A Reliable System Performance Through NeuToMa with Spiking Neural Network Transformation in Network on Chip (NoC) Communication

The NeuToMa transformation is all about changing a trained Spiking Neural Network (SNN) into a graph representation where neurons are nodes and synapses are edges. The INTMSN model is an integrated framework that aims to improve the performance and reliability of neuromorphic computing systems. It is made up of three major components: NeuToMa, Network on Chip (NoC) packet design, and the Reliability System Performance (RSP) model. NeuToMa alters SNNs to a graph representation that organizes neuron firing patterns that are optimized, and thus, communication overhead is reduced. The NoC packet design is all about making communication more efficient by having a well-defined structure for packet formation that will be used in data transfer management between processing elements and HBM. INTMSN achieved better latency results compared to MOATM and OCATN at rates of 0.03 and 0.05. Even at a 0.10 rate, INTMSN outperformed others with 27 cycles and significant improvement. In Transpose1, INTMSN surpassed others by 25% with a throughput of 0.061. For Transpose2 and Hotspot, INTMSN achieved throughputs of 0.070 and 0.064, respectively, outperforming MOATM and OCATN.