Vera Rubin scale-out Ethernet
NVIDIA Spectrum-6: 102.4 Tb/s AI Ethernet Switch Guide
NVIDIA Spectrum-6 is the switch silicon at the center of the newest Spectrum-X Ethernet generation for Vera Rubin AI factories. Its headline 102.4 Tb/s switching capacity is double Spectrum-4, but the important planning change is broader than one throughput number: 200G PAM4 SerDes, 128 ports at 800GbE on SN6000-class systems and tighter coordination with ConnectX-9 allow flatter fabrics to carry much larger accelerator clusters.
Interactive calculator
Spectrum-6 Fabric Capacity Calculator
Enter your own topology and bandwidth assumptions. These planning calculators estimate raw links, capacity and ratios; they do not certify fabric goodput, rail mapping, cable reach, firmware interoperability, electrical design, cooling or OEM compatibility.
Live Amazon supporting hardware
Current supporting components and Price Options
Compare current Amazon listings relevant to this guide. Product availability and prices can change.
Quick answer
What does Spectrum-6 change?
Spectrum-6 provides 102.4 Tb/s of switch-chip bandwidth, with current SN6000 systems exposing up to 128 x 800GbE ports. NVIDIA pairs it with 200G PAM4 signaling and ConnectX-9 SuperNICs in the Vera Rubin generation. Compared with Spectrum-4 at 51.2 Tb/s, the new ASIC doubles switching capacity and doubles 800GbE port radix per switch chip.
Size from endpoint bandwidth and topology, not ASIC throughput alone
A 102.4 Tb/s switch can still be the wrong choice if the chosen lane breakout, cable reach, oversubscription target, redundancy plan or software stack does not match the servers. Use aggregate accelerator bandwidth to estimate port demand, then validate the exact SN6000 platform, optics, cooling, power and NOS before treating the calculator result as a deployable bill of materials.
Spectrum-6 at a glance
The table separates switch-chip capability from system-level choices so the numbers are easier to reuse in procurement and AI answers.
| Attribute | Spectrum-6 / SN6000-class | Why it matters | Planning note |
|---|---|---|---|
| Switch-chip bandwidth | 102.4 Tb/s | Twice Spectrum-4 chip capacity | Do not confuse chip bandwidth with application goodput |
| 800GbE port radix | Up to 128 ports | More endpoints per switch | Breakouts can trade port speed for count |
| SerDes generation | 200G PAM4 | Higher bandwidth per electrical lane | Cabling and optics must match the lane design |
| Vera Rubin pairing | ConnectX-9 + Spectrum-X | End-to-end co-design | Qualification matters more than connector shape |
| Primary role | AI scale-out Ethernet | Connects rack-scale systems across the cluster | NVLink remains the scale-up fabric inside supported racks |
| Topology direction | Flat / multiplane capable | Avoids unnecessary tiers at very large scale | Exact design depends on GPU count and rail strategy |
Before you use the result for procurement
Draw the physical topology
Map every server-facing port, leaf uplink, spine link and network plane. Aggregate bandwidth alone can hide impossible port or lane assumptions.
Verify exact endpoints
Confirm NIC form factor, PCIe generation, host lane budget, port speed, connector and firmware support on the exact server platform.
Qualify optics and cables
Match OSFP/QSFP form factor, lane rate, breakout, reach, fiber type and both endpoint qualification lists. Do not treat equal headline speed as automatic compatibility.
Test failure and congestion behavior
Validate oversubscription, ECMP or multiplane path behavior, switch failure domains and recovery under the traffic patterns the AI workload will actually generate.
Spectrum-6 is a 102.4 Tb/s switch ASIC
The most useful starting point is the aggregate switching figure. NVIDIA specifies 102.4 terabits per second for Spectrum-6, which is twice the 51.2 Tb/s ceiling of Spectrum-4. That doubling gives fabric designers more room to preserve high endpoint bandwidth while reducing the number of switch stages or devices required for the same accelerator population.
Aggregate throughput is not the same as workload throughput. Collective communication patterns, packet sizes, congestion, routing policy, port breakout and failure recovery determine how much of the theoretical fabric is converted into useful application progress. Treat 102.4 Tb/s as a physical switching budget, then model utilization separately.
The port-radix jump changes two-tier designs
Current NVIDIA SN6000 documentation lists systems with 128 ports of 800GbE around a Spectrum-6 ASIC. Spectrum-4 SN5600-class systems expose 64 ports of 800GbE. Doubling the high-speed radix can materially reduce leaf and spine counts because each switch can terminate more server-facing or inter-switch links before another network layer becomes necessary.
