<\/span><\/h2>\n\n\n\nNVLink compatible cards are NVIDIA GPUs fitted with the NVLink interconnect, which lets two or more GPUs exchange data directly instead of routing it through PCIe and host memory. Compatibility is defined by generation. A card supports the NVLink version built into its architecture, and linked cards need matching bridges or matching switch ports.<\/p>\n\n\n\n
NVLink exists because PCIe became a bottleneck in multi-GPU systems. With NVLink, GPUs pass data and share memory peer to peer, and the CPU, host memory, and PCIe bus drop out of the path. Across six generations the aggregate bandwidth per GPU has grown from 160 GB\/s to 3.6 TB\/s. NVIDIA’s NVLink product page tracks the current specifications.<\/p>\n\n\n\n<\/span>NVLink, NVSwitch, and the Two Kinds of Compatibility<\/strong><\/strong><\/span><\/h3>\n\n\n\nThe word “compatible” means two different things depending on the card.<\/p>\n\n\n\n
On consumer and workstation cards, NVLink compatibility is physical. Two matching GPUs are joined by a removable NVLink bridge<\/strong> that clips over the top edges of both cards. The bridge comes in 2-slot, 3-slot, and 4-slot spacings so it can span the gap between adjacent cards.<\/p>\n\n\n\nOn data-center cards, compatibility is fabric-based. GPUs connect through NVSwitch<\/strong> ASICs, which create an all-to-all fabric so any GPU can reach any other at full NVLink bandwidth. NVIDIA introduced NVSwitch with the DGX-2 in 2018, and a rack-scale domain such as the GB200 NVL72 now treats up to 72 GPUs as one large device. That is roughly 130 TB\/s of fabric bandwidth in a single cabinet.<\/p>\n\n\n\n<\/span>How NVLink Compatibility Differs From SLI<\/strong><\/strong><\/span><\/h3>\n\n\n\nNVLink and SLI are often confused, and the confusion changes what hardware you buy. SLI<\/strong> (Scalable Link Interface) was a rendering technology that combined two or more GeForce cards for gaming. It used its own bridge, the SLI HB bridge on the Turing generation. NVLink<\/strong> is a general-purpose data interconnect for memory sharing and compute workloads.<\/p>\n\n\n\nThe two are not interchangeable. An RTX 3090 does not take an SLI HB bridge. It takes an NVLink bridge, in a 3-slot or 4-slot length. If you are enabling NVLink on an RTX 3090, an SLI bridge will not fit the connector and will not work.<\/p>\n\n\n\n
<\/span>Why Bridge Slots Matter on Legacy Cards<\/strong><\/strong><\/span><\/h3>\n\n\n\nSlot spacing is only relevant to the legacy GeForce cards, and it is decided by your motherboard layout rather than by the GPU. A 3-slot bridge spans a one-card gap, while a 4-slot bridge spans a wider gap left by thicker coolers or a riser. Getting this wrong is the most common reason a used RTX 3090 NVLink pair will not link, so measure the gap between the two card edges before ordering.<\/p>\n\n\n\n
