Home
Tech
The 1.6Tbps of Nvidia change to prop up the meteoric increase of the AI factories with millions of GPU, but will have to wait

The 1.6Tbps of Nvidia change to prop up the meteoric increase of the AI factories with millions of GPU, but will have to wait

Quantum-X and Spectrum-X switches of NVIDIA reduce the dependence of traditional optical transceptions
Silicon Photonics increases the efficiency of the Nvidia network by 3.5x
Quantum-X and Spectrum-X switches offer a yield up to 400 TBPS for Hipperscala factories

Nvidia is advancing its network technology by integrating copeshed optics (CPO) in its infiniband quantum and Spectrum Ethernet switch, a movement that is expected to reduce energy consumption and cost in AI data centers.

In its GTC 2025 event, Nvidia detailed its plans to implement silicon photonics, which will improve efficiency by reducing the need for traditional optical transceives.

Instead of trusting traditional transceptions connectable, Nvidia is integrating photonics directly into the ASIC of change, reducing the use of energy and minimizing signal loss. These advances benefit from Hipperscala, and could also improve small businesses with similar efficiency profits.

Nvidia seeks to reduce the power of the AI data center by more than 50%.

The Spectrum-X and Quantum-X switches in NVIDIA use silicon photonics to offer a higher bandwidth and lower energy consumption, which supports up to 1.6 terabitos per second (TBP) per port to efficiently connect millions of GPU.

Photonics Quantum-X and Spectrum-X switches offer configurations ranging from 128 ports to 800 Gbps to 512 ports to 800 Gbps, delivering total yields of up to 400 Tbps.

While the Ethernet Spectrum-X platform improves multipurpose hyperscale implementations, the Quantum-X Infiniband switches offer higher signal integrity and resilience, which makes them contenders for the best network switch.

“AI factories are a new class of data centers with extreme scale, and network infrastructure should be reinvented to maintain rhythm. When integrating silicon photonics direct Nvidida.

The new NVIDIA switches improve energy efficiency in a 3.5 factor, while reducing signal degradation. In a typical AI data center with 400,000 GPU, conventional network configurations require millions of optical transceptions, which consume significant power.

The NVIDIA approach reduces the total power of the 72 megawatt network to 21.6 megawatts, drastically improving sustainability. These profits could also improve commercial smartphones, allowing faster and more reliable connectivity.

While Nvidia is making the transition to optical networks, Copper remains relevant in specific configurations.

Systems such as GB200 NVL72 still use thousands of copper cables to link GPUs and CPU through NVLink 5, offering lower energy consumption at the grid level.

However, As Nvidia progresses to NVLink 6, copper limitations will become more evident, which reinforces the need for photonic solutions in the implementations of large -scale AI tools.

The new NVIDIA switches are established for its launch at the end of 202 and 2026. The first model, the infiniband quantum 3450-LD switch, which will be launched at the end of 2025, will provide 144 connectivity ports of 800 GB/sec and a total bandwidth of 115 TB/sec.

In 2026, the Spectrum SN6810 Ethernet Switch will debut with 128 ports at 800 GB/sec and an aggregate bandwidth of 102.4 TB/sec. A larger spectrum SN6800 will also arrive in 2026, with 512 ports of 800 GB/sec and a total yield of 409.6 TB/sec.

Through Nextplatform

You may also like...