NVIDIA MMA4Z00-NS Compatible 800G OSFP SR8 Finned-Top (IHS) Twin-Port Transceiver — Dual MPO-12/APC, for Quantum-2 QM9700/QM9790

Finned-Top (IHS) — the switch-side module, not the NIC-side module

NVIDIA MMA4Z00-NS compatible — 800G OSFP SR8 twin-port, 850nm multimode, breaks out to 2 x 400G

→ NVIDIA OSFP part-number crosswalk — Mellanox MMA/MMS numbers mapped to the new NVIDIA 980- numbers, with form factor, optics and the correct part for each end of the link.

OEM Equivalent: NVIDIA / Mellanox MMA4Z00-NS
Description: Twin-port OSFP SR8 optical transceiver, 800Gb/s, finned-top / Integrated Heat Sink (IHS), dual MPO-12/APC, multimode

This is the finned-top module — the one that goes in the switch, not the server. Air-cooled Quantum-2 switch ports have no heatsink of their own, so the module has to carry its own fins. Put this in a ConnectX-7 and it will not seat. Put a flat-top module in an air-cooled switch and it seats fine, looks like it worked, and has no cooling path at all.

NVIDIA now calls this form factor IHS, for Integrated Heat Sink — the fins are integrated into the module's own metal top. Older documentation and existing BOMs still say "finned-top." Same part.

It is a twin-port module: 800G total, presenting as 2 x 400G. One of these serves two servers.

Which One Do I Need?

The fins are either on the module or on the equipment. Never both, never neither.

And OSFP does not go in a BlueField-3 at all. NVIDIA states that BlueField-3 adapters accept QSFP112 devices only, and that OSFP is not for use in BlueField-3 DPUs. Neither this module nor the 400G flat-top will seat in one — that port needs the MMA1Z00-NS400, the same optics in a QSFP112 shell, on the same MPO-12/APC fiber. A twin-port module in the switch can feed QSFP112 adapters and OSFP adapters in any combination — the fiber is identical, only the far-end shell changes.

The canonical topology: one twin-port finned module in the switch → two MPO-12/APC patch cords → two flat-top modules in two ConnectX-7 servers. When you are sizing an order, that is the ratio — for every two server ports, one switch module.

Technical Specifications

Note on Reach: Why You'll See Both 50m and 100m

NVIDIA specifies 50m over OM4, and that is the figure to design against in an InfiniBand fabric. The optics are built to the SR MSA, which permits up to 100m over OM4 — both numbers are correct, they answer different questions. The standard describes what the optics can do under ideal conditions; NVIDIA's number is what they will support. NVIDIA's own documentation for this twin-port family describes straight multimode crossover cables up to 100m and splitter cables up to 50m, which is the same budget described from the cable's end rather than the module's.

Usable reach past 50m depends on fiber grade and how many MPO matings sit in the path, because every additional mating costs optical budget. Design to 50m for NVIDIA support parity. Planning a longer run? Talk to us first.

Good to Know

• This module will not seat in a ConnectX-7 OSFP port. The port cage carries its own riding heatsink and there is no clearance for fins.
• 800G is not an upgrade at the server end — the ConnectX-7 OSFP port is a 400G port. If you are buying for a NIC, you want the 400G flat-top module, or the QSFP112 version if the adapter is a BlueField-3.
• Populating DGX H100 front cages? Those are twin-port flat-top cages. NVIDIA's documentation for this family specifies the flat-top 800G variant, MMA4Z00-NS-FLT, for the DGX H100 — the finned part is for Quantum-2 and Spectrum-4 switches, where airflow is reduced. The same flat-top logic applies to liquid-cooled deployments, where a cold plate contacts a flat surface. We do not stock MMA4Z00-NS-FLT — but that is the part number to specify, and if someone has quoted you the finned part for a DGX front cage, that quote is wrong.
• MMS4X00-NM and MMS4X00-NS are the 800G single-mode DR8 variants. The MMA / MMS prefix is the tell: MMA is multimode, MMS is single-mode.
• Running single-mode? This is not your module. Tell us what the fiber plant is and we will point you at the right part rather than sell you this one.

Quick Checklist — Is This the Right Module?

You are populating a port on an air-cooled Quantum-2 InfiniBand switch (QM9700 or QM9790). The port has no heatsink of its own. Your fiber plant is multimode (OM3 or OM4) with MPO-12/APC connectors. Your runs are 50m or shorter. You want one switch port to feed two 400G server ports.

Availability

In stock at Resilient Tec · Additional units ship on a 7–10 day lead time.
Building a cluster? Tell us how many ports you are populating and we will quote the switch-side and server-side modules together. Request a quote →

Warranty: lifetime, backed by Resilient Tec — not a manufacturer term. When a module fails you reach a named seller in Knoxville, Tennessee, not a ticket queue.

You Might Also Need

• 400G OSFP SR4 Flat-Top (RHS) transceiver — the server-side pair. Order two per module of these.
• MMA1Z00-NS400 — 400G QSFP112 SR4 — the adapter-side module when the far end is a BlueField-3 or a ConnectX-7 QSFP112
• MPO-12/APC OM4 multimode trunk cable, Type B, female-to-female — two per twin-port module, $59.00 to $139.00
• MPO-12/APC 1:2 splitter cables — built to order, when one twin-port switch port feeds four adapters at NDR200
• OSFP DAC cables — when the run is short enough that copper beats optics
• Product guides — structured cabling, switches and the rest of the rack

Related Guides

Flat-Top or Finned-Top OSFP? How to Tell Which Module Your ConnectX-7 Actually Needs

MPO-12/APC vs UPC: Why Your Existing Fiber Won't Work With NDR Optics

NVIDIA OSFP Part-Number Crosswalk: Mellanox MMA/MMS to NVIDIA 980- Numbers

OSFP vs. QSFP-DD: A Practical Breakdown for GPU Clusters