NVIDIA (NVDA) Co-Packaged Silicon Photonics Switches for Gigawatt AI Factories Webinar summary

Keynote and infrastructure overview

• Emphasized the shift from single-processor computing to data center-scale AI supercomputers, where network architecture defines performance.

• Outlined four major infrastructures: NVLink Scale-Up, Spectrum-X Ethernet Scale-Out, BlueField DPU-based context memory storage, and Scale-Across for inter-data center connectivity.

• Spectrum-X Ethernet designed specifically for AI workloads, eliminating jitter and synchronizing GPU operations for optimal distributed computing.

• Highlighted the need for end-to-end infrastructure, focusing on RDMA and SuperNICs to control data flow and avoid network hotspots.

• Achieved 1.4x performance and 3x expert dispatch improvement for AI workloads by eliminating jitter and ensuring predictable, synchronized GPU communication.

Co-packaged optics technology and deployment

• Co-packaged optics move the optical engine into the switch package, reducing power consumption by up to 5x and increasing data center resiliency.

• Innovations include micro-ring modulators, high-power lasers, and advanced packaging with TSMC for reliability and mass production.

• Spectrum-X and Quantum-X switches with co-packaged optics support up to 409 Tb/s and 2,000 ports, enabling million-scale GPU AI factories.

• Liquid cooling and dense switch designs further optimize power and performance for large-scale AI deployments.

• Initial deployments with partners like CoreWeave, Lambda, and Texas Advanced Computing Center are set for this year, with broader rollouts in the second half.

Reliability, flexibility, and adoption considerations

• Co-packaged optics improve reliability by eliminating human touch and external transceiver handling, validated through rigorous testing.

• Technology supports both short and long-range connections within and across data center buildings, replacing a wide range of pluggable transceivers.

• Concerns about flexibility and pay-as-you-go models are addressed by optimizing switch utilization in AI supercomputers, reducing both CapEx and OpEx.

• Annual innovation cadence will drive larger radix switches, higher port densities, and further integration of liquid cooling and flexible rack designs.

• Spectrum-X Ethernet supports multiple operating systems and customizable designs for diverse customer requirements.