This paper presents a soft-error-tolerant memory-control circuit for SRAM-based field programmable gate arrays (FPGAs). A potential obstacle to applying such FPGAs to safety-critical industrial control systems is their low tolerance. The main reason is that soft errors damage circuit-configuration data stored in SRAM-based configuration memory. To overcome this obstacle, the soft-error tolerance must thus be improved while suppressing the circuit area overhead, and data stored in external memory must be protected when a fault occurs on the FPGA. Therefore, a memory-control circuit was developed on the basis of a dual-modular-redundancy (DMR) architecture. This memory controller has a repair and retry scheme that repairs damaged circuit-configuration data and re-executes unfinished accesses after the repair. The developed architecture reduces circuit redundancy below that of a commonly used triple-modular-redundancy (TMR) architecture. Moreover, a write-invalidation circuit was developed to protect data in external memory, and an external-memory-state recovery circuit was developed to enable resumption of memory access after fault repair. The developed memory controller was implemented in a prototype circuit on an FPGA and evaluated using the prototype. The evaluation results demonstrated that the developed memory controller can operate successfully for 1.03×109 hours (at sea level). In addition, its circuit area overhead was found to be sufficiently smaller than that of the TMR architecture.

The reliability of deep neural networks (DNN) against hardware errors is essential as DNNs are increasingly employed in safety-critical applications such as automatic driving. Transient errors in memory, such as radiation-induced soft error, may propagate through the inference computation, resulting in unexpected output, which can adversely trigger catastrophic system failures. As a first step to tackle this problem, this paper proposes constructing a vulnerability model (VM) with a small number of fault injections to identify vulnerable model parameters in DNN. We reduce the number of bit locations for fault injection significantly and develop a flow to incrementally collect the training data, i.e., the fault injection results, for VM accuracy improvement. We enumerate key features (KF) that characterize the vulnerability of the parameters and use KF and the collected training data to construct VM. Experimental results show that VM can estimate vulnerabilities of all DNN model parameters only with 1/3490 computations compared with traditional fault injection-based vulnerability estimation.