Data Transfer Limitations and Accuracy Trade-Offs in Neuromorphic Real-Time Systems

Neuromorphic computing is an event-driven and sparse, biologically inspired computing paradigm, offering strong potential for resource-efficient real-time inference. However, event-based sensors produce highly variable event rates, which can challenge predictable deployment on performance constrained systems. In this work, we analyze data-transfer limitations and accuracy trade-offs for DVS128 Gesture classification on SpiNNaker2 using a passive input buffer and buffer-aware training. We show that a buffer could constrain the number of input events to mitigate event fluctuations and still preserves performance. On SpiNNaker2 (High PL, 1000 timesteps, 1 ms tick), buffered models achieve 89.39–91.29% on-chip accuracy, compared to 91.67% for the no-buffer baseline. At the same time, average on-chip energy is reduced from 0.800 J without buffering to 0.510 J with buffering. These results demonstrate that buffer-aware deployment can substantially improve energy efficiency while maintaining competitive accuracy, enabling more predictable and robust neuromorphic real-time systems.