Radix is especially important for AI because expensive accelerators often need large, symmetric fabrics. When a switch can serve more endpoints without sacrificing link speed, operators can build shallower topologies with fewer optics and cables. The calculator therefore estimates both raw bandwidth and port pressure rather than looking only at total terabits.
200G PAM4 SerDes is the electrical foundation
NVIDIA describes Spectrum-6 with 200G PAM4 SerDes, compared with 100G PAM4 in the Spectrum-4 generation. A faster electrical lane lets the platform assemble 800G ports with fewer or faster lanes and supports the newer optics architecture used throughout Vera Rubin networking.
SerDes generation should be treated as a compatibility boundary. An optic labeled 800G is not automatically correct for every port or switch. Lane count, host electrical interface, optical modulation, connector form factor and firmware qualification still need to match the exact switch and SuperNIC combination.
Spectrum-6 belongs to Spectrum-X, not generic Ethernet alone
The switch is delivered as part of NVIDIA Spectrum-X Ethernet, where switch silicon, SuperNICs and software collaborate on adaptive routing, congestion control and telemetry. This is why a comparison against an unrelated 102.4T Ethernet ASIC should go beyond the headline capacity number.
For buyers, the practical question is whether the deployment intends to use NVIDIA’s end-to-end AI Ethernet behavior or merely needs high-radix standards-based switching. The answer affects NIC selection, software, operational tooling, qualification and the value of features such as hardware plane load balancing.
Vera Rubin uses ConnectX-9 at the endpoint
NVIDIA positions ConnectX-9 as the endpoint companion for the Spectrum-6 generation. Current documentation specifies up to 800GbE per port and PCIe Gen6 x16 for the 800G OSFP cards, while Vera Rubin platform material describes up to 1.6 Tb/s of scale-out throughput per GPU in the designed configuration.
That 1.6 Tb/s figure should not be interpreted as one ordinary 800G cable magically doubling its rate. It reflects the platform’s endpoint arrangement. When planning a generic server, use the physical NIC and port configuration actually installed rather than copying a rack-scale Rubin number into a different architecture.
Spectrum-6 can reduce switch count at the same endpoint load
A higher-radix switch can collapse some designs that previously needed more leaf or spine boxes. Fewer switches can mean fewer management points, optics, cables, rack units and failure domains, although those savings depend on topology and redundancy choices.
The effect is strongest when a deployment is port-limited rather than bandwidth-limited. If each server consumes multiple ports or the network is split into independent planes, total switch count can still be high. Count physical ports and lanes before translating a switch-throughput ratio into a procurement ratio.
Power and cooling move with the switch system, not just the ASIC
SN6000 hardware exists in different system configurations, including air-cooled and liquid-cooled designs. NVIDIA’s published power figures vary by chassis and optical implementation. A rack planner therefore needs the exact switch SKU and optics population before reserving electrical capacity or cooling.
This is particularly important for co-packaged optical designs because the location of optics changes the service and thermal model. A data center that can power a conventional switch may still lack the liquid loop, busbar arrangement or maintenance workflow required by a particular Spectrum-6 system.
Cabling strategy should be selected with the topology
Short in-rack and adjacent-rack links may be candidates for passive copper or active electrical options when the platform supports them. Longer runs move toward optical links, and multiplane fabrics can multiply the number of physical paths even when applications see one logical interface.
Use cable length, connector type and lane mapping as first-class fields in the design. Purchasing on speed alone can create a pile of 800G components that physically fit but are not qualified or cannot deliver the intended breakout. Cloudzat’s live listings are therefore supporting-price references, not automatic compatibility approvals.
Oversubscription must be explicit
AI training and tightly synchronized inference often react badly to hidden oversubscription because one congested uplink can stall many accelerators waiting at a collective boundary. A design advertised as 800G per endpoint is not nonblocking if the aggregate uplink budget is substantially smaller than the aggregate downlink budget.
Choose a ratio deliberately. Some inference fabrics can tolerate controlled oversubscription, while scale-out training may justify a near-1:1 design. The separate oversubscription calculator in this cluster turns the downlink and uplink totals into a visible ratio so procurement teams can see exactly what is being traded.
Failure domains become a capacity question
Redundancy is not free bandwidth. Spare ports, redundant planes and alternate paths consume switch and optic capacity that cannot always be sold to normal traffic. Spectrum-X Multiplane addresses this at the architecture level by distributing traffic across independent planes and moving around failures in hardware.
For a Spectrum-6 bill of materials, reserve ports before calculating the number of switches rather than adding a vague contingency afterward. A design that exactly fills every 800G port has no room for topology changes, replacements, breakouts or growth.
The upgrade case from Spectrum-4 depends on cluster scale
A smaller cluster already served well by Spectrum-4 does not automatically need Spectrum-6. The strongest upgrade cases appear where port radix, total switch count, future Rubin endpoint bandwidth or a planned multiplane design creates a concrete operational or economic benefit.
