Integrated Photonics and Liquid Liquid Storage: NVIDIA Announces 800G Switches Spectrum-X and Quantum-X

The race to AI is changing the face of data centers: network infrastructure is becoming more and more complex in the pursuit of high throughput and low latency. This comes at the cost of increased power consumption to operate optical transceivers. That’s why NVIDIA has introduced a new generation of switches with integrated silicon photonics that should solve this problem, while also providing increased reliability and speed of network infrastructure deployment.

NVIDIA estimates that a traditional cloud data center for every 100,000 servers consumes 2.3 MW of energy to support the operation of optical transceivers, but in AI clusters, where each accelerator needs its own fast network connection, this value can reach 40 MW, i.e. up to 10% of the total energy consumption of the entire complex. It would be much more reasonable to spend this energy on computing rather than network infrastructure.

Source here and Next: NVIDIA

The new Spectrum-X and Quantum-X switches should solve this problem radically. They use new ASICs that combine a chip switch and photonic modules on a single substrate. This approach allows us to abandon several links of the traditional chain that are part of a classic optical transceiver. A modern high-speed transceiver includes eight lasers that consume about 10 W, and a DSP block that requires 20 W.

Integrated photonics allows you to get by with just two external lasers to ensure the operation of one 1.6 Tbit/s port. The lasers are connected in this scheme directly to the photonic module on board the new ASICs. The optical engine itself as part of the ASIC consumes only 7 W, and the laser requires another 2 W. The difference in energy consumption is at least threefold.

In addition, the simplified wiring helps improve reliability: NVIDIA claims a 63-fold improvement in signal integrity, which does not have to travel through multiple electrical connections from the ASIC to the transceiver and within the transceiver, and a ten-fold increase in overall network reliability. While a traditional design can lose 22 dB of signal along its electrical path, the photonic module design only loses 4 dB.

The new ASIC packaging scheme is quite complex: it implements detachable optical connectors that enable scenarios with different switch port configurations, with speeds from 200 to 800 Gbps. The flagship Spectrum SN6800 switch includes 512 800GbE ports with an aggregate switching speed of 409.6 Tbps. The SN6810 model is more compact, it offers 128 800GbE ports and switching up to 102.4 Tbps.

The Quantum-X series is currently represented by the Quantum 3450-LD model: 144 800G InfiniBand ports with a total throughput of 115 Tbps. The combination of high density and such speeds required the development and integration of a custom liquid cooling system. The new Quantum-X switches will be available in the second half of this year, and Spectrum-X in the second half of 2026.

NVIDIA’s proprietary optical engines use micro-ring modulators (MRMs), which became possible thanks to NVIDIA’s collaboration with TSMC in the packaging of the COUPE “multi-story” chips. In addition to TSMC, Browave, Coherent, Corning Incorporated, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric Industries, and TFC Communication participated in the creation of the new switches.

The advantages of the new scheme are especially significant on a large scale, at the level of hundreds of thousands of accelerators. The deployment time is reduced by 1.3 times, and the overall reliability of the network becomes an order of magnitude higher. However, for now we are talking only about switches – optical cables will be directly connected to their ports. However, the other end of the cable will still go to the transceiver servicing a separate accelerator or node. Also, there are no plans to transfer NVLink to “optics” yet, since it is still easier and more profitable to work with “copper” inside the node and NVL rack.