Compared with networks using traditional transceivers, Spectrum-X Ethernet silicon photonics technology delivers a 5x improvement in energy efficiency, a 5x increase in AI uptime, and 1.3x faster deployment time.
NVIDIA Spectrum-X Ethernet silicon photonics technology is now in full-scale production. The next-generation Spectrum-X switches are built using co-packaged optics (CPO) technology, enabling the NVIDIA Vera Rubin platform to scale horizontally within data centers and deploy AI factories across regions.
This platform achieved mass production through deep co-engineering collaboration with semiconductor and systems ecosystem partners in Taiwan, with Taiwan Semiconductor, SPIL, TFC, and Foxconn each making critical contributions across key layers of the process—from silicon photonics to full system integration:
Taiwan Semiconductor’s advanced silicon photonics manufacturing technology translates breakthrough designs into production-ready chips.SPIL’s chip-level packaging, assembly, and testing technologies integrate electrical and optical components with micron-level precision.
TFC’s laser chips, packaged into modules, deliver the reliability required for AI workloads operating 24/7 year-round.
Foxconn’s system assembly integrates the Spectrum-X CPO switch into a complete rack-scale networking platform.
NVIDIA AI Factory systems are unboxed, installed, and powered on within NVIDIA-owned and operated AI factories, validating end-to-end workflows prior to customer shipment.
Spectrum-X Ethernet silicon photonics technology exemplifies NVIDIA’s full-stack co-design approach. Compared to networks using traditional transceivers, Spectrum-X delivers 5x higher energy efficiency, 5x improvement in AI uptime, and 1.3x faster deployment.
With its simplified design that frees up more power for compute, NVIDIA’s co-packaged optics networking infrastructure underpins million-GPU-scale AI factories. Companies such as CoreWeave, Lambda, and Oracle Cloud Infrastructure have already adopted this technology.
Through large-scale deployment of CPO, NVIDIA has eliminated the power, reliability, and deployment time ceilings of optical interconnects—addressing one of the key constraints limiting AI cluster growth.
Editor/Lambor