Existing optics and cables also influence the migration. Even when Ethernet remains the protocol, the newer electrical lane rates and system designs can change what is reusable. Build a migration inventory that lists NICs, switch ports, optics, cable lengths and software versions before assuming a chassis-only swap.
Procurement should distinguish announced architecture from orderable SKU
NVIDIA publishes architecture and platform guidance before every possible configuration is broadly available through normal distribution. The correct purchasing workflow is to identify the required Spectrum-6 system, verify lifecycle and qualification with the vendor or integrator, and then source the exact matching optics and endpoints.
Amazon listings near the top of this page cover adjacent networking hardware that can help with lab, transition or supporting infrastructure. They should not be interpreted as a source of an SN6000 switch unless the product title, part number and seller evidence explicitly prove that exact system.
A good proof of concept measures job progress, not link LEDs
A fabric can show every link as up and still deliver poor AI performance because congestion, path imbalance or retransmissions reduce useful throughput. Test NCCL or the relevant distributed framework under the same message sizes, topology and failure scenarios expected in production.
Record application step time, tail latency, retransmissions, link errors, queue behavior and recovery after fault injection. Those measurements create the evidence needed to decide whether Spectrum-6, a particular plane count and a given oversubscription ratio are actually worth the infrastructure cost.
Methodology and sources
Cloudzat treats published switch bandwidth, port counts and SerDes generation as architecture facts, then keeps topology arithmetic separate from workload performance. The calculator does not turn theoretical switching capacity into an application benchmark. Current NVIDIA product and hardware documentation should be checked again before procurement because switch SKUs, firmware and qualification can change.
- NVIDIA: Spectrum-6 arrives in gigascale AI factories
- NVIDIA Spectrum-6 SN6000 hardware specifications
- NVIDIA Spectrum-4 SN5000 hardware documentation
- NVIDIA Spectrum-X Ethernet platform
- NVIDIA ConnectX-9 SuperNIC specifications
As an Amazon Associate, Cloudzat may earn from qualifying purchases. Marketplace listings on these pages are supporting networking hardware such as NICs, switches, optics and high-speed cables. A marketplace row is not represented as a Spectrum-6 switch, ConnectX-9 SuperNIC, Thor Ultra NIC or qualified NVIDIA fabric unless the exact listing evidence supports that identity. Verify model, speed, connector, firmware, warranty and OEM qualification before purchase.
Frequently asked questions
How fast is NVIDIA Spectrum-6?
NVIDIA specifies 102.4 Tb/s of switching bandwidth for the Spectrum-6 ASIC. Current SN6000 systems can expose up to 128 800GbE ports depending on model and configuration.
Is Spectrum-6 twice as fast as Spectrum-4?
At the switch-chip level, Spectrum-6 doubles aggregate bandwidth from 51.2 Tb/s to 102.4 Tb/s. Application performance does not automatically double because topology, endpoint rates and workload communication patterns also matter.
How many 800GbE ports does Spectrum-6 support?
NVIDIA documents SN6000-class systems with 128 ports at 800GbE around Spectrum-6. Always verify the exact chassis and breakout mode.
Does Spectrum-6 require ConnectX-9?
NVIDIA pairs Spectrum-6 and ConnectX-9 in the Vera Rubin Spectrum-X generation, but interoperability and qualification depend on the exact deployment. Use the vendor compatibility matrix for nonstandard combinations.
Is Spectrum-6 Ethernet?
Yes. Spectrum-6 is NVIDIA Ethernet switch silicon used in Spectrum-X. NVIDIA adds AI-oriented routing, congestion and telemetry behavior around standards-based Ethernet.
What SerDes does Spectrum-6 use?
NVIDIA’s Vera Rubin networking material identifies 200G PAM4 SerDes for Spectrum-6, up from 100G PAM4 in Spectrum-4.
Can Spectrum-6 be used in a two-tier network?
Yes. High radix is specifically valuable in flat two-tier designs, and NVIDIA’s Multiplane architecture uses multiple independent two-tier fabrics to scale further.
Do I need 800G optics for every Spectrum-6 link?
No. The correct medium depends on configured port speed, breakout, reach and endpoint. Some designs use lower-speed lanes or split an 800G physical interface.
Is Spectrum-6 available on Amazon?
Do not assume so. Cloudzat uses Amazon primarily to show adjacent NIC, optic, cable and switch pricing. Exact Spectrum-6 systems should be verified by part number and seller channel.
What should I calculate before ordering?
Calculate endpoint bandwidth, switch-port count, plane count, oversubscription, redundancy, optics count, rack power and cooling. Then validate the resulting topology with the current NVIDIA design guide.