NVIDIA Spectrum-X Silicon Photonic Chip: AI Data Center Energy Efficiency Drink About Three times

At the recent Hot Chips Conference, NVIDIA demonstrated Spectrum-X silicon photonic chips. This new platform is regarded as the "heart-saving" of AI data centers because it can effectively solve the bandwidth, delay and power bottles faced by traditional networks in super-large AI factories, and becomes the key foundation for future generations of data centers.

From copper line to optics, to silicon photons

In the past, most of the connections between servers and exchangers depend on copper lines. Although the cost is low, when transmitting high-speed signals of hundreds of Gb/s, the power was severe, energy consumption was high, and even additional digital signal processing chips (DSPs) were needed to supplement, which limited efficiency and stability. To improve these problems, the Traditional Data Center began introducing an optical module to convert the signal to the optical signal for long distance transmission.

However, traditional optical modules still need to go through PCB and plug-in receiver, and the signal paths and energy consumption are relatively high. The emergence of Silicon Photonics allows optical components to be directly integrated next to the ASIC packaging of the exchanger to form the so-called Co-Packaged Optics (CPO). This design greatly shortens the signal path, reduces power consumption and delays, and improves reliability.

(Source: NVIDIA)

Features of Spectrum-X

According to the NVIDIA blog, Spectrum-X Photonics uses 200G/lane SerDes, which can provide up to 409.6 Tb/s bandwidth, and a single machine supports up to 512 800G ports. Compared with traditional architecture, Spectrum-X can enable data centers to achieve 3x more energy efficiency improvements, 10x system reliability, and reduce single port power consumption from about 30W to 9W. The overall design is more power-efficient and reliable, and it is a continuity solution for AI factories that require high-density GPU sets.

In addition to Spectrum-X, NVIDIA has also simultaneously launched Quantum-X Photonics, a new generation exchange platform for InfiniBand. Although its switching capacity (115 Tb/s, 144 port 800G) is slightly lower than Spectrum-X, it has built-in SHARP v4 in-network computing capabilities, especially suitable for high-speed training of supercomputers and a single large model.

And Spectrum-X Photonics is scheduled to be launched in 2026, becoming the core technology to support the next-generation AI data center.

NVIDIA’s Spectrum-X Ethernet Photonics Debuts as the World’s First 200G Co-Packaged Optics, Supercharged the Future of AI