NVIDIA Architecture Shift Optical Interconnects to Replace Copper in AI Scaling
NVIDIA presents its upcoming AI computing rack design which establishes a new standard by employing optical interconnects to replace copper data links which have reached their maximum capability. Recent research from TrendForce shows that GPU developers are now designing their future chips with increased interconnect density to achieve the extensive data transfer requirements of their growing AI cluster systems. The new system works for both internal rack expansion and external rack extension purposes.
The NVIDIA NVLink 6 protocol defines a peak transmission rate of 400G SerDes per lane which enables GPUs to reach a maximum data throughput of 3.6 TB/s. Copper cables start to lose their electrical signal quality when they reach distances that exceed their maximum operating range. Current data shows that copper links can work at these frequencies only for distances shorter than one meter.
Broadcom together with other industry leaders predicts that copper will continue to dominate ultra short reach connections in single racks until 2028 because of its lower costs. The only method to achieve optimal performance with sustained bandwidth at 576 GPU clusters spanning multiple racks requires the use of optical transmission.
Optical transmission requires Wavelength Division Multiplexing WDM technology because it needs to send multiple signals through a single fiber at a density level which copper cannot reach. TrendForce predicts a steady rise in Co Packaged Optics CPO penetration within AI data centers
- 2026 Forecast CPO is expected to account for roughly 0.5% of optical transceiver modules
- 2030 Forecast Market penetration is projected to reach approximately 35%
- Adoption Path The Rubin generation will start with inter rack transmission of scale out systems while later generations will use it for scale up systems
NVIDIA has maintained its presence in the optical sector through a two billion dollar investment which it divided between Lumentum and Coherent. The two parties established multiple year contracts which provide them with priority access to high tech laser and optical components. The company uses TSMC’s COUPE 3D packaging technology to advance its development of silicon photonics systems. The process includes
- 3D Stacking The method decreases space requirements by stacking logic and photonics chips
- Micro Ring Modulators The silicon photonics die receives 200G PAM4 MRM integration which boosts bandwidth density
- Efficiency The system lowers power expenses while achieving maximum data rate transmission through its optical engine
Optical technology now counts as a necessary component which high scale AI infrastructure depends upon to function correctly. The future standard inter rack connections for computing clusters will use silicon photonics and CPO packaging technology once both systems reach their complete development. The new system enables data transmission to synchronize with the rapid increase of GPU processing power which establishes the foundation for widespread optical I/O implementation and advanced interconnect systems that will dominate the next ten years.
