Optimized co-scheduling of mixed-precision neural network accelerator for real-time multitasking applications

Jiang, Wei; Song, Ziwei; Zhan, Jinyu; He, Zhiyuan; Wen, Xiangyu; Jiang, Ke

doi:10.1016/j.sysarc.2020.101775

Cited by 38 publications

(7 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Hence, sometimes there is a need to perform additional tests to determine the balance between the rate of action and the level of confidence in assessing the risk of emergencies. W. Jiang et al dealt with optimising neural network operation for real-time multitasking applications [15]. This work will develop general guidelines for the neural network model training using different methodologies with full and partial supervision for the Prescriptive Maintenance strategy (RxM).…”

Section: Analysis Of Neural Network Training Algorithms For Implement...mentioning

confidence: 99%

Analysis of Neural Network Training Algorithms for Implementation of the Prescriptive Maintenance Strategy

LEMPA¹

2022

Terotechnology XII

View full text Add to dashboard Cite

Abstract. This paper presents a proposal to combine supervised and semi-supervised training strategies to obtain a neural network for use in the prescriptive maintenance approach. It is required in this approach because of only partially labelled data for use in supervised learning, and additionally, this data is predicted to expand quickly. The main issue is the decision on which are suitable training methodologies for supervised learning, having in mind using this data and methods for semi-supervised learning. The proposed methods of training neural networks with supervised and semi-supervised training to receive the best results will be tested and compared in further work.

show abstract

Section: Analysis Of Neural Network Training Algorithms For Implement...mentioning

confidence: 99%

Analysis of Neural Network Training Algorithms for Implementation of the Prescriptive Maintenance Strategy

LEMPA¹

2022

Terotechnology XII

View full text Add to dashboard Cite

show abstract

“…The PEs (processing elements) support bit-serial multiplication and the weight bits are sent serially with additional PEs used to compensate for the lower throughput of the bit-serial architecture, which is much simpler than bit-parallel multiplication hardware. The concept of using mixed-precision hardware is also explored in [16] which proposes a scheduling strategy to distribute real-time tasks associated with sensor data acquisition, inference and action on a heterogenous system that combines an FPGA and CPU. The FPGA inference accelerator supports precisions from 8 to 64 bits, resulting in different computation times that the scheduler needs to take into account.…”

Section: Introductionmentioning

confidence: 99%

Energy-efficient neural networks with near-threshold processors and hardware accelerators

Nunez-Yanez

Howard

2021

Journal of Systems Architecture

View full text Add to dashboard Cite

“…A Bluetooth 5 powered architecture was demonstrated by Fraga‐Lamas et al, 8 which not only has the ability to quickly response to critical events but also can handle computing‐intensive complex tasks. Jiang et al 9 focused on reducing the execution time of artificial intelligent applications on hybrid CPU and field‐programmable‐gate‐array platforms. However, all the above research works 4‐9 fail to take into account the energy budgets and reliability requirements in designing CPS algorithms.…”

Section: Introductionmentioning

confidence: 99%

“…Jiang et al 9 focused on reducing the execution time of artificial intelligent applications on hybrid CPU and field‐programmable‐gate‐array platforms. However, all the above research works 4‐9 fail to take into account the energy budgets and reliability requirements in designing CPS algorithms.…”

Section: Introductionmentioning

confidence: 99%

Exploring reliable edge‐cloud computing for service latency optimization in sustainable cyber‐physical systems

et al. 2021

View full text Add to dashboard Cite

In recent years, the advance in information technology has promoted a wide span of emerging cyber‐physical systems (CPS) applications such as autonomous automobile systems, healthcare monitoring, and process control systems. For these CPS applications, service latency management is extraordinarily important for the sake of providing high quality‐of‐experience to terminal users. Edge‐cloud computing, integrating both edge computing and cloud computing, is regarded as a promising computation paradigm to achieve low service latency for terminal users in CPS. However, existing latency‐aware edge‐cloud computing methods dedicated for CPS fail to jointly consider energy budgets and reliability requirements, which may greatly degrade the sustainability of CPS applications. In this article, we explore the problem of minimizing service latency of edge‐cloud computing coupled CPS under the constraints of energy budgets and reliability requirements. We propose a two‐stage approach composed of static and dynamic service latency optimization. At static stage, Monte‐Carlo simulation with integer‐linear‐programming technique is adopted to find the optimal computation offloading mapping and task backup number. At dynamic stage, a backup‐adaptive dynamic mechanism is developed to avoid redundant data transmissions and executions for achieving additional energy savings and service latency enhancement. Experimental results show that our solution is able to reduce system service latency by up to 18.3% compared with representative baseline solutions.

show abstract

Optimized co-scheduling of mixed-precision neural network accelerator for real-time multitasking applications

Cited by 38 publications

References 19 publications

Analysis of Neural Network Training Algorithms for Implementation of the Prescriptive Maintenance Strategy

Analysis of Neural Network Training Algorithms for Implementation of the Prescriptive Maintenance Strategy

Energy-efficient neural networks with near-threshold processors and hardware accelerators

Exploring reliable edge‐cloud computing for service latency optimization in sustainable cyber‐physical systems

Contact Info

Product

Resources

About