NeTS: Accelerating AI Workloads with Programmable Photonic Interconnects

This project develops networking algorithms and programmable photonic interconnects for accelerating artificial intelligence (AI) training and inference workloads. Modern AI systems increasingly rely on large number of graphical processing units (GPUs) that must exchange enormous amounts of data during training and inference. Existing electrical interconnects are becoming increasingly constrained by power consumption and bandwidth scaling, as AI models continue to grow in scale. This project investigates how programmable photonic interconnects can improve communication efficiency by dynamically adapting network connectivity to the communication structure of AI workloads. In particular, the project studies collective communication operations such as AllReduce and All-to-All, which are fundamental building blocks in modern distributed training systems. The research is organized into three technical thrusts. The first thrust studies collective communication algorithms for static photonic topologies and investigates whether new algorithms can reduce communication completion times beyond existing logarithmic-step approaches while accounting for realistic propagation delays and congestion effects. The second thrust focuses on programmable photonic interconnects with reconfigurable circuit switching. This thrust develops topology synthesis techniques, switching schedules, and optimization frameworks that explicitly incorporate photonic constraints such as reconfiguration delay, insertion loss, switch radix limitations, and multi-hop routing effects. The third thrust studies workload-level implications in Mixture-of-Experts models, including token routing and expert placement problems under photonic interconnects. The project combines theoretical analysis, optimization, packet-level simulation using NS3 and Astra-Sim, and hardware experimentation using an eight GPU testbed connected to a programmable photonic switching prototype. The project advances the foundations of collective communication and programmable interconnect design for future AI infrastructure. The resulting algorithms and system insights can improve communication efficiency, reduce energy consumption, and enable more scalable training and inference systems. The project will release software prototypes, benchmarks, simulation extensions, and datasets to support reproducible research in networking and distributed systems. Educational activities include integrating project outcomes into graduate networking courses and engaging undergraduate students in research through photonic switching experiments, collective communication benchmarking, and simulations. The project will additionally contribute to workforce development in networking, distributed systems, and artificial intelligence infrastructure by providing hands-on systems research opportunities involving programmable hardware, GPUs, and large-scale simulation environments. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2551372 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Vamsi Addanki | Institution: Purdue University, WEST LAFAYETTE, IN | Award Amount: $427,000 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2551372 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2551372.html