Neural Networks for Fast Design Space Exploration of On-Chip Interconnect Networks

In this chapter, we will look at how artificial neural network (ANN) models may be used to design on-chip interconnects for integrated circuits and systems. Conduct of on-chip copper (Cu) interconnect networks is restricted by dispersive processes such as grain boundary scattering, surface roughness scattering, top/bottom surface, and sidewall scattering when minimum interconnect width scales below the 22 nm mark. The resistance of the interconnects, measured over a unit of length, is greatly inflated over the bulk value due to these scattering phenomena. Enhanced signal attenuation, delay, and power loss are all consequences of such extremely resistive interconnects. Augmenting to causing discontinuities in the interconnect line, the migration of copper ions into the dielectric layer increases dielectric conductivity and leakage losses. To prevent copper ions from diffusing away from the copper conductor, a barrier layer is often positioned around it. Recently, many studies have investigated the potential of employing graphene nanoribbons as a replacement barrier material for Cu interlinks. At ambient temperature, the mean free route for electrons in graphene nanoribbons is much greater than in copper. Therefore, in addition to the copper conductor, the barrier layer provides extra low-resistance routes for conducting electrons. It is becoming more important to do SPICE-based simulations of hybrid Cu-Graphene on-chip interconnect networks due to their rising popularity. Simulations of interconnects using SPICE are notoriously time-consuming and computationally intensive. Using surrogate models is one way to deal with the computational overhead of exploring design space. Due to their capacity to mimic the extremely nonlinear input–output correlations of electronic packaging systems, surrogate templates based on machine learning (ML) regression are now quite sought-after. We will also collate the results of our ML-based models to those of our comprehensive EM and SPICE simulations.