Abstract
Resistive crossbar arrays have been shown to enable the implementation of energy-efficient in-memory computing accelerators suitable for the diffusion of artificial neural networks (ANNs) at the edge. However, the effect of line parasitic resistances can degrade the performance of ANNs if not appropriately considered during the design of the accelerator and each time a new task is targeted. Poised by these limitations, several crossbar line parasitic resistance models have been proposed in the literature. However, a comparative study of the performance of these models is still missing. In this work, we compare and benchmark over a broad range of operating conditions several crossbar line parasitic resistance models from the literature. The results emphasize a practical trade-off: compact analytical models are fast but provide reliable estimates only for small (i.e., 32 × 32) array sizes; conversely, iterative numerical models require more computing time but are more accurate under all conditions. Thus, iterative methods remain indispensable for larger or more complex designs. In addition, the implications of using these models in line parasitic-aware training of the ANN are analyzed on an MNIST classification task. The results further indicate that despite providing a considerable speedup of the ANN training, analytical models preserve accuracy only when small crossbar tiles are used, highlighting the need for more dependable yet fast compact models.
Funding Information
  • NextGenerationEU (ECS_00000033_ECOSISTER)

This publication has 28 references indexed in Scilit: