HMR-NEureka: Hybrid Modular Redundancy DNN Acceleration in Heterogeneous RISC-V SoCs

On-board Artificial Intelligence (AI) processing is widely considered to be necessary for the next generation of satellites. Adequate levels of protection against radiation-induced errors, at an affordable cost, are thus required for AI accelerators in satellites. This paper presents HMR-NEureka, a Hybrid Modular Redundant accelerator designed for Deep Neural Network (DNN) inference in heterogeneous RISC-V System-on-Chips (SoCs). HMR-NEureka features two operational modes: a redundancy mode leveraging dual modular redundancy (DMR) with low-overhead hardware-based recovery, and a performance mode that repurposes redundant datapaths to improve throughput for non-critical operations. Implemented as an extension of the opensource NEureka accelerator, HMR-NEureka integrates into a multi-core RISC-V cluster with Error Correction Codes (ECC), ensuring end-to-end fault protection. Our results, based on a GlobalFoundries 12 nm technology implementation, demonstrate a 93% reduction in faulty executions in redundancy mode with moderate area (< 10%) and performance (∼ 5%) overhead compared to a non-fault-tolerant baseline. These findings establish HMR-NEureka as a flexible and efficient AI accelerator for mission-critical space applications.