NVIDIA Mellanox MCX556M-ECAT-S25 (NVIDIA P/N 900-9X5AD-0056-DT1) - ConnectX-5 VPI Socket Direct adapter: dual-port QSFP28, EDR 100Gb/s InfiniBand or 100GbE, 2 x PCIe Gen 3.0 x8 with a 25 cm Slim-Line SAS harness, tall bracket. Condition: Refurbished, tested by T.E.S IT-SOLUTIONS.
Technical Identity
The MCX556M-ECAT-S25 is a two-board NVIDIA ConnectX-5 VPI Socket Direct assembly: a low-profile adapter card (167.65 x 68.90 mm) with two QSFP28 cages, plus an Auxiliary PCIe Connection Card (113.0 x 40.00 mm), joined by a 25 cm Slim-Line SAS harness. Silicon is MT28808A0-FCCF-EVM. Each QSFP28 port is Virtual Protocol Interconnect: InfiniBand SDR/DDR/QDR/FDR/EDR or Ethernet 1/10/25/40/50/100 GbE. The "M" in the OPN denotes the Socket Direct multi-slot build and "S25" the 25 cm harness - the S35A sibling uses a 35 cm harness.
Engineering Context
NVIDIA splits the ConnectX-5 16-lane PCIe bus into two 8-lane buses: one reaches the host through the card's x8 edge connector, the other through an x8 parallel connector to the Auxiliary PCIe Connection Card. Both are PCIe Gen 3.0, SERDES @ 8.0 GT/s, x8 (Gen 2.0 and 1.1 compatible) - there is no Gen 4 path. The two boards go into two adjacent PCIe x8 slots joined by the harness. Because each x8 bus is enumerated separately, the two 100Gb/s physical ports appear to the operating system as four netdevices of up to 50Gb/s each. NVIDIA is explicit that full 100Gb/s performance is only achieved when both the adapter card and the auxiliary card are installed in the dual PCIe x8 connectors. EDR is four lanes at 25.78125 Gb/s NRZ with 64b/66b encoding for 100 Gb/s effective; IBTA v1.3 compliant. Typical power 17.1 W passive, maximum 20.0 W passive and 25.5 W with 2.5 W active cables; 5 W per QSFP28 cage. Airflow heatsink-to-port: 400 LFM passive, 900 LFM with 5 W modules. Operating 0 to 55 deg C. Offloads: RDMA and RoCEv2, GPUDirect RDMA, NVMe-oF target offload, erasure coding, T10-DIF, SR-IOV, ASAP2 with VXLAN/NVGRE.
Deployment & Use Cases
Dual-socket HPC and AI nodes where NIC-to-CPU NUMA locality matters: MPI over NVIDIA HPC-X or OpenMPI on EDR InfiniBand; GPU nodes using GPUDirect RDMA; NVMe-over-Fabrics endpoints tuned per NUMA node; Ceph and VMware vSAN clusters on RoCEv2; Lustre and Spectrum Scale (GPFS) clients.
Technical Specifications
Brand / family: NVIDIA Mellanox ConnectX-5 VPI
MPN / OPN: MCX556M-ECAT-S25
NVIDIA P/N: 900-9X5AD-0056-DT1
Product type: PCIe Socket Direct InfiniBand/Ethernet adapter
Ports: 2 x QSFP28, copper and optical
Per-port speed: EDR 100Gb/s InfiniBand or 100GbE Ethernet
InfiniBand: SDR/DDR/QDR/FDR/EDR, IBTA v1.3
Ethernet: 1/10/25/40/50/100 GbE
Host interface: Socket Direct 2 x PCIe Gen 3.0 x8 @ 8.0 GT/s
Boards / harness: adapter 167.65 x 68.90 mm, auxiliary 113.0 x 40.00 mm, 25 cm Slim-Line SAS
Netdevices presented: 4, up to 50Gb/s each
Power / airflow: 17.1 W typical passive, 5 W per port; 400 LFM passive heatsink-to-port
Condition: Refurbished - tested by T.E.S IT-SOLUTIONS
Compatibility & Parent Systems
InfiniBand switches: NVIDIA Switch-IB 2 SB7800 (36 x EDR 100Gb/s QSFP28, managed) and SB7890 (externally managed). Quantum HDR QM8700 / QM8790 ports also link at EDR or HDR100 with the correct cable.
