Mapping and virtual neuron assignment algorithms for MAERI accelerator

To date, some different deep learning accelerators (DLAs) have proposed to solve challenges caused by increasing deep neural networks’ layers. GPU-based systems almost faced energy efficiency problems due to the parallel computing operations and increasing memory accesses that led to memory capacity, bandwidth requirement, and delay challenges. DLAs-based systems tried to overcome the challenges and improve the parameters, which their flexibility remains a challenge. Some case studies investigated the proposed DLAs and demonstrated an impressive effect of different mapping methods on reducing energy consumption and delay. We analyze MAERI’s role in solving the issue and the impact of mapping methods to face the challenges induced by implementing different DNN trained models using the accelerators. This work proposes an algorithm for mapping and assigning virtual neurons (VNs) on the MAERI accelerator to improve its performance and cost. The simulation results demonstrate the reducing energy consumption and delay of approximately 21–92% and 14–21% caused by implementing AlexNet and VGG-16 on MAERI, respectively. The mapping method has a significant effect in increasing the proposed DLAs’ performance and reducing their cost without redesign their structures. The proposed VNs assignment approach helps to support different DNN trained models and increases DLA-based systems' flexibility.