SELF TEST AND IMPROVED SELF REPAIR FOR AI ACCELERATOR THROUGH RECURSIVE APPROACH

Many AI accelerators have been proposed as a result of the rapid advancement in artificial intelligence (AI) and the resulting growthin demand for data-intensive analytics. However, the dependability of AI-based solutions has become a crucial concern as they havebeen tailored to applications that demand precision and consistency. Self-testing and self-recovery for artificial intelligenceaccelerators are therefore suggested. With fewer test patterns and less test power, the suggested self-test can achieve 100%test accuracy for both delay and stuck-at faults. It is advanced utilizing scan chains made out of functional paths. The hybrid BISTself-test comprises logical BIST to confirm processor unit functionality and memory-BIST to test accumulators, registers, and buffers.Furthermore, the structural and functional properties of the systolic array in the AI accelerator are used to carry out self-diagnosisand self-recovery. Real-time progress is made using the suggested self-test and self-diagnosis, which enables precise fault localizationand fault type characterization. Efficient pruning for defective elements of processing with four processing levels is used to advancethe suggested self-recovery. It consists of row/column swapping, weight distribution, PE-bypass, and the removal of PE rows/columnswith more faults, or, alternatively, recursively employing PE without fault. Both sparse and dense networks can effectively apply thesuggested recovery technique. AI accelerators can significantly improve their reliability while barely sacrificing performance. Acomprehensive approach with little area overhead is offered by the suggested artificial intelligence assessment and recovery method.