<\/span>NVLink Compatible Cards: The Complete 2026 List<\/strong><\/strong><\/span><\/h2>\n\n\n\nThe table below is the master compatibility list. Bandwidth is bidirectional aggregate per GPU. “Discontinued” means NVIDIA no longer builds the NVLink connector into that product family, even though used cards still work.<\/p>\n\n\n\nCard<\/strong><\/td>Architecture and year<\/strong><\/td>NVLink gen<\/strong><\/td>Bandwidth per GPU<\/strong><\/td>NVLink status<\/strong><\/td>Interconnect needed<\/strong><\/td><\/tr>TITAN RTX, RTX 2080 Ti, RTX 2080, RTX 2080 Super<\/td> Turing, 2018<\/td> NVLink 2<\/td> ~100 GB\/s<\/td> Discontinued<\/td> 2-slot or 3-slot NVLink bridge<\/td><\/tr> RTX 3090 (base model)<\/td> Ampere, 2020<\/td> NVLink 3<\/td> 112.5 GB\/s<\/td> Discontinued, last GeForce<\/td> 3-slot or 4-slot NVLink bridge<\/td><\/tr> RTX A5000, RTX A6000<\/td> Ampere pro, 2021<\/td> NVLink 3<\/td> ~112.5 GB\/s<\/td> End of line<\/td> 2-slot NVLink bridge<\/td><\/tr> RTX PRO 6000 Blackwell<\/td> Blackwell, 2025<\/td> NVLink 5<\/td> ~1.8 TB\/s<\/td> Current<\/td> NVLink 5 link between paired cards<\/td><\/tr> A100<\/td> Ampere, 2020<\/td> NVLink 3<\/td> 600 GB\/s<\/td> Legacy<\/td> NVSwitch fabric on copper<\/td><\/tr> H100, H200<\/td> Hopper, 2022 to 2024<\/td> NVLink 4<\/td> 900 GB\/s<\/td> Current<\/td> NVSwitch 3; 1.6T OSFP between racks<\/td><\/tr> B200, GB200, GB300<\/td> Blackwell, 2024 to 2026<\/td> NVLink 5<\/td> 1.8 TB\/s<\/td> Current<\/td> NVSwitch 4; copper in rack, 800G and 1.6T optics between racks<\/td><\/tr> Vera Rubin<\/td> Rubin, 2026<\/td> NVLink 6<\/td> 3.6 TB\/s<\/td> Shipping 2H 2026<\/td> NVLink 6 fabric; co-packaged optics on the roadmap<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<\/span>Consumer GeForce and TITAN Cards With NVLink<\/strong><\/strong><\/span><\/h3>\n\n\n\nFour NVLink compatible cards came from the Turing generation: the TITAN RTX, the RTX 2080 Ti, the RTX 2080, and the RTX 2080 Super. Each links to one partner card over a bridge at roughly 100 GB\/s. The RTX 2070 Super is often listed as NVLink capable, but it has no NVLink connector and never did.<\/p>\n\n\n\n
The Ampere generation narrowed that to a single card. The RTX 3090 (base model)<\/strong> is the last consumer GeForce GPU with an NVLink connector, offering 112.5 GB\/s between two cards. The RTX 3090 Ti shipped with the physical connector but NVIDIA disabled NVLink support in the drivers, so treat the base RTX 3090 as the practical end of the line.<\/p>\n\n\n\n<\/span>Workstation Cards: RTX A-Series and RTX PRO Blackwell<\/strong><\/strong><\/span><\/h3>\n\n\n\nProfessional Ampere cards kept NVLink after GeForce dropped it. The RTX A5000 and RTX A6000 both link over a 2-slot bridge at about 112.5 GB\/s, and both are correctly classed as professional Ampere cards rather than GeForce 30-series parts.<\/p>\n\n\n\n
The line resumed with the RTX PRO 6000 Blackwell<\/strong>, which brings NVLink 5 to a workstation card at roughly 1.8 TB\/s. For studios and labs, this is the only current-generation way to buy NVLink on a single-slot-class card rather than in a server.<\/p>\n\n\n\n<\/span>Data-Center Cards: A100, H100, B200, GB200, and GB300<\/strong><\/strong><\/span><\/h3>\n\n\n\nThe data-center line is where NVLink compatible cards perform at full specification. These cards do not use bridges; they attach to a switch fabric.<\/p>\n\n\n\n
