Scalability of Ultralow-Loss Calibration-Free Silicon Photonic Mach-Zehnder Switches

Abstract

With the rapid development of artificial intelligence (AI) based on deep neural networks, large-scale photonic switches are essential components for the fast and efficient communication of unprecedentedly large amounts of data between processing units and memories. In this paper, a comprehensive Monte Carlo analysis is provided on the scalability of Mach–Zehnder switch (MZS) networks utilizing Benes topologies as an example, employing the transfer matrix method. The results show that iterative calibration algorithms with high time complexities are infeasible for large-scale MZSs with significant random phase imbalances, which, instead of the excess loss, is the dominant fundamental obstacle for scaling up MZS. Therefore, calibration-free MZSs are crucial for scaling up. To further validate the key assumptions of the Monte Carlo analysis above, ultralow-loss 2 × 2 MZSs and 4 × 4 Benes MZSs fabricated with standard 180-nm silicon photonics foundry processes are systematically characterized. Drawing from the statistical experimental results of random phase imbalance and excess loss, the scalability of the Benes topology is projected and concludes that it is promising to realize large-scale, low-excess-loss, calibration-free N × N photonic switches (e.g., N ≥ 64) based on these proposed MZS for agile, flexible, and scalable optical packet/burst switching (OPS/OBS) in data centers.