# NVIDIA Spectrum-6

> Source: https://aiwiki.ai/wiki/nvidia_spectrum_6
> Updated: 2026-07-25
> Categories: AI Hardware, AI Infrastructure, Data Centers, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

[NVIDIA](/wiki/nvidia) Spectrum-6 is an [Ethernet](/wiki/ethernet) switch ASIC and system architecture for large [AI infrastructure](/wiki/ai_infrastructure). A single Spectrum-6 chip has 102.4 terabits per second (Tb/s) of switching capacity, twice the 51.2 Tb/s of the preceding Spectrum-4 generation. NVIDIA uses the chip in its SN6000 family, ranging from single-chip 102.4 Tb/s switches to a four-chip, 409.6 Tb/s system.[1][2]

Spectrum-6 is not the name of only one switch. It is the silicon foundation for several pluggable-optics and co-packaged-optics systems, and it anchors the next generation of [NVIDIA Spectrum-X](/wiki/nvidia_spectrum_x). NVIDIA developed it alongside the [NVIDIA Vera Rubin](/wiki/nvidia_vera_rubin) compute platform, with the switch handling communication between racks and sites while NVLink handles tightly coupled communication inside a rack.[1][2]

## Role in the Vera Rubin platform

NVIDIA formally presented Spectrum-6 in January 2026 as one of the six original Vera Rubin chips. The other components were the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, and BlueField-4 DPU. This division gives Spectrum-6 a specific job: it provides the high-radix scale-out network that connects many rack-scale systems into a larger cluster.[3]

That role differs from [NVLink](/wiki/nvlink), which connects accelerators within a Vera Rubin NVL72 rack at very high bandwidth. For Ethernet scale-out, Spectrum-6 switches work with [ConnectX-9](/wiki/connectx) adapters at the endpoints. Together with NVIDIA's routing, congestion-control, telemetry, and management software, they form the Spectrum-X fabric. A Spectrum-6 switch can therefore be deployed as standards-based Layer 2 or Layer 3 Ethernet hardware, but NVIDIA's intended AI configuration uses coordinated behavior in both the switch and the network adapter.[1][4]

The target traffic also differs from ordinary enterprise traffic. Distributed training and inference create synchronized collective operations and bursty all-to-all exchanges. A slow or congested path can leave many accelerators idle. NVIDIA says Spectrum-X responds with adaptive routing, coordinated congestion control, hardware-assisted traffic isolation, and fine-grained telemetry. These are fabric-level functions, not extra raw bandwidth in the Spectrum-6 ASIC.[2]

## Switch silicon and interfaces

Each Spectrum-6 chip exposes 512 lanes running at 200 Gb/s with PAM4 signaling. Multiplying 512 by 200 Gb/s gives the advertised 102.4 Tb/s switching capacity. The lanes can be presented as 512 interfaces at 200 Gb/s or combined into higher-rate interfaces, including 128 at 800 Gb/s. NVIDIA's documentation describes leaf, spine, and superspine uses and full wire-speed forwarding.[2][4]

The 200G lane rate is also the main generational change from Spectrum-4, which used 100G SerDes and delivered 51.2 Tb/s per chip. It allows a Spectrum-6 system to double switching capacity without simply doubling the number of external connectors. In the pluggable SN6600 designs, one physical OSFP cage carries two logical 800 GbE interfaces. This is why the same hardware may be described as 64 physical 1.6 Tb/s cages or 128 logical 800 GbE ports.[4][5]

That wording matters in standards terms. IEEE 802.3df-2024 defines 800 GbE and permits an eight-lane interface to be divided into lower-rate ports. The follow-on IEEE P802.3dj project covers 200 Gb/s electrical and optical signaling and rates through 1.6 Tb/s.[6] NVIDIA's public SN6000 documentation advertises logical Ethernet interfaces through 800 GbE. The 1.6 Tb/s figure used for a twin-800G OSFP cage is its aggregate physical capacity, not evidence that every such cage is operating as one 1.6 TbE logical port.

