Key Takeaways
1. Nvidia launched Spectrum-X Photonics Ethernet and Quantum-X Photonics InfiniBand switches to enhance data center connectivity using light technology, targeting AI factories and reducing energy consumption.
2. Scaling up data center performance is challenging due to signal loss in copper connections, which can reach around 22 decibels, increasing power requirements and failure risks.
3. Co-packaged optics (CPO) improve efficiency by placing optical engines next to switch chips, reducing electrical loss to about four decibels and power usage per port to around nine watts, leading to significant gains in energy efficiency and signal quality.
4. Spectrum-X Photonics offers high bandwidth for Ethernet networks, achieving up to 400 Tb/s in large configurations, while Quantum-X Photonics focuses on 800 Gb/s InfiniBand connections, featuring liquid cooling and in-network computing capabilities.
5. Nvidia’s future plans include advancing optical technology through stages, aiming for higher speeds and lower latencies, and collaborating with various companies to streamline manufacturing and enable scalable networks with millions of GPUs.
Nvidia is on a mission to change how data centers connect by using light technology. In March 2025, they launched the Spectrum-X Photonics Ethernet and Quantum-X Photonics InfiniBand switches. These devices are made to link extensive “AI factories” located in various areas and support millions of GPUs while cutting down energy consumption and expenses. The main goal is to combine optical engines with switch chips, which helps in removing unnecessary electrical components.
The Challenge of Scaling Up
Scaling up is quite a big challenge. Once speeds reach 800 gigabits per second or higher, copper connections between servers and switches start to hinder performance. Signals weaken as they make their way through boards and connectors. This loss occurs even before the signal gets to the optical module. Nvidia mentions that this loss is around 22 decibels on 200-gigabit channels. As a result, more power is required, with each port consuming roughly 30 watts. Additional components also heighten the chances of failures.
Innovations with Co-packaged Optics
Co-packaged optics, or CPO, transforms this arrangement. By positioning the optical engine right next to the switch chip, signals are sent to the fiber almost instantly. This setup minimizes electrical loss to about four decibels and cuts down power use per port to around nine watts. Nvidia claims that, when scaled up, this method delivers roughly 3.5 times better energy efficiency, over 60 times improved signal quality, 10 times greater resilience due to fewer active components, and around 30 percent quicker setup time, since there’s less to construct and keep running.
Spectrum-X Photonics for Ethernet
For Ethernet applications, Spectrum-X Photonics is targeted at large, multi-tenant networks. Nvidia states it provides about 1.6 times more bandwidth per area than traditional Ethernet. The options include configurations such as 128 ports at 800 Gb/s or 512 ports at 200 Gb/s, achieving a total of 100 Tb/s. Larger configurations can expand to 512 ports at 800 Gb/s or up to 2,048 ports at 200 Gb/s, reaching a total of 400 Tb/s.
Quantum-X Photonics for InfiniBand
Focusing on InfiniBand, Quantum-X Photonics centers around 800 Gb/s connections and employs liquid cooling, which uses liquid coolant to dissipate heat. Its leading switch features 144 ports and can manage 115 Tb/s of data. It also incorporates in-network computing, allowing data processing within the network, rated at 14.4 trillion floating-point operations per second. The system utilizes Nvidia’s latest SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) technology to speed up group tasks across the network.
Future Plans and Partnerships
Nvidia asserts that this new generation is twice as fast and five times more scalable for AI networks compared to the last one. The scalability improvement is closely tied to TSMC’s COUPE platform and modern packaging techniques. In the first stage, optical engines in OSFP modules will achieve 1.6 Tb/s. The second stage will introduce co-packaged optics on the motherboard, with capabilities of 6.4 Tb/s. The third stage aims for 12.8 Tb/s within processor packages, which will further minimize latency and energy usage. Nvidia plans to roll out CPO-based Quantum-X switches in early 2026 and Spectrum-X Photonics later that year, both featuring liquid cooling.
Nvidia is collaborating with several companies, including TSMC, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric, TFC, and others for manufacturing, optics, and assembly. The aim is to eliminate thousands of individual components from large clusters, accelerate setup, and enable networks with a million GPUs without excessive power usage.
Source:
Link
Leave a Reply