http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
이기종프로세서 사용수명 신뢰성 평가를 위한 모델링 방법론
정수안(Suan Jung),황정민(Jeongmin Hwang),김보길(Bogil Kim),송진호(William J. Song) 대한전자공학회 2022 대한전자공학회 학술대회 Vol.2022 No.11
Research and industry have adopted heterogeneous architecture designs to improve performance and energy efficiency of computing systems. However, different types of processing units playing distinct roles complicates the lifetime reliability modeling. It is difficult to estimate system-level lifetime reliability of heterogeneous processors under various design specifications and operating conditions. The paper proposes a framework RelSim for lifetime reliability evaluation of the heterogeneous system. The proposed framework is capable of modeling various designs of heterogeneous processors with a various mix of failure mechanisms and statistical distributions under multiple execution scenarios. The framework is also capable of reliability management schemes. In addition, GPU parallelization accelerates the Monte Carlo simulation. The flexible framework expects users extend research into lifetime reliability.
이기종 프로세서 사용수명을 고려한 신경망 가속 스케쥴링
황정민(Jeongmin Hwang),정수안(Suan Jung),김보길(Bogil Kim),송진호(William J. Song) 대한전자공학회 2022 대한전자공학회 학술대회 Vol.2022 No.11
In this paper, we analyze the impact on the lifetime reliability of heterogeneous processors while accelerating neural network execution. Most edge devices are designed with heterogeneous processors that integrate CPUs and accelerators to process vast amounts of data quickly. Although performance has been improved due to device expansion, energy efficiency has been reduced due to increased power and energy consumption. Therefore, lifetime reliability management has become more difficult due to complex processor design structures. As neural networks become more common in various industries, performance, energy efficiency and lifetime reliability management become more important. This paper presents parallel scheduling of CPUs and accelerators in neural networks acceleration, named cooperative scheduling. We propose a method to improve performance while satisfying the required lifetime reliability through adopting a combination of improved scheduling and lifetime reliability management methods during neural network acceleration.