\nA100 (Ampere, 2020)<\/strong>\u00a0uses NVLink 3 at 600 GB\/s per GPU across 12 links, connected by second-generation NVSwitch into 8-GPU domains.<\/li>\n\n\n\nH100 and H200 (Hopper)<\/strong>\u00a0use NVLink 4 at 900 GB\/s per GPU across 18 links. Third-generation NVSwitch adds in-network reduction, and external NVLink Switch modules extend a domain to 256 GPUs.<\/li>\n\n\n\nB200, GB200, and GB300 (Blackwell)<\/strong>\u00a0use NVLink 5 at 1.8 TB\/s per GPU. A GB200 NVL72 rack connects 72 GPUs through nine NVSwitch trays as one domain, and GB300 Blackwell Ultra is the volume platform through 2026.<\/li>\n<\/ul>\n\n\n\n<\/span>The 2026 Flagship: Vera Rubin and NVLink 6<\/strong><\/strong><\/span><\/h3>\n\n\n\nVera Rubin brings NVLink 6, which doubles per-GPU bandwidth again to 3.6 TB\/s across as many as 36 links. The first system, Vera Rubin NVL144, is scheduled to ship in the second half of 2026, and Rubin Ultra adds co-packaged optics to the roadmap. For anyone buying cabling, this is the transition to watch, because NVLink 6 shifts more of the fabric onto high-speed optical and active electrical links. NVIDIA’s NVLink Fusion program opens the fabric to partner silicon and carries it further over optics, which is why the 1.6T OSFP layer matters for future deployments. Our companion guide on NVLink and NVSwitch generations and topology covers the switch fabric history in depth.<\/p>\n\n\n\n
<\/span>Why NVIDIA Dropped NVLink on GeForce After the RTX 3090<\/strong><\/strong><\/span><\/h2>\n\n\n\nThe single most useful fact about NVLink compatible cards in 2026 is this: NVIDIA removed NVLink from consumer GeForce cards after the RTX 3090.<\/strong> The RTX 40-series (Ada, 2022) and the RTX 50-series (Blackwell, 2025) have no NVLink connector.<\/p>\n\n\n\n<\/span>What the RTX 40-Series and RTX 50-Series Changed<\/strong><\/strong><\/span><\/h3>\n\n\n\nNVIDIA reclaimed the die area and board space that NVLink consumed and spent it on Tensor Cores, DLSS hardware, and higher clocks instead. The calculation was simple. Very few GeForce buyers ran two cards in a shared-memory configuration, while every buyer benefits from more compute and better rasterization. NVLink survived only where memory sharing drives the workload: professional, workstation, and data-center silicon.<\/p>\n\n\n\n
<\/span>What This Means If You Still Run RTX 3090s<\/strong><\/strong><\/span><\/h3>\n\n\n\nThe RTX 3090 remains a legitimate NVLink card for local AI work in 2026, and it is the last consumer option for a bridged pair. Two RTX 3090s joined by an NVLink bridge expose 48 GB of combined framebuffer, which is enough to hold models that a single 24 GB card cannot. Software support is mature, and tensor parallelism in DeepSpeed and vLLM can split a model across the pair.<\/p>\n\n\n\n
If you are building this kind of system, plan the bridge and the intervening cabling together. A 3-slot or 4-slot NVLink bridge is a hard requirement, and the bridge only carries the 112.5 GB\/s link. Storage, host traffic, and any multi-node work still travel over PCIe or Ethernet.<\/p>\n\n\n\n
<\/span>NVLink Generations and Bandwidth: Which Card Supports What<\/strong><\/strong><\/span><\/h2>\n\n\n\nEvery NVLink compatible card belongs to a generation, and that generation sets its bandwidth ceiling. The generation is fixed by architecture rather than by configuration. NVIDIA has revised the NVLink specification six times since 2016, each time roughly doubling the per-GPU rate. The table below condenses all six.<\/p>\n\n\n\n