Ethernet switches: NVIDIA Spectrum SN2700 (32 x 100GbE QSFP28), SN2410 and SN2100; Spectrum-2 SN3700C and SN3800; Spectrum-3 SN4600C.
Cabling: QSFP28 - NVIDIA LinkX MCP1600-E series EDR passive DAC, MFA1A00-E series 100G AOC, MMA1B00-E100 multimode transceiver.
Host requirement, read before ordering: two adjacent free PCIe x8 or wider slots in the same chassis, ideally one mapped to each CPU, both within the 25 cm harness reach. Typical hosts: Dell PowerEdge R640/R740, HPE ProLiant DL360/DL380 Gen10, Lenovo ThinkSystem SR630/SR650, Supermicro dual-socket 1U/2U.
Operating systems: NVIDIA lists OpenFabrics Enterprise Distribution (MLNX_OFED) as the supported stack for ConnectX-5 Socket Direct - plan on Linux.
Port mode: mlxconfig -d <device> set LINK_TYPE_P1=<IB or ETH> LINK_TYPE_P2=<IB or ETH>, then reboot.
What This Is NOT
NOT a single-slot x16 adapter. It needs two adjacent PCIe x8 slots plus the 25 cm harness; it cannot be deployed as a conventional single-edge-connector NIC.
NOT full speed with only the primary card fitted. NVIDIA states 100Gb/s per port is reached only when both boards are installed in the dual x8 connectors.
NOT presented as two interfaces. The OS sees four netdevices at up to 50Gb/s each - size bonding and MPI rank placement accordingly.
NOT PCIe Gen 4. Both paths are Gen 3.0 x8 at 8.0 GT/s; a Gen 4 slot negotiates down.
NOT HDR or NDR capable. EDR 100Gb/s per port maximum; HDR needs ConnectX-6, NDR needs ConnectX-7.
NOT a QSFP56, QSFP-DD, OSFP or SFP28 card, NOT Multi-Host, NOT a BlueField DPU, and NOT the 35 cm S35A variant.
NOT supplied with transceivers, DAC/AOC cables or PCIe risers.
Frequently Asked Questions
Q. What does the MCX556M-ECAT-S25 require from the host server?
Two adjacent free PCIe x8 or wider slots in the same chassis, ideally one mapped to each CPU of a dual-socket server, both within reach of the 25 cm harness. A single-slot chassis cannot host it, and placing both connectors under the same root complex loses the NUMA benefit.
Q. How many network interfaces does the MCX556M-ECAT-S25 present to the OS?
Four. Each PCIe x8 bus is enumerated separately and each sees both physical ports, so the two 100Gb/s QSFP28 ports appear as four netdevices of up to 50Gb/s each. This is expected ConnectX-5 Socket Direct behaviour, not a fault.
Q. What switch models is the MCX556M-ECAT-S25 compatible with?
InfiniBand: NVIDIA Switch-IB 2 SB7800 and SB7890 (36 x EDR 100Gb/s QSFP28); Quantum QM8700/QM8790 also link at EDR or HDR100. Ethernet: Spectrum SN2700, SN2410 and SN2100, Spectrum-2 SN3700C and SN3800, Spectrum-3 SN4600C.
Q. Is the MCX556M-ECAT-S25 PCIe Gen 4, and how does it differ from the S35A?
Neither path is Gen 4 - both are PCIe Gen 3.0 x8 at 8.0 GT/s. The only difference between MCX556M-ECAT-S25 and MCX556M-ECAT-S35A is harness length, 25 cm versus 35 cm; electrically, mechanically and in port configuration they are identical, and NVIDIA lists MLNX_OFED as the supported stack for both.
Q. What is supplied by T.E.S IT-SOLUTIONS?
Refurbished. Each assembly is bench-tested for PCIe enumeration on both x8 buses, link training on both QSFP28 ports and harness continuity, then packed anti-static. EU stock. Cables and optics are separate line items.
Why T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS (tes-itsolutions.com) supplies NVIDIA Mellanox networking hardware from EU stock. Socket Direct assemblies are checked for PCIe enumeration on both x8 buses, link training on both ports and harness continuity before dispatch. Handling 1-3 business days, plus consultation on slot mapping and NUMA pinning.