NVIDIA SN4600C Specs & Datasheet (MSN4600-CS2RC)

The SN4600C is NVIDIA's 64-port 100GbE switch built on the Spectrum-3 ASIC. It packs 64 QSFP28 cages into a 2U chassis, which is what makes it a spine and super-spine platform rather than a top-of-rack leaf: at that port density a single pair of switches can aggregate a large number of leaf uplinks without chaining tiers.

It is an open Ethernet switch — the hardware and the network operating system are decoupled, so the platform ships with a NOS preloaded rather than welded to a vendor stack. That decoupling is why the same physical switch appears under several ordering part numbers, and it is the single most common source of confusion when buying one (see the part-number decoder below).

The SN4600C is not a GPU or an accelerator. It is the network layer around a GPU cluster — a distinction worth making, because it is frequently listed alongside compute in secondary-market inventory and is often assumed to be part of the accelerator itself.

The SN4600C platform specification. Port count, speed, switching capacity, and packet-processing rate are fixed characteristics of the Spectrum-3 platform; airflow direction and preloaded operating system vary by ordering part number and are not field-changeable.

All 64 ports are QSFP28 cages running at 100GbE. Because QSFP28 is a four-lane form factor, each port also breaks out to lower speeds — 10, 25, 40, and 50GbE — using breakout cables or splitter optics.

That breakout capability is what lets one platform serve two roles as a cluster grows. Used natively at 100G, the SN4600C is an aggregation and spine switch. Broken out to 25G, a single 64-port chassis fans out to a much larger number of lower-speed access ports, which is how the same switch ends up in storage and management roles rather than only in the spine.

The practical planning note: breakout consumes port budget in groups, not singly. Decide the speed mix before you size the switch count, because a design that assumes 64 usable 100G ports and then breaks half of them out is a different switch count than the spreadsheet started with.

The SN4600C is the platform name. What you actually order is an MSN4600 part number, and the suffix encodes two things that cannot be changed after delivery:

• Airflow direction — F = P2C (air moves from the power-supply side toward the ports), R = C2P (air moves from the ports toward the power supplies).
• Network operating system — a trailing C marks the Cumulus Linux build. Part numbers without it were the Onyx builds.

So MSN4600-CS2RC reads as: 100GbE, 2 power supplies, C2P airflow, Cumulus Linux. Its sibling MSN4600-CS2FC is the identical switch with P2C airflow.

This matters more than it looks. Airflow direction has to match the hot-aisle/cold-aisle orientation of the rack it lands in, and it is a manufacturing option, not a configuration setting — the wrong variant is a returned shipment, not a support ticket. The full trade-off is covered in MSN4600-CS2RC vs MSN4600-CS2FC.

Current SN4600C ordering part numbers ship with Cumulus Linux preloaded. The earlier Onyx (formerly MLNX-OS) part numbers for this platform — the ones without the trailing C — have been withdrawn, and NVIDIA's ordering documentation for the SN4000 family now lists only the Cumulus variants for the SN4600C.

That lifecycle split is a live procurement issue rather than a historical footnote, because Onyx-based SN4600C units still circulate on the secondary market. A unit is not defective for being an Onyx build, but it is a discontinued ordering configuration, and the operating-system entitlement is the thing to verify rather than assume. Cumulus Linux vs Onyx covers what each OS is and what actually transfers with a resold unit.

The SN4600C platform itself remains in mass production — it is the OPN-level Onyx configurations that were retired, not the switch.

A GPU cluster runs more than one network, and they are not interchangeable. The east-west compute fabric that carries gradient and collective traffic between GPUs is the highest-bandwidth network in the building. Alongside it sit a storage fabric and an in-band management network, both of which are typically 100GbE.

The SN4600C is generally specified for the second group. NVIDIA's DGX BasePOD and SuperPOD reference architectures use SN4600-series switches for the 100GbE storage and in-band management fabrics, with the higher-bandwidth compute fabric handled separately. This is why a 100GbE switch shows up in GPU-cluster bills of materials: it is the network around the accelerators, and the cluster does not run without it.

The planning consequence is that switch count follows node count and storage design, not GPU count alone. A cluster sized purely on accelerators routinely under-orders the fabric it needs to actually operate.

The SN4600C is a fixed-configuration catalog part rather than a built-to-order system, which makes it one of the few items where a listed price is meaningful — see the MSN4600-CS2RC listing for current availability and lead time, or browse the full networking catalog.

Three things are worth confirming before any order, whatever the source: the airflow variant matches your rack orientation, the operating-system entitlement transfers, and the date codes are what you expect on units described as new. Tell us the fabric you are standing up and we will come back on availability and fit.