Georgia Tech researchers reported 2.7-fold speedups using ferroelectric tuning for wafer-scale optical interconnects in mixture-of-experts training.

Georgia Institute of Technology researchers published a technical paper examining thermal-tuning overhead in wafer-scale optical interconnects for large-language-model mixture-of-experts training. The analysis identified repeated tuning stalls during communication phases.

The researchers proposed ferroelectric-based tuning as the mitigation mechanism. Their paper reported speedups of 2.7 times in the studied setting, with the figure referring to the paper’s cross-layer analysis rather than a general performance claim for every mixture-of-experts workload.

The work specifically concerns tuning behavior in wafer-scale optical interconnects, not GPU arithmetic throughput or CPU performance. The source summary does not provide the tested model size, wafer dimensions, node count, power savings, or an independent reproduction of the result.