AI communication revolution: Nvidia aims to utilize light by 2026 for data transfer among AI GPUs through silicon photonics and co-packaged optics in future AI data centers
In a significant move towards advancing data center technology, Nvidia has announced plans to introduce CPO-based optical interconnection platforms for both Ethernet and InfiniBand technologies. This development is set to revolutionize the way data is transmitted within AI data centers, aiming for higher transfer rates at lower power consumption.
The company's next-generation platforms will utilise silicon photonics interconnects with co-packaged optics (CPO), a technology that embeds the optical conversion engine alongside the switch ASIC, thereby sidestepping penalties associated with traditional pluggable optical modules.
The Spectrum-X Platform
The Spectrum-X platform, a key component of Nvidia's next-generation AI platforms, will be powered by the Spectrum-6 ASIC. This platform will support two devices: the SN6810 and SN6800. The SN6810 offers an impressive 102.4 Tb/s of bandwidth with 128 ports operating at 800 Gb/s each, while the SN6800 scales to 409.6 Tb/s and 512 ports at the same rate.
Both Spectrum-X devices will use liquid cooling, a feature that promises to keep the systems running efficiently even under heavy loads.
The Quantum-X InfiniBand Switches
Nvidia also plans to introduce Quantum-X InfiniBand switches in early 2026. These switches will deliver an astounding 115 Tb/s of throughput and support 144 ports operating at 800 Gb/s each.
TSMC's COUPE Roadmap
Nvidia's roadmap for CPO appears to closely follow TSMC's COUPE roadmap. The first generation of TSMC's COUPE roadmap offers 1.6 Tb/s data transfer while reducing power consumption. The second generation moves into CoWoS packaging with co-packaged optics, enabling 6.4 Tb/s at the motherboard level. The third generation aims for 12.8 Tb/s within processor packages and targets further cuts in power and latency.
AI Clusters and Benefits
AI clusters using Quantum-X InfiniBand and Spectrum-X Photonics will eliminate thousands of discrete components, resulting in faster installation, easier servicing, and reduced power consumption per connection. These clusters are intended for powering new AI clusters for generative AI applications.
Competitive Landscape
Notably, AMD has recently acquired Enosemi, a move that could be seen as a response to Nvidia's optical interconnect advancements. It is worth mentioning that companies like Intel, Nvidia, and Cisco are actively developing co-packaged optics solutions, and startups like Celestial AI have raised significant funding related to optical interconnects for AI.
However, the development order for optical interconnects for Nvidia's next-generation AI platforms is not explicitly mentioned in the search results.
Power Savings and Improved Metrics
Nvidia's CPO-based switches are expected to improve metrics like time-to-turn-on, time-to-first-token, and long-term reliability. CPO reduces electrical loss to 4 decibels and per-port power consumption to 9W, making it a promising technology for the future of AI data centers.
In 2026, Nvidia will introduce the Spectrum-X Photonics platform for Ethernet in the second half of the year. Meanwhile, the next-generation Quantum-X and Spectrum-X photonics interconnection solutions are slated for release in 2026.
As Nvidia moves towards using light for communication in its next-generation AI platforms, the future of AI data centers is set to become faster, more efficient, and more reliable.
Read also:
- Understanding Hemorrhagic Gastroenteritis: Key Facts
- Expanded Community Health Involvement by CK Birla Hospitals, Jaipur, Maintained Through Consistent Outreach Programs Across Rajasthan
- Abdominal Fat Accumulation: Causes and Strategies for Reduction
- Deepwater Horizon Oil Spill of 2010 Declared Cleansed in 2024?
