·NVIDIA

The Future is Optical: NVIDIA to Revolutionize AI Clusters with Light by 2026

NVIDIA has announced a groundbreaking shift for its future AI data center platforms. According to a report from Tom's Hardware, the company plans to fully embrace silicon photonics and Co-Packaged Optics (CPO) by 2026, replacing traditional electrical signals with light for high-speed interconnectivity between GPUs.

To address the ever-increasing communication demands of massive GPU clusters, NVIDIA is accelerating the adoption of optical interconnects. The company's next-generation rack-level AI platform, slated for a 2026 launch, will be built on silicon photonics and CPO. This innovative approach involves integrating optical engines directly alongside the switch chip, a design that promises significantly higher bandwidth and lower power consumption.

Compared to conventional pluggable optical modules, CPO technology offers substantial advantages. Beyond reducing the number of components and enhancing system reliability, it dramatically lowers signal loss from 22 decibels to just 4 decibels. Furthermore, it slashes per-port power consumption from 30 watts down to an efficient 9 watts.

The core benefit of this technology lies in avoiding the power consumption and signal degradation inherent in long-distance electrical transmission. In large-scale AI clusters, copper cables are no longer viable for the 800Gb/s speeds required, making fiber optic connections a necessity. While traditional optical modules involve long electrical paths before signal conversion, adding complexity and energy use, CPO performs the electrical-to-optical conversion almost instantaneously after the signal is generated.

Looking ahead, NVIDIA plans to launch its Quantum-X InfiniBand switch in early 2026. This liquid-cooled powerhouse will deliver a total bandwidth of 115Tb/s across 144 ports at 800Gb/s each. It will also feature 14.4 TFLOPS of in-network computing capabilities and the fourth-generation SHARP protocol to minimize latency, making it ideal for generative AI and other hyperscale clusters.

Following this, in the second half of 2026, the company is set to release its Spectrum-X photonic Ethernet platform. Based on the Spectrum-6 chip, this platform will include two liquid-cooled models: the SN6810, offering 102.4Tb/s with 128 ports, and the SN6800, providing a massive 409.6Tb/s with 512 ports.

In conclusion, NVIDIA's strategic pivot to optical interconnects marks a fundamental evolution in data center architecture. By overcoming the physical limitations of copper, this move is essential for building the next generation of powerful, efficient, and scalable AI systems required to drive the future of artificial intelligence.

Torna al blog

Related Posts

  • _{region} ( _{language} )

    • _{area}

    _{region}
    _{language}