What Is Cumulus Linux?

Cumulus Linux is a network operating system (NOS) — the software that turns switch hardware into a working network device. Its defining idea is right there in the name: it is actually Linux, a Debian-based distribution, not a proprietary switch OS with a Linux kernel hidden underneath.

That distinction is the whole point. On a traditional switch you manage the box through a vendor-specific command line and configuration idiom that exists nowhere else. On Cumulus Linux, interfaces, routing, and configuration are Linux constructs managed with tooling a systems engineer already knows — so the same automation stack used for servers (configuration management, version control, CI-style validation) extends straight to the network layer instead of requiring a separate discipline and a separate team.

NVIDIA acquired Cumulus Networks in 2020 and folded the OS into its networking line. It is now the default NOS preloaded on NVIDIA Spectrum Ethernet switches.

Cumulus Linux only makes sense in the context of open networking — the model where the switch hardware and the network operating system are decoupled and sold separately, rather than welded into one vendor stack.

On a Spectrum switch, the silicon and the NOS are independent layers. That is why the same physical chassis can ship running different operating systems, and why "which OS" is a purchasing decision encoded in the ordering part number rather than something you pick after the box arrives. The main options that run on Spectrum hardware are Cumulus Linux (NVIDIA's Linux-native NOS), Onyx (the older Mellanox-lineage OS, now largely withdrawn on current platforms), and the community-standard SONiC.

This is the opposite of the classic model where buying the switch and buying the software were a single, inseparable transaction. The full comparison of the two NVIDIA-shipped options is in Cumulus Linux vs Onyx.

Because Cumulus Linux is Linux, you can configure it two ways.

• Directly, as Linux. Interfaces, routing daemons (it uses FRR for BGP, OSPF and the rest), and system state are ordinary Linux configuration, editable and scriptable like any other Debian host.
• Through a unified management layer. NVIDIA provides NVUE (NVIDIA User Experience), a structured configuration model and CLI/REST interface that gives the switch a single coherent object model to manage — a more approachable path than hand-editing files, and one that fits automation cleanly. NVUE succeeded the earlier NCLU (Network Command Line Utility) that longtime Cumulus users will recognize.

Either way, the underlying state is Linux, so a change can be templated, reviewed, and rolled out across a fleet of switches the same way a server configuration would be — which is exactly why large and cloud-style buildouts reach for it.

Open networking needs a standard way to install an OS onto bare switch hardware, and that is ONIE — the Open Network Install Environment. ONIE is a small boot loader pre-installed on open switches that lets you install a compatible NOS image onto the hardware, the way a server's firmware lets you install an operating system.

In practice this means a Spectrum switch is a hardware platform that boots ONIE and then loads a NOS image — Cumulus Linux, or another supported OS. It is the mechanism that makes the hardware/OS separation real rather than theoretical: the switch you buy is entitled to a particular NOS, and ONIE is how that image lands on it. What you are entitled to run is set by the part number, not by what the silicon can technically boot — a distinction that matters most on secondary-market hardware, where the OS license and update entitlement don't automatically travel with the box.

Cumulus Linux runs on NVIDIA Spectrum Ethernet switches — the merchant-silicon platforms built on the Spectrum ASIC family. In current ordering, that spans the SN2000 (Spectrum-1), SN4000 (Spectrum-3, including the SN4600C), and SN5000 (Spectrum-4) families, from top-of-rack leaf switches up to spine and super-spine platforms.

It shows up constantly in AI and data-center fabrics for a specific reason: the network around a GPU cluster — the leaf-and-spine Ethernet fabric carrying scale-out, storage, and management traffic — is exactly the kind of large, repetitive switch fleet that benefits from being managed as fleet infrastructure rather than as individually-configured appliances. That fabric is a first-class part of a buildout decision, not an afterthought; where Ethernet fits against InfiniBand for that role is covered in InfiniBand vs Ethernet for AI. To browse the switches themselves, see the GPU and networking catalog.