## SN6000 systems

NVIDIA's detailed SN6000 hardware manual covers four Spectrum-6 systems. Its current Ethernet portfolio also lists the rack-integrated SN6200-LD. Published maximum configurations are summarized below.[4][5][7]

| System | Optical interface | Cooling and height | Spectrum-6 chips | Switching capacity | Published interface configuration |
| --- | --- | --- | ---: | ---: | --- |
| SN6600-LD | 64 twin-800G OSFP cages, pluggable | Liquid, 2U | 1 | 102.4 Tb/s | 128 x 800 GbE, with 400/200/100 GbE breakout options |
| SN6600 | 64 twin-800G OSFP cages, pluggable | Air, 3U | 1 | 102.4 Tb/s | 128 x 800 GbE, with 400/200/100 GbE breakout options |
| SN6810-LD | 128 MMC-12 connectors, co-packaged optics | Liquid, 2U | 1 | 102.4 Tb/s | 128 x 800 GbE or 512 x 200 GbE |
| SN6800-LD | 512 MMC-12 connectors, co-packaged optics | Liquid, 5U | 4 | 409.6 Tb/s | 512 x 800 GbE or 2,048 x 200 GbE in NVIDIA's technical overview |
| SN6200-LD | 32 twin-800G OSFP front cages plus rack backplane links | 1U | 1 | 102.4 Tb/s | Front-facing ports plus 256 x 200G backplane connections |

The first four systems show that optics and cooling are configuration choices rather than fixed properties of the ASIC. SN6600 provides an air-cooled, pluggable option. SN6600-LD keeps pluggable modules but moves to liquid cooling and a denser chassis. SN6810-LD and SN6800-LD integrate silicon photonics with the switch package and use liquid cooling. The SN6800-LD combines four Spectrum-6 ASICs and internally rearranges their lanes to reach 409.6 Tb/s.[4][5]

Published power figures also vary substantially by form factor. NVIDIA lists typical power of 1.96 kW and maximum power of 2.2 kW for the single-chip CPO SN6810-LD. The four-chip SN6800-LD is listed at 8.3 kW typical and 10.5 kW maximum. For the pluggable SN6600 systems, consumption depends on the type of retimed optics installed, so a chassis-only comparison would be misleading.[5]

## Pluggable optics and co-packaged optics

Spectrum-6 supports both conventional pluggable optical modules and co-packaged optics (CPO). In a pluggable system, high-speed electrical signals travel from the switch ASIC across the circuit board to a replaceable optical module. CPO moves the electrical-to-optical conversion next to the switch ASIC. The Spectrum-6 CPO package described by NVIDIA contains 32 silicon-photonic engines at 3.2 Tb/s each, with micro-ring modulators, detachable fiber connectors, and external laser arrays.[2]

NVIDIA says this arrangement removes DSP retimers from the optical path and reduces the number of active components. The company reports that optical loss falls from about 22 decibels to about 4 decibels and describes the result as up to 64 times better signal integrity. Its current photonics material also claims about five times better network power efficiency than pluggable transceivers and five times longer sustained AI application runtime. The July 2026 Spectrum-6 announcement separately claims a tenfold improvement in mean time between incidents.[1][8]

Those figures are vendor comparisons for a CPO network, not measured guarantees for every Spectrum-6 switch. They also do not apply to the pluggable SN6600 models merely because those systems use the same ASIC. CPO trades the field-replaceable module model for tighter integration, different fiber connectors, external lasers, and liquid-cooled system designs. Independent reporting before launch similarly treated CPO as a response to the electrical loss and power cost of moving 200G signals across a board, while noting that NVIDIA's Ethernet CPO systems were scheduled for the second half of 2026.[9]

NVIDIA first announced the associated [Spectrum-X Photonics](/wiki/spectrum_x_photonics) roadmap in March 2025. At that time it used rounded figures of 100 Tb/s and 400 Tb/s for the two CPO system sizes and projected availability in 2026. Later technical documentation gives the exact 102.4 Tb/s and 409.6 Tb/s values. The earlier announcement also claimed 3.5 times better power efficiency, while 2026 material uses a five-times figure, illustrating why the date and comparison baseline matter when citing these claims.[8][10]

