Researchers from Georgia Tech have examined a bottleneck in wafer-scale optical interconnects used for Mixture-of-Experts LLM1. local linear models2. Large Language ModelThe model class referenced in the long-context serving demo. training: repeated stalls caused by thermal tuning during communication phases. In their paper, they frame the problem as a cross-layer overhead that directly interferes with training flow rather than as a minor control-side inefficiency.
The study focuses on wafer-scale optical interconnects and the tuning behavior required to keep them operating correctly. According to the abstract, these tuning events recur often enough to create measurable pauses while communication is taking place. That makes the tuning mechanism itself part of the training cost, especially in workloads that rely on frequent expert routing and communication.
To address this, the authors evaluate a ferroelectric-based mitigation approach. The paper reports that eliminating the tuning stalls leads to a speedup of 2.7x for Mixtral 8x7B in the evaluated scenario. The result is presented as evidence that tuning overhead can dominate the performance of optical interconnects in this class of training system.
The technical paper, titled “Thermal Tuning Overhead in Wafer-Scale Optical Interconnects for LLM1. local linear models2. Large Language ModelThe model class referenced in the long-context serving demo. MoE Training: A Cross-Layer Analysis and Ferroelectric-Based Mitigation,” was published in August 2026 by Seongwon Yoon, Pin-Jun Chen, and Shimeng Yu. It ties the performance issue to thermal tuning specifically and proposes ferroelectric devices as the mitigation path studied in the work.