NVIDIA releases MMS1X00-N5400 QSFP112 optical modules supporting 400Gbps interconnect over 500m distances.

NVIDIA recently supplied us with MMS1X00-N5400 QSFP112 1310NM 400Gbps optical modules to connect two NVIDIA GB300 AI Station systems to an NVIDIA Spectrum-X switch. Given the investment these represent, we wanted to document their specifications and performance characteristics.

The module uses the QSFP112 form factor—a quad-channel pluggable standard with 112G PAM4 signaling per channel, delivering 400Gbps total throughput. This compact form factor is considerably smaller than the OSFP modules increasingly common at 800Gbps and 1.6Tbps speeds. The connector is an MPO-12, though a 4×112G design uses only eight of the twelve available fibers.

NVIDIA distinguishes these modules with external heatsinks mounted outside the pluggable cage. While many 400Gbps ports rely on internal cooling within the device's cages, NVIDIA's external heatsink approach allows the modules to extend slightly beyond the NIC and switch, improving thermal dissipation in demanding environments.

The QSFP112 form factor differs meaningfully from QSFP56-DD alternatives. Lower-cost PAM4 systems running 56Gbps per lane require eight lanes to reach 400Gbps and thus demand a different connector. Critically, QSFP112 optics like these support direct connectivity without requiring a gearbox inside the module—a significant advantage when deploying 400Gbps 500m DR4 links that need to connect to a switch without additional signal processing.

Testing on an NVIDIA GB300 AI Station connected to a high-end switch confirmed 400Gbps connectivity, with the module autonegotiating almost instantaneously on the NVIDIA ConnectX-8 NIC. The module reported a 10W maximum power draw and maintained mid-40°C operating temperatures under load—well below thermal thresholds.

The investment was justified by operational flexibility. With extensive DR4 optics already deployed across the lab, these modules enable easy switching between systems or direct GB300-to-GB300 connections via ConnectX-8 NICs and standard 12-fiber MPO/MTP-12 singlemode cable. In large AI clusters, NVIDIA optics proved more thermally robust at higher ambient temperatures than generic alternatives—a meaningful consideration for liquid-cooled GB300 systems where active optical module cooling is minimal. From an integration standpoint, the reliability gain warranted the cost.