The impact of the soft errors in convolutional neural network on GPUs: Alexnet as case study

Khalid Adam, Izzeldin I. Mohd, Younis M. Younis
2021 Procedia Computer Science  
Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads,
more » ... ch makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions. Abstract Convolutional Neural Networks (CNNs) have been increasingly deployed in many applications, including safety critical system such as healthcare and autonomous vehicles. Meanwhile, the vulnerability of CNN model to soft errors (e.g., caused by radiation induced) rapidly increases, thus reliability is crucial especially in real-time system. There are many traditional techniques for improve the reliability of the system, e.g., Triple Modular Redundancy, but these techniques incur high overheads, which makes them hard to deploy. In this paper, we experimentally evaluate the vulnerable parts of Alexnet mode (e.g., fault injector). Results show that FADD and LD are the top vulnerable instructions against soft errors for Alexnet model, both instructions generate at least 84% of injected faults as SDC errors. Thus, these the only parts of the Alexnet model that need to be hardened instead of using fully duplication solutions.
doi:10.1016/j.procs.2021.02.012 fatcat:jwloam4rcbhbpncybbdmwwszzi