The answer, first
Flat-top goes in the server. Finned-top goes in the switch.
If you're populating an OSFP port on a ConnectX-7 or ConnectX-8 adapter, a BlueField-3 DPU, or a DGX system, you need the flat-top module — Mellanox part MMA4Z00-NS400, which NVIDIA now also numbers 980-9I51S-00NS00, for 400G SR4.
If you're populating a port on an air-cooled Quantum-2 switch — a QM9700 or QM9790 — you need the finned-top module, part MMA4Z00-NS, which is twin-port and carries 800G as 2 × 400G.
That's the whole rule. The rest of this post is why, what happens when you get it wrong, and the third variant that catches people out.
Why the fins matter
The fins aren't a cooling upgrade. They're a statement about which end of the link the module belongs in.
A ConnectX-7 OSFP port has a heatsink built into the cage. It sits down on top of whatever module you insert. That's why NVIDIA now calls the flat-top form factor RHS — Riding Heat Sink. The equipment's heatsink rides on the module. There is no clearance for a module with its own fins, so a finned module simply won't seat.
An air-cooled switch port has no such heatsink. The module has to bring its own. NVIDIA calls that IHS — Integrated Heat Sink, because the fins are integrated into the module's metal top.
The fins are either on the module or on the equipment. Never both, never neither.
If you learned this as "flat-top and finned-top," you're not out of date — those are still the common names. But NVIDIA's documentation increasingly says RHS and IHS, so it's worth knowing both, because you'll see both in the same procurement thread.
What a wrong order actually costs
The finned-top-in-a-NIC mistake is the common one, and it's usually reasoning that sounds right: 800G is a bigger number than 400G, so surely it's the better module. But the twin-port 800G finned module is a switch part. It won't fit the NIC, and the 800G isn't usable at the server end anyway — the ConnectX-7 OSFP port is a 400G port.
The reverse mistake — flat-top in an air-cooled switch — is worse, because the module physically seats. It looks like it worked. There's just no cooling path.
Neither mistake is expensive per unit. Both are expensive at cluster scale, because nobody orders one. They order forty-eight, and the mismatch surfaces on install day.
The canonical topology
The two parts are designed as a pair:
One twin-port finned module in the switch → two MPO-12/APC patch cords → two flat-top modules in two ConnectX-7 servers.
One switch module serves two servers. When you're sizing an order, that's the ratio: for every two server ports, one switch module.
The third variant nobody stocks
Here's the one that isn't in most catalogs.
There is an 800G SR8 module in flat-top — part MMA4Z00-NS-FLT. It exists for liquid-cooled switch deployments, where a cold plate needs to contact a flat surface and integrated fins are exactly wrong.
So the full picture is three variants, not two:
| Deployment | Module |
|---|---|
| ConnectX-7 / ConnectX-8 / DGX server port | 400G flat-top (RHS) — MMA4Z00-NS400 |
| Air-cooled Quantum-2 switch | 800G twin-port finned-top (IHS) — MMA4Z00-NS |
| Liquid-cooled switch | 800G flat-top — MMA4Z00-NS-FLT |
If you're building liquid-cooled and someone quoted you the finned part, that quote is wrong. We don't stock the flat-top 800G — we'd rather tell you the part number than sell you the one we have.
Part numbers that look alike but aren't
A few near-misses worth checking your PO against:
- MMA1Z00-NS400 — one character off, and it's QSFP112, a different cage entirely. It will not fit an OSFP port.
- MMS4X00-NS400 — single-mode DR4 at 1310nm, 500m reach. Different optics, different fiber plant.
- MMS4X00-NM / MMS4X00-NS — the 800G single-mode DR8 variants.
The MMA / MMS prefix is the tell: MMA is multimode, MMS is single-mode. Everything in this post is multimode SR. If your plant is single-mode, none of these are your part — and any seller who doesn't ask which you're running isn't paying attention.
→ Full NVIDIA OSFP part-number crosswalk — every part number in this range mapped to form factor, optics, reach and which end of the link it belongs in.
Why you'll see 50m on one sheet and 100m on another
Worth knowing before you plan a run. NVIDIA qualifies these modules at 50m over OM4. The SR MSA the optics are built to permits up to 100m over OM4. Both numbers are correct — the standard describes what the optics can do under ideal conditions, and NVIDIA's number is what they will support in an InfiniBand fabric.
The practical part: every additional MPO mating in the path eats optical budget. A 90m run through three patch panels is exactly where link margin quietly disappears. Design to 50m and you keep support parity with NVIDIA and a margin you can troubleshoot.
Ready to Buy?
We stock both parts, new, and ship from Knoxville, Tennessee. Lifetime warranty, backed by Resilient Tec.
- 400G OSFP SR4 Flat-Top (RHS) — MMA4Z00-NS400 compatible — the ConnectX-7 / ConnectX-8 / DGX server-side module
- 800G OSFP SR8 Finned-Top (IHS) — MMA4Z00-NS compatible — the Quantum-2 air-cooled switch-side module
Populating a cluster? Tell us the port count and we'll quote the pair together. Running single-mode, or building liquid-cooled? Ask — we'd rather point you at the right part than sell you the wrong one.
You Might Also Need
- OSFP DAC cables — when the run is short enough that copper beats optics
- MPO-12/APC multimode patch cords, Type B polarity — the fiber between the two modules above
- QSFP112 stacking cables for DGX Spark and the DGX Spark interconnect reference
- Cat8 and structured cabling for the rest of the rack
- Network switches — Arista, Cisco, HPE/Aruba and Mellanox, new and tested pre-owned