## Fabric software and operation

The SN6000 hardware can run NVIDIA Cumulus Linux, which the hardware manual lists as the preinstalled option. SONiC is also offered as an open-source network operating system, although NVIDIA cautions that it may not be supported on every switch. Firmware updates are delivered through the management software.[11]

Across the broader Spectrum portfolio, NVIDIA offers NetQ for visibility, troubleshooting, and lifecycle management, and DSX Air for simulating network designs before deployment. The portfolio also advertises RDMA over Converged Ethernet (RoCE). In a Spectrum-X deployment, the important distinction is that the switch is not expected to make routing decisions in isolation. Spectrum-6 and ConnectX-9 exchange telemetry and coordinate path selection and congestion response.[1][7]

A 2026 paper by NVIDIA researchers describes the Spectrum-X design as a multiplane network that replaces additional hierarchical depth with parallel paths. It places hardware-accelerated load balancing in both network adapters and switches so the fabric can react at microsecond time scales. The reported evaluation reached 98% of theoretical line rate and saw a 7% latency increase when 10% of fabric links failed.[12] These results support the design rationale for Spectrum-X, but the paper does not identify Spectrum-6 as the switch generation under test, so they should not be treated as Spectrum-6-specific benchmarks.

NVIDIA makes broader platform claims for the current generation. It says Spectrum-X can deliver up to 1.6 times the AI networking performance of off-the-shelf Ethernet, sustain up to 95% network efficiency in clusters larger than 100,000 GPUs, and use 1.7 times fewer switches with a multiplane topology.[1] Public material does not make those figures a universal per-switch result, and actual performance will depend on topology, workload, endpoint configuration, software version, and failure conditions.

For facilities separated by longer distances, NVIDIA positions [Spectrum-XGS](/wiki/nvidia_spectrum_xgs) as the scale-across layer. It adds distance-aware congestion control and multi-site telemetry to the Spectrum-X design.[2] This capability belongs to the wider networking platform; the 102.4 Tb/s Spectrum-6 switch capacity by itself does not determine wide-area performance.

## Availability and adoption

The product's rollout spans several announcements. NVIDIA disclosed its CPO switch roadmap in March 2025, formally introduced Spectrum-6 with Vera Rubin in January 2026, and said Rubin partner systems would become available in the second half of 2026.[3][10] On July 21, NVIDIA said Spectrum-6 was arriving in large AI deployments and named [CoreWeave](/wiki/coreweave), [Microsoft](/wiki/microsoft), [Nebius](/wiki/nebius), SpaceXAI, and Tesla as early adopters. It specified that CoreWeave, Microsoft, and Nebius would be among the first providers to deploy Vera Rubin infrastructure with Spectrum-6.[1] Independent trade reporting repeated the launch and adopter list, but it drew those details from NVIDIA rather than documenting five separate deployments.[13]

The strongest public deployment confirmation came from CoreWeave on the same day. The cloud provider said it had deployed liquid-cooled SN6600-LD switches for Vera Rubin NVL72. CoreWeave describes each switch as 102.4 Tb/s across 64 physical twin-800G OSFP cages and its fabric as nonblocking, multiplane, and multirail. It reported fitting 16 switches in a 48U rack, for 1.64 petabits per second of aggregate switching capacity.[14] The account confirms a specific Spectrum-6 system in operation, although its utilization and efficiency comparisons are CoreWeave's own measurements.

The evidence for the other named adopters is less specific. Microsoft said in March 2026 that it had powered on a Vera Rubin NVL72 system in its labs and planned to roll the platform into liquid-cooled Azure data centers over the following months, but that announcement did not identify Spectrum-6.[15] NVIDIA included a Nebius executive statement in its July article. No separate public Spectrum-6 deployment statement from SpaceXAI or Tesla was identified in the sources used here. The five-company list is therefore best read as NVIDIA's adoption announcement, not as proof that all five had production switches serving customers on the announcement date.

## Assessment

Spectrum-6 doubles NVIDIA's per-chip Ethernet switching capacity and gives the company one silicon base for air-cooled pluggable, liquid-cooled pluggable, and liquid-cooled CPO systems. Its importance to Vera Rubin comes from the complete Spectrum-X design: switch silicon, endpoint adapters, topology, telemetry, congestion control, and operations software working together. Raw capacity alone does not establish application performance.

