NVIDIA Mellanox® MFS1S50-H005E — LinkX HDR 200Gb to 2x HDR100 QSFP56 AOC Splitter, 5m
Technical Identity
The NVIDIA Mellanox® MFS1S50-H005E is a LinkX active optical cable (AOC) splitter that breaks a single QSFP56 200Gb HDR InfiniBand trunk into two independent QSFP56 100Gb HDR100 adapter legs over a 5-metre multi-mode fibre assembly. The "1S50" digits in the MPN encode a 1-to-2 splitter topology; the "H" prefix denotes HDR-class signalling at 200Gb (4× 50Gb PAM4 lanes) on the trunk side, with each leg carrying 2× 50Gb PAM4 lanes at HDR100 or 100GbE. Integrated 850nm VCSEL multi-mode optics are factory-fused on every connector, hot-pluggable, host-powered, and field-non-serviceable. Condition: Refurbished Grade A, checked by T.E.S IT-SOLUTIONS (see Inspection and checks).
Specifications
Manufacturer documentation: NVIDIA MFS1S50-H0xxE 200Gb/s to 2x100Gb/s QSFP56 AOC product specifications. T.E.S IT-SOLUTIONS is independent of NVIDIA.
What This Is NOT
NOT a 1-to-1 cable. 1× to 2× splitter only — single QSFP56 trunk feeds two independent QSFP56 legs; a single leg cannot carry the full 200Gb.
NOT a passive copper DAC. Active optical with integrated multi-mode VCSEL optics on every connector. Copper DAC reach limits (~3m at HDR) do not apply.
NOT a transceiver pair. Optics are permanently fused into a fixed 5m fibre assembly — they cannot be removed or reused on other cables.
NOT an OSFP-to-QSFP hybrid splitter. Both ends are QSFP56 form factor. This is not an NDR 400Gb cable and will not seat in OSFP cages on Quantum-2 / Spectrum-4 switches.
NOT for NDR / 400G fabrics. Maximum trunk rate is 200Gb HDR. Do not specify for greenfield AI training clusters built on Quantum-2 / Spectrum-4 NDR400.
NOT plug-and-play on every cage. Breakout requires the switch port-pair to be configured for 1-to-2 mode (OpenSM / UFM for InfiniBand, NVOS / Cumulus for Ethernet).
NOT bundled with ConnectX adapters, transceivers, or rail kits — sold as a standalone cable assembly.
Compatibility & Constraints
Trunk switches (HDR InfiniBand): NVIDIA Quantum-1 (QM8700, QM8790) with QSFP56 cages. Port must be set to 1× 200Gb HDR with paired downstream ports configured for 2× HDR100 breakout via OpenSM / UFM.
Trunk switches (200GbE): NVIDIA Spectrum-2 (SN3700, SN3700V) and Spectrum-3 (SN4700) running breakout-capable Cumulus Linux or NVOS firmware.
Adapter side: NVIDIA ConnectX-6 / ConnectX-6 Dx (MCX653x / MCX656x families) with QSFP56 ports at HDR100 or 100GbE.
Cage mismatch (will NOT work): Quantum-2 (QM9700 / QM9790) and Spectrum-4 (SN5400 / SN5600) use OSFP cages and reject QSFP56 connectors.
Length / reach: 5m fixed — supports intra-rack server-to-ToR and adjacent-rack server-to-EoR runs. Not suited for spine-leaf inter-row links; use 15m, 20m, or 30m MFS1S50-H variants.
Routing: OM3 / OM4 multi-mode bend-radius rules; do not coil tighter than 30mm under load.
Verify before deploy: switch firmware, MLNX_OFED / DOCA driver, and adapter port-mode must match NVIDIA's LinkX firmware compatibility matrix.
Frequently Asked Questions (FAQ)
What is the MFS1S50-H005E?
MFS1S50-H005E: An NVIDIA Mellanox LinkX active optical cable that splits a single 200Gb HDR InfiniBand (or 200GbE Ethernet) QSFP56 trunk into two independent 100Gb HDR100 (or 100GbE) QSFP56 legs over a 5-metre multi-mode fibre assembly with integrated 850nm VCSEL optics.
Is the MFS1S50-H005E a 1-to-1 cable or a splitter?
The MFS1S50-H005E is a splitter: 1 trunk to 2 legs. The trunk has a single QSFP56 at 200Gb HDR; the adapter side has two QSFP56 connectors each at 100Gb HDR100. The split is permanent — it cannot be reconfigured to a 1-to-1 link.
Does this work with NDR / 400G hardware like Quantum-2 or Spectrum-4?
No. Trunk side is QSFP56 HDR at 200Gb maximum. Quantum-2 and Spectrum-4 use OSFP cages at NDR 400Gb; the MFS1S50-H005E will not seat in those cages.
Can the MFS1S50-H005E carry Ethernet, or only InfiniBand?
MFS1S50-H005E: Both. On a Spectrum-2 / Spectrum-3 switch in QSFP56 200GbE mode with breakout enabled, the two legs become 100GbE ports. On a Quantum-1 InfiniBand switch, the legs carry HDR100. Correct operation depends on switch and adapter port configuration.
Which NVIDIA adapters terminate the split legs?
MFS1S50-H005E: Each leg terminates in a QSFP56 cage on a ConnectX-6 or ConnectX-6 Dx adapter (MCX653x / MCX656x families) at HDR100 or 100GbE. The two legs can feed two separate single-port adapters, or both ports of one dual-port card.
What workload class is 200Gb / HDR appropriate for?
HDR / 200Gb fabrics are the workhorse for HPC compute clusters, NVMe-oF storage targets, mid-scale AI inference, and east-west links inside existing HDR training clusters. Frontier-scale AI training today specifies NDR 400Gb; MFS1S50-H005E suits brownfield HDR expansion, not greenfield NDR.
Where in the data centre does a 5m AOC typically live?
MFS1S50-H005E: 5m supports intra-rack server-to-ToR and adjacent-rack server-to-EoR runs. Longer spine-leaf or inter-row links should use 15m, 20m, or 30m AOC variants in the MFS1S50-H family.
What condition is supplied by T.E.S IT-SOLUTIONS?
MFS1S50-H005E: Refurbished Grade A. Every unit is checked by T.E.S IT-SOLUTIONS before it ships (see Inspection and checks). Ships from Haifa, Israel with anti-static packaging.
Why buy from T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS is a Haifa-based data-centre networking specialist serving HPC, AI cluster, and enterprise procurement teams since 2011. Every cable, transceiver, adapter, and switch we ship is identified by MPN, NVIDIA P/N, and ASIC generation before dispatch, and every order includes free pre-sale compatibility advice on switch / adapter / cable pairing. Multi-currency B2B invoicing with worldwide express options, and direct technical contact with our networking engineers — not a call-centre. See tes-itsolutions.com.