<\/span>The NVLink 1 to 6 Bandwidth Table<\/strong><\/strong><\/span><\/h3>\n\n\n\nGeneration<\/strong><\/td>Year<\/strong><\/td>First GPU<\/strong><\/td>Links per GPU<\/strong><\/td>Bandwidth per GPU<\/strong><\/td><\/tr>NVLink 1<\/td> 2016<\/td> Tesla P100 (Pascal)<\/td> 4<\/td> 160 GB\/s<\/td><\/tr> NVLink 2<\/td> 2017<\/td> Tesla V100 (Volta)<\/td> 6<\/td> 300 GB\/s<\/td><\/tr> NVLink 3<\/td> 2020<\/td> A100 (Ampere)<\/td> 12<\/td> 600 GB\/s<\/td><\/tr> NVLink 4<\/td> 2022<\/td> H100 (Hopper)<\/td> 18<\/td> 900 GB\/s<\/td><\/tr> NVLink 5<\/td> 2024<\/td> B200 (Blackwell)<\/td> 18<\/td> 1.8 TB\/s<\/td><\/tr> NVLink 6<\/td> 2026<\/td> Vera Rubin<\/td> up to 36<\/td> 3.6 TB\/s<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<\/span>Reading the Numbers: Per-GPU Aggregate vs Per-Link<\/strong><\/strong><\/span><\/h3>\n\n\n\nTwo details prevent most misreading of NVLink specs.<\/p>\n\n\n\n
First, published bandwidth is bidirectional aggregate<\/strong>, meaning transmit plus receive. A figure of 900 GB\/s on an H100 is 450 GB\/s in each direction, not 900 GB\/s each way.<\/p>\n\n\n\nSecond, a card can share a generation without sharing the bandwidth. The RTX 3090 and the A100 both use NVLink 3, yet the RTX 3090 delivers 112.5 GB\/s while the A100 delivers 600 GB\/s. The consumer card devotes one link pair to the bridge, while the data-center card puts 12 links on a switch fabric. Never assume two cards with the same generation number perform alike.<\/p>\n\n\n\n
<\/span>Scale-Up vs Scale-Out: Where NVLink Ends and Optics Begin<\/strong><\/strong><\/span><\/h2>\n\n\n\nEvery NVIDIA AI cluster has two networks, and they use different physical media. The two layers also decide what kind of transceiver, DAC, or copper assembly a given NVLink compatible card needs. Getting the boundary right is the difference between a working fabric and a rack of dropped links.<\/p>\n\n\n\n
<\/span>Inside the Rack: Copper NVLink<\/strong><\/strong><\/span><\/h3>\n\n\n\nScale-up<\/strong> is NVLink, and inside a rack it runs on copper. A GB200 NVL72 uses copper NVLink at the rear of the cabinet because the distances are short, under a few meters, and the power saving is large. Running that fabric optically would add roughly 20 kW per rack. Those copper runs are high-speed twin-axial assemblies, effectively large DAC cables, that plug into the compute trays and the NVSwitch trays.<\/p>\n\n\n\n<\/span>Between Racks: InfiniBand NDR and 800G OSFP<\/strong><\/strong><\/span><\/h3>\n\n\n\nScale-out<\/strong> is the network that carries jobs, storage, and traffic between racks, and it runs on optics. DGX H100 systems use 400G ConnectX-7 InfiniBand, GB300 moves to 800G ConnectX-8, and the switches are Quantum-X800 for InfiniBand or Spectrum-X800 for Ethernet. Where an NVLink domain must cross a rack boundary, as in the two-cabinet GB200 NVL36x2, the NVSwitch trays connect through 1.6T OSFP cages carrying real NVLink traffic.<\/p>\n\n\n\nThe practical rule is simple. Keep NVLink on copper where the distance is short and the power budget is tight, switch to optics where reach or density demands it, and standardize on 800G and 1.6T OSFP so the cabling carries across the GB300 and Vera Rubin generations.<\/p>\n\n\n\n<\/span>The Cabling Each NVLink Compatible Card Needs<\/strong><\/strong><\/span><\/h3>\n\n\n\nThis table maps the compatibility list to the physical layer, which is the part a cabling supplier can actually help with.<\/p>\n\n\n\n