation
- Leveraging intentional sparsity through preprocessing for enhanced hardware efficiency.
-
Face Recognition Neural Network Implementation
- Natural sparsity utilization in MAC multipliers for cost-effective implementations.
-
Thresholding Sparsity in Face Recognition
- Intentional sparsity introduction through thresholding preprocessing for further efficiency gains.
Stats
The DS2-like algorithmic sparsities improve PPC block implementation costs significantly.
Applying DS16 creates a 93% sparsity with acceptable output quality degradation.
Quotes
"As both paradigms are inspired from biological brain operation, they can be utilized complementarily."
"Utilization of natural sparsity does not degrade system accuracy while improving implementation costs."