Nvidia unveiled its Spectrum-X Photonics and Quantum-X Photonics network Nvidia new Silicon Photonics based 400 tb/s
Nvidia unveiled its Spectrum-X Photonics and Quantum-X Photonics network switch platforms at GTC 2025 that will be used in exascale data centers based on silicon photonics. The new platforms boost data transfer rates to 1.6 Tb/s per port and up to 400 Tb/s total, enabling millions of GPUs to work together efficiently. Nvidia claims the switches provide higher bandwidth, less power loss, and greater reliability than traditional network technologies.
The Spectrum-X Photonics InfiniBand and Ethernet platforms deliver 1.6 Tb/s per port, twice the maximum rate of current high-end copper Ethernet solutions, with a total bandwidth of up to 400 Tb/s across different port combinations.
The Spectrum-X Photonics switches come in several combinations ranging from 128 ports at 800 Gb/s or 512 ports at 200 Gb/s, delivering a total bandwidth of 100 Tb/s. An even bigger variant offers 512 ports at 800 Gb/s or 2,048 ports at 200 Gb/s, delivering 400 Tb/s of bandwidth. The Quantum-X Photonics series offers 144 ports at 800 Gb/s InfiniBand, employing 200 Gb/s SerDes to deliver maximum data transfer.
Quantum-X delivers twice the performance of previous network solution generations and increases AI compute scalability by five times, and is capable of meeting demanding workloads and constructing larger AI clusters.
Quantum-X InfiniBand switches incorporate a liquid cooling system to maintain onboard silicon photonics chips in efficient operation without overheating. The new network platforms therefore deliver 3.5 times better energy efficiency, 10 times better network reliability, and 63 times better signal integrity, reducing power consumption and enhancing long-term performance. Additionally, the speed of deployment is increased by 1.3 times, which makes the switches a preferable solution for hyperscale AI data centers, Nvidia stated.
Nvidia’s Spectrum-X Photonics Ethernet and Quantum-X Photonics InfiniBand solutions leverage TSMC’s silicon photonics solution, the Compact Universal Photonic Engine (COUPE), where a 65nm electronic integrated circuit (EIC) is integrated with a photonic integrated circuit (PIC) using SoIC-X packaging technology. Nvidia has also teamed with Coherent, Corning, Foxconn, Lumentum, Senko, and other partners to create a dedicated silicon photonics ecosystem with a stable supply chain so that AI clusters and data centers that were previously not possible with proprietary hardware could be constructed.
Nvidia is set to release the Quantum-X InfiniBand switches later in 2025, followed by the Spectrum-X Photonics Ethernet switches in 2026.
Though these innovations have a lot of potential, integrating silicon photonics on this level is not without its challenges, and widespread adoption will mean companies will have to replace existing networking hardware. Despite this, Nvidia’s silicon photonics is a huge advancement in AI networking. While Nvidia has not yet spoken about future plans for its network hardware, TSMC’s COUPE has a promising future roadmap, which Nvidia could choose to copy. Read more