As of July 2026, the public record includes detailed hardware manuals and a direct CoreWeave deployment account, so Spectrum-6 is beyond a paper roadmap. Important limits remain. NVIDIA has not published pricing, broad independent production benchmarks, or equally detailed deployment evidence for every named adopter. Several power, reliability, efficiency, and switch-count comparisons are NVIDIA or partner claims with specific, sometimes unstated baselines. They are useful indicators of design goals, but independent measurements across different network topologies are still needed.

## References

1. NVIDIA, "Built for Vera Rubin, NVIDIA Spectrum-6 Arrives in Gigascale AI Factories," July 21, 2026. https://blogs.nvidia.com/blog/nvidia-spectrum-six-arrives-in-gigascale-ai-factories/
2. NVIDIA Technical Blog, "Inside the NVIDIA Vera Rubin Platform: Six New Chips, One AI Supercomputer," January 5, 2026, updated March 16, 2026. https://developer.nvidia.com/blog/inside-the-nvidia-rubin-platform-six-new-chips-one-ai-supercomputer/
3. NVIDIA Newsroom, "NVIDIA Kicks Off the Next Generation of AI With Rubin," January 5, 2026. https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer
4. NVIDIA, "NVIDIA Spectrum-6 SN6000 Switch Systems Hardware User Manual: Introduction," updated July 20, 2026. https://networking-docs.nvidia.com/sn6000hw/introduction
5. NVIDIA, "NVIDIA Spectrum-6 SN6000 Switch Systems Hardware User Manual: Specifications," updated July 1, 2026. https://networking-docs.nvidia.com/sn6000hw/specifications
6. IEEE Standards Association, "Ethernet's Next Bar is Now - 800 Gb/s!" April 23, 2024. https://standards.ieee.org/beyond-standards/ethernets-next-bar/
7. NVIDIA, "Accelerated Ethernet Switching for AI and the Cloud," accessed July 25, 2026. https://www.nvidia.com/en-us/networking/ethernet-switching/
8. NVIDIA, "NVIDIA Silicon Photonics," accessed July 25, 2026. https://www.nvidia.com/en-us/networking/products/silicon-photonics/
9. Tom's Hardware, "Nvidia outlines plans for using light for communication between AI GPUs by 2026," August 2025. https://www.tomshardware.com/networking/nvidia-outlines-plans-for-using-light-for-communication-between-ai-gpus-by-2026-silicon-photonics-and-co-packaged-optics-may-become-mandatory-for-next-gen-ai-data-centers
10. NVIDIA Newsroom, "NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUs," March 18, 2025. https://nvidianews.nvidia.com/news/nvidia-spectrum-x-co-packaged-optics-networking-switches-ai-factories
11. NVIDIA, "NVIDIA Spectrum-6 SN6000 Switch Systems Hardware User Manual: Software Management," updated April 26, 2026. https://networking-docs.nvidia.com/sn6000hw/software-management
12. Sajy Khashab et al., "High-speed Networking for Giga-Scale AI Factories," arXiv:2605.21187, May 20, 2026. https://arxiv.org/abs/2605.21187
13. SDxCentral, "Nvidia debuts Spectrum-6 switches to power next-gen AI networking," July 21, 2026. https://www.sdxcentral.com/news/nvidia-debuts-spectrum-6-switches-to-power-next-gen-ai-networking/
14. CoreWeave, "Liquid-Cooled Switching Doubles AI Network Bandwidth Per Rack," July 21, 2026. https://www.coreweave.com/blog/liquid-cooled-switching-doubles-ai-network-bandwidth-per-rack
15. Microsoft, "Microsoft at NVIDIA GTC: New solutions for Microsoft Foundry, Azure AI infrastructure and Physical AI," March 16, 2026. https://blogs.microsoft.com/blog/2026/03/16/microsoft-at-nvidia-gtc-new-solutions-for-microsoft-foundry-azure-ai-infrastructure-and-physical-ai/