NVIDIA shifts AI competition from computing performance to system interconnects, leveraging CPO and silicon photonics to break I/O bottlenecks. By integrating high-speed fabric with advanced packaging, it redefines data center efficiency and accelerates the optical supply chain transition.

Key Highlights

• I/O Bottleneck: AI performance now hinges on interconnect capacity rather than raw compute, as traditional copper wiring hits physical limits in power, reach, and signal integrity.
• CPO Necessity: Co-Packaged Optics shortens electrical paths to reduce SerDes power and latency, offering a practical path to sustainable high-bandwidth scaling beyond 1.6T.
• System-Level Fabric: By integrating silicon photonics and advanced switches, NVIDIA treats entire racks as a single, massive GPU, optimizing data flow from chip to backbone.
• Market Trajectory: Adoption will accelerate with future generations, shifting the supply chain focus toward silicon photonics and high-density fiber arrays as CPO penetration rises.

Table of Contents

1. TrendForce’s View
   • CPO Penetration Rate in Optical Modules
2. NVIDIA Fabric Architecture
   • NVIDIA Scale-Out Fabric
   • Differences between Pluggable Optical Modules and CPO
   • Overview of NVIDIA’s GPU Generations and Rack Interconnect Architectures
   • DWDM Photonics For Interconnects
   • Example – 200Gbps vs. 32Gbps circuitry
   • Optics on Interposer w/ DWDM
   • Optics on Interposer
   • Spectrum-X Photonics delivers 64x better signal integrity
   • Overview of NVIDIA’s Three-Layer Interconnect Architecture

< Total Pages: 13>