Two DGX H100 or H200 nodes can be linked directly over NDR InfiniBand, with no Quantum-2 switch in between. The catch is the cabling. The DGX cluster ports are flat-top, twin-port 800G OSFP cages, and most of the "DGX cables" sold online, including the finned-top MCP4Y10 DAC, will not seat in them. This guide covers what fits, what does not, and what a two-node setup needs besides cables.
Short answer
Use flat-top twin-port optics at both ends, not a copper DAC
Put a flat-top 800G twin-port module in each DGX cage you use, such as the MMS4X00-NM-FLT (single-mode, 500 m), and join them with MPO-12/APC OS2 trunks, two per module. NVIDIA states that DACs "are generally not used" on these cages because of internal chassis cabling losses. The MCP4Y10 DAC is finned on both ends and is built for switch-to-switch links only.
Why the DGX H100 Is Different From a ConnectX-7 Server
In an HGX server from Dell, HPE or Supermicro, each GPU usually has its own ConnectX-7 PCIe card with a single-port OSFP cage. In a DGX H100 or H200, the eight ConnectX-7 cluster interfaces sit on internal mezzanine cards and are routed to four twin-port OSFP cages. Each cage carries two 400G links, and each cage is flat-top: the heat sink is built into the cage, so the module must have a flat top.
Two consequences follow. A finned-top module or cable is physically too tall for the cage. And because the signal has already travelled through internal cabling before it reaches the cage, NVIDIA does not use passive copper DACs here. Its LinkX guide says DACs "are generally not used due to the internal chassis cabling losses" and lists a single 3-meter active copper cable as the only copper option.
What Fits, and What Does Not
| Part | Top style | DGX H100 / H200 cage? |
|---|---|---|
| MMS4X00-NM-FLT 800G 2×DR4, single-mode | Flat | Yes, in stock |
| MMA4Z00-NS-FLT 800G 2×SR4, multimode | Flat | Yes, by quote |
| NVIDIA 3 m active copper cable (ACC) | Flat | Yes, the one copper option NVIDIA lists |
| MCP4Y10-N001 / N002 passive DAC | Finned, both ends | No. Switch-to-switch only |
| MMA4Z00-NS / MMS4X00-NM (finned) | Finned | No. These go in the switch |
How Many Parts Two Nodes Need
Each DGX has four cluster cages carrying eight 400G links. To connect every rail, you need four flat-top twin-port modules per node, eight in total, and eight MPO-12/APC trunks: one per 400G link, joining rail 1 on node A to rail 1 on node B, and so on. Both ports of a twin-port module must use the same fiber type, and single-mode DR modules need OS2 single-mode trunks, not OM4. You can start with fewer rails; NCCL will use the links that are up, at lower bandwidth.
What You Need Besides Cables
- A subnet manager. An InfiniBand fabric needs one, and with no switch there is nothing else to run it. Run OpenSM on one of the two nodes, one instance per rail.
- Two nodes is the ceiling. A direct link joins exactly two ports. A third DGX means a Quantum-2 switch such as the QM9700, and at that point the switch side takes finned-top modules.
- Test before you trust it. Run nccl-tests across both nodes and confirm all eight rails carry traffic.
- Expect to own the support. NVIDIA's reference designs (BasePOD, SuperPOD) are switched. A switchless pair works at the InfiniBand level, but it is not a configuration NVIDIA documents.
For the full Hopper cabling picture, including switch-side parts and breakout cables for HGX servers, see the H100 / H200 interconnect matrix.
FAQ
Can two DGX H100 systems be connected without a switch?
Yes. InfiniBand ports can be linked directly, so two DGX H100 or H200 nodes can be cabled back-to-back rail to rail. One node must run a subnet manager such as OpenSM. A third node requires a Quantum-2 switch.
Does the MCP4Y10 DAC fit a DGX H100?
No. The MCP4Y10-N001 and N002 are twin-port 800G passive DACs with finned tops on both ends, made to link two Quantum-2 or Spectrum-4 switches. The DGX H100 cluster cages are flat-top, so a finned connector will not seat.
What cable connects two DGX H100 nodes directly?
Flat-top twin-port 800G optics in each DGX cage, such as the MMS4X00-NM-FLT single-mode module, joined by MPO-12/APC trunk cables. NVIDIA says passive DACs are generally not used on these cages, and lists a single 3-meter active copper cable as the only copper option.
How many modules and cables do two DGX H100s need?
For all eight rails: four flat-top twin-port modules per node, eight modules in total, and eight MPO-12/APC trunk cables, one per 400G link.
Ready to Buy?
- MMS4X00-NM-FLT flat-top 800G 2×DR4 — $999, in stock
- MPO-12/APC OS2 Type B trunk — from $55, in stock in six lengths
- MPO-12/APC OM4 Type B trunk — from $59, for multimode SR builds
Speccing a two-node pair or a full pod? Send us your node count and we will quote the modules and fiber together. Prices checked November 2, 2026.