With AI model size doubling every four months, the size of AI compute clusters is growing rapidly to keep up. Traditional data center front-end networks rely on optical interconnect for data movement with Terabits per second of connectivity and have been doubling data rate every 4 years. As the models grow it is also expected that the back-end GPU networks expand beyond the confines of the rack. Optical interconnects will be needed at distances of tens of meters with 10Tb/s data transfer rates and doubling every 2 years will be expected. How should optical interconnects best achieve this bandwidth increase and what is the best performance axis to push next? In this panel discussion we pose the question - “Is it better to scale with more wavelengths or by adding more fibers?” Join us to see if any of the panelists can convince you they have the right approach.

The Optoelectronic Technology Consortium (OETC) 16 Gbit/s, 32-channel parallel fiber optic bus introduces new requirements for modeling and simulation of an optical bus which were not present for previous single-channel optical link systems. These requirements include the simulation of the statistical variation in component parameters, timing skew, and crosstalk. Due to the required simulation of statistical variation in component parameters, a time efficient simulation approach such as the quasi-analytical simulation methodology is vital. In addition, the adoption of a block-oriented simulation environment facilitates mixed-level simulation which minimizes the computational requirements of the bus simulation while achieving the required accuracy in the simulation of all bus components. Simulation results indicate that while noise is not a limiting factor in the performance of the optical bus, timing skew, and device uniformity across the channels of the bus are the most important factors for satisfactory performance. Experimental measurements confirm these simulation results.