|
Abstract
The development of the building blocks for the latest technology in multiple areas (AI, robotics, communication, gaming, automotive, scientific research, etc) demands a fast and high-performance computing solution for the EM field. PeakView offers an efficient flow and methodology that leverages GPU cards to accelerate the EM simulation during the design process of high-speed layouts. These designs are increasingly being developed on TSMC’s advanced process technologies such as the COmpact Universal Photonic Engine (COUPE), which integrates electronic (EIC) and photonic chips (PIC) in an EIC-on-PIC configuration. With the goal of reducing power consumption and increasing speed, the design in these new technologies require EM simulation at high frequencies with 3D accuracy.
In this paper, we present the brute-force 3D EM simulation study on the metalfill impacts in the photonic design with computations massively accelerated by GPU and distributed computing. Previous brute-force metalfill EM simulations showed a degradation of up to ~13% on the performance of some devices; therefore, accounting for this impact during the photonic design signoff is critical. The PIC performance was analyzed at different temperatures, and all results included not only the nport file output, but also an equivalent RLCK circuit model that helps designers to gain insight in the photonic transmission line parasitics. The simulation approach was based on the implementation of NVIDIA GPU cards and distributed computing capabilities to establish an optimum design flow that guarantees rapid EM simulations with sufficient accuracy. We obtained up to 6X speed-up by a single GPU card compared to regular CPU-based nodes. This includes both NVIDIA’s RTX and the DGX GPU platforms, and this performance can be even extrapolated with running frequencies in parallel with distributed computing. For effective use of GPU cards, single precision was selected from PeakView’s settings considering that some GPU cards may be optimized to provide better performance and memory consumption with single precision. This methodology was used to compare the s-parameters of the photonic transmission line at different frequencies and temperatures, achieving a small shift in the temperature response in front of 175℃ temperature increase. This work exposes some of the multiple factors that may impact the performance of EIC-on-PIC integrated designs, starting from the EM accuracy, suggested high frequency simulation, temperature variability, and design flow integration. Considering all these factors creates a solid methodology for the development of future IC and 3DIC photonic applications.
|
