InfiniBand

professionalwidespread

InfiniBand is an extremely fast, low-latency networking technology that connects the machines of supercomputers and AI data centers so they can work as a single large machine.

In plain terms

Training a large AI model is not done by one computer, but by several thousand, which must continuously exchange data. InfiniBand is the cabling and rule system on which this exchange takes place. This is therefore not the internet, but the internal network of the data center.

Imagine it like a factory's internal pneumatic tube system. The traditional network is like city mail: the package passes through many hands, sometimes gets lost, and must be resent. The tube system delivers the item directly to the recipient's desk, without queues or lost packages.

Its main advantage is very low and predictable latency. If thousands of graphics processors (GPUs) work together, the slowest connection holds back the entire work, so network speed is directly measurable in money.

Its best-known alternative is Ethernet, which you know from home and office networks. Ethernet is cheaper and offered by many more manufacturers, while InfiniBand is designed specifically for peak performance.

How it works

Dedicated network cards are installed in the servers, and these are connected to InfiniBand switches by copper or optical cables.
A central control program, the subnet manager, maps the entire network and pre-determines which data goes where.
The sending machine only starts transmitting data when the receiver signals that it has free space. This prevents packets from being lost due to congestion and rarely requires time-consuming retransmission.
Data bypasses the operating system, moving directly from one machine's memory to another. This solution is called Remote Direct Memory Access (RDMA).
Multiple lanes run in parallel within a connection, so modern systems achieve speeds of several hundred gigabits per second per connection.

An everyday example

When you type a question to a chat AI, the model behind it may have been trained in a data center where GPUs were connected by InfiniBand, although many large clusters now rely on Ethernet. You see none of this. You encounter it in the news when reports are written about the construction of a new AI data center or supercomputer.

Who is affected

developerIf you train or run a large model on many GPUs, the network type noticeably affects runtime and cost. → Code

industrial worker, engineerData center design, cabling, and operation require specific InfiniBand expertise, which differs from standard Ethernet knowledge. → Industry

entrepreneur, executiveWhen acquiring your own AI infrastructure, the network represents a significant portion of the cost, and the choice can bind you to a single supplier for a long time.

bank customer, investorThe market for AI data center networks is a significant item in the business results of chip manufacturers and networking companies, so it frequently appears in investor news. → Money

A common misconception

MythInfiniBand is just a faster internet cable that could be installed at home.

In realityInfiniBand does not connect to the internet but links machines within a data center. It requires dedicated cards, switches, and control software, making it unsuitable and uneconomical for home use.

Limits and risks

The market is practically dominated by a single manufacturer, Nvidia, which could increase procurement costs and create supplier dependency.
It is more expensive than Ethernet, and its operation requires fewer specialists.
It is meant only for relatively short distances inside a data center; Ethernet and the internet network are still needed for connections to the outside world.
A fast network alone does not speed things up: if the software cannot efficiently distribute work across machines, the advantage is lost.

Similar, but not the same

Ethernet – Ethernet is a general-purpose, multi-vendor networking standard from home to data center, while InfiniBand is specifically an internal network for high-performance computing clusters.
RDMA – RDMA is a data transfer method (direct memory-to-memory copying) that InfiniBand uses by default, but can also be implemented over Ethernet with appropriate hardware.
NVLink – NVLink connects GPUs within a single server or rack, while InfiniBand connects servers to each other.

Related concepts

Broader

Data center, Computer Networks, high-performance computing

Narrower

subnetwork manager

On the edge

Ethernet, RDMA, GPU, Supercomputer, GPU cluster, NVLink

A brief history

The InfiniBand Trade Association, which maintains the standard, was formed in 1999 by the merger of two competing industry initiatives; the first specification appeared in 2000. The main manufacturer of the technology became Mellanox, which Nvidia acquired in 2020.

Where things stand

By the end of 2026, InfiniBand remains a standard network for supercomputers and large AI training clusters. However, in new AI data centers, AI-optimized Ethernet is a serious competitor, and many new large clusters are built on it.

Status: 2026 November

Also known as

InfiniBand · Infiniband · IB · InfiniBand network · InfiniBand network · InfiniBand network · InfiniBand switch · Nvidia Quantum InfiniBand

Sources

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