RePAIR: Reconfigurable Platform for AI Resilience within RISC-V Ecosystem

Recently, platforms combining RISC-V processors with accelerators for deep-learning applications have gained popularity even for high-reliability applications such as avionics and space. However, for high-performance safety-critical systems, it is mandatory to couple high-performance architecture with reliable mechanisms for coping with errors and faults. We propose the first FPGA-based architecture that combines a RISC-V processor with a systolic array-based accelerator, a fault detection, fault correction, and an execution recovery mechanism. The proposed solution corrects faults in the systolic array datapath by exploiting a partial reconfiguration mechanism. When an error is detected, the RISC-V processor can trigger the accelerator reconfiguration, correcting the fault. Furthermore, the approach allows resuming the inference from the last correctly executed step, significantly reducing the availability overhead. The approach results in a high-performance and high-reliable platform that can autonomously detect and correct faults, providing execution continuity and minimal system downtime.