NOMA: A Novel Reliability Improvement Methodology for 3-D IC-based Neuromorphic Systems

Hardware is the foundation of trust for all real-time applications and is especially crucial in Artificial Intelligence (AI). The growing reliance on AI for various tasks has raised concerns about the reliability of the hardware. Extensive research has shown that hardware faults arising from the manufacturing process or device variation can significantly impact the precision and accuracy of AI applications. This also applies to 3-D IC (Three-Dimensional Integrated Circuit)-based Spiking Neural Networks (SNNs), despite their noise resilience, low-power, and low memory footprint advantages. This is because, in addition to the mentioned hardware faults, thermal dissipation on 3-D ICs can negatively affect the memory stacked on upper dies. Hence, this article proposes a methodology called NOMA (Network-of-Memory Architecture) to enhance the reliability of 3-D IC-based SNNs by replacing the defective critical synaptic weights in high-priority layers with lower-priority ones. The method is assessed on multiple memory-on-logic architectures with various memory technologies such as SRAM, eDRAM, STT-RAM, and RRAM. As a result, for the 45nm SRAM library, the accuracy degradation improves by roughly 56% at a BER of 0.05 with a timing overhead of 3.538μs and a power overhead of 390.796nJ.