NVLink

in English: Nvidia NVLink

professionalwidespread

NVLink is Nvidia's own high-speed connection that links multiple graphics processors (and their associated central processors) so they work almost like a single giant machine.

In plain terms

The largest language models do not fit on a single chip: running them requires several, and training them requires several thousand or even tens of thousands of graphics processing units (GPUs) working together. These must continuously exchange data, and if the exchange is slow, the expensive chips wait.

NVLink is like a pass-through window built between chefs in a kitchen. Instead of carrying the plate around the hallway (which would be the traditional network), they hand it directly to each other, so the kitchen operates as a single coordinated team.

It differs from standard computer connections in that it was designed specifically for traffic between GPUs and between the GPU and the processor: it transfers much more data with much lower latency. It is Nvidia's own technology, so for a long time it worked almost exclusively with the company's chips; an exception was some IBM server processors.

How it works

GPUs are connected via separate, high-speed data lanes that bypass the computer's general, slower internal connectors.
This allows chips to directly access each other's memory, so a large model distributed across multiple GPUs runs as if it were in one place.
In larger systems, dedicated NVLink switch chips direct traffic, so that all of the several dozen GPUs in a full server rack can see one another directly.
Traffic between cabinets over longer distances is handled by other networks (InfiniBand or Ethernet): NVLink handles connections within a cabinet, while the network handles data center-level connectivity.

An everyday example

When you ask a chat AI a long question and the answer arrives in seconds, several GPUs connected via NVLink may be working on it simultaneously in the background (not all providers use Nvidia chips). You see nothing of this, only that the answer is fast.

Who is affected

developerIf you train or run a large model on multiple GPUs, the connection speed determines how much you gain from the extra chips. → Code

entrepreneur, executiveWhen purchasing or renting an AI server, NVLink is a key factor in price and performance, and it also ties you to Nvidia.

bank customer, investorNVLink is one of Nvidia's market advantages, making it a recurring subject in chip industry news and analysis. → Money

industrial worker, engineerWhen designing data centers, powering and cooling densely connected GPU racks is a separate engineering challenge. → Industry

A common misconception

MythNVLink is the same as gamers' old SLI bridge, just with a new name.

In realitySLI was a solution for games that allowed two graphics cards to calculate an image together. NVLink is a more general and much faster data connection. Some earlier high-end graphics cards did have an NVLink connector, and SLI on those operated via an NVLink bridge, but the technology's true domain is the data center, where it connects dozens of GPUs. It has disappeared from newer consumer cards.

Limits and risks

This is Nvidia's proprietary technology: since NVLink Fusion, chips from other companies can connect to it, but only alongside Nvidia products, making it harder for those building on this to switch manufacturers.
It operates only over short distances, typically within a single server or rack, and does not connect the entire data center.
Systems densely interconnected with NVLink are very expensive, consume significant power, and mostly require liquid cooling.
You hardly encounter it on home computers nowadays, so it does not speed up AI running on your own computer.

Similar, but not the same

InfiniBand – InfiniBand connects servers and racks within the data center, while NVLink connects GPUs within a single server or rack.
PCI Express – PCI Express is a general-purpose, standardized internal connector found in almost every computer. NVLink is much faster, but is designed solely for traffic between GPUs and between GPUs and the processor.
UALink – UALink is a joint open standard from several chip manufacturers for the same task, while NVLink is Nvidia's own solution.

Related concepts

Broader

AI infrastructure, Data center, high-performance computing

Narrower

NVLink Fusion, NVLink Switch

On the edge

Grafikus processzor, InfiniBand, UALink, AI factory, GPU-klaszter

A brief history

Nvidia announced NVLink in 2014, with the first products launching in 2016. In 2025, the company partially opened the technology under the name NVLink Fusion to processors and custom chips from other manufacturers.

Where things stand

By the end of 2026, this is a basic component of Nvidia-based AI servers and server racks, and a significant part of large AI data centers is built on it. The standard for its open rival, UALink, was published in 2025, but systems built on it are currently far less widespread.

Status: 2026 November

Also known as

NVLink · Nvidia NVLink · NVIDIA NVLink · NVLink Fusion · NVLink Switch · NVLink-kapcsolat · NVLink interconnect · NVLink bridge

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.