ISQED'20 Best Papers

Xianwei Cheng, Hui Zhao, Mahmut Kandemir, Saraju P. Mohanty, Beilei Jiang, Marzieh Vaeztourshizi, Mehdi Kamal, Massoud Pedram, Ming Feng Hsieh, Mohammad Nasim, Imtiaz Khan, Yuen Yee Cheng, Sung P. Lin, Abdullah Ash Saki, Swaroop Ghosh2020 · DOI

TL;DR

This document highlights three best papers from ISQED'20. The first paper optimizes DNN execution on GPUs by leveraging opportunistic computing, aiming to improve efficiency. The second paper explores the accuracy-energy trade-off in hardware implementations of error-resilient applications. The third paper presents a novel morphable physically unclonable function and true random number generator based on commercial magnetic memory.

What Problem Does This Paper Address?

The ISQED'20 conference showcases cutting-edge research, and this document specifically calls out three "Best Papers" that address critical challenges in computing and hardware design.

The first paper tackles the growing demand for efficient Deep Neural Network (DNN) execution on Graphics Processing Units (GPUs). As DNNs become more complex and widely used, optimizing their performance and energy consumption on hardware accelerators like GPUs is crucial. The challenge lies in maximizing throughput and minimizing latency without sacrificing accuracy, especially in resource-constrained or performance-critical applications.

The second paper addresses the fundamental trade-off between accuracy and energy efficiency in hardware implementations of error-resilient applications. Many applications, particularly in areas like embedded systems or fault-tolerant computing, can tolerate a certain level of error in their computations in exchange for significant power savings. The problem is to systematically explore and quantify this trade-off to enable designers to make informed decisions about hardware design and application deployment.

The third paper focuses on the development of secure and efficient hardware primitives for random number generation and physical unclonability. Physically Unclonable Functions (PUFs) are essential for hardware security, providing unique device fingerprints. True Random Number Generators (TRNGs) are vital for cryptography and simulation. The challenge is to create robust, reliable, and cost-effective solutions for these functionalities, potentially leveraging emerging memory technologies.

Methods

The document outlines the contributions of three distinct papers:

Paper 1: Alleviating Bottlenecks for DNN Execution on GPUs via Opportunistic Computing This paper proposes a method to optimize DNN execution on GPUs by employing "opportunistic computing." While the specifics of the opportunistic computing strategy are not detailed, it implies a dynamic allocation or utilization of GPU resources based on availability or specific computational needs to overcome existing bottlenecks.

Paper 2: EGAN: A Framework for Exploring the Accuracy vs. Energy Efficiency Trade-off in Hardware Implementation of Error Resilient Applications This paper introduces "EGAN," a framework designed to analyze the relationship between computational accuracy and energy consumption in hardware implementations of error-resilient applications. The framework likely involves modeling or simulation techniques to quantify these trade-offs, enabling designers to select optimal configurations.

Paper 3: A Morphable Physically Unclonable Function and True Random Number Generator using a Commercial Magnetic Memory This paper presents a novel approach to creating a morphable PUF and TRNG by utilizing commercial magnetic memory. The "morphable" aspect suggests that the PUF's characteristics or behavior can be altered, potentially for enhanced security or adaptability. The use of magnetic memory as the underlying technology is a key methodological innovation.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Opportunistic Computing for DNNs: The idea that intelligently leveraging available GPU resources in an "opportunistic" manner can significantly alleviate performance bottlenecks in DNN execution is a key insight. This suggests a move beyond static resource allocation towards more dynamic and adaptive strategies.
  • Quantifying the Accuracy-Energy Trade-off: The development of a framework like EGAN to systematically explore and quantify the accuracy-energy trade-off in error-resilient applications is crucial. This moves the design process from intuition to a data-driven approach, allowing for optimized hardware choices based on application requirements.
  • Magnetic Memory for Security Primitives: The innovative use of commercial magnetic memory for building both PUFs and TRNGs is a significant advancement. This highlights the potential of emerging memory technologies to serve dual purposes, offering both data storage and security functionalities.
  • Morphable PUFs: The concept of a "morphable" PUF is particularly interesting. It implies a level of programmability or adaptability in hardware security components that could lead to more robust and versatile security solutions, perhaps allowing for re-configuration against specific attacks or for different security levels.

Key Results

Not discussed in the paper.

Limitations and Future Work

Not discussed in the paper.

Summary updated

BibTeX
@misc{cheng2020isqed,
  author = {Xianwei Cheng and Hui Zhao and Mahmut Kandemir and Saraju P. Mohanty and Beilei Jiang and Marzieh Vaeztourshizi and Mehdi Kamal and Massoud Pedram and Ming Feng Hsieh and Mohammad Nasim and Imtiaz Khan and Yuen Yee Cheng and Sung P. Lin and Abdullah Ash Saki and Swaroop Ghosh},
  title = {{ISQED'20 Best Papers}},
  year = {2020},
  doi = {10.1109/isqed48828.2020.9137010},
}
Figure