NVIDIA Mellanox MCX654106A-HCAT (NVIDIA P/N 900-9X6B4-0058-DT0) - ConnectX-6 VPI adapter kit: dual-port QSFP56, HDR 200Gb/s InfiniBand or 200GbE per port, Socket Direct 2 x PCIe x16, tall brackets. Condition: New, factory-sealed.
Technical Identity
The MCX654106A-HCAT is a two-board NVIDIA ConnectX-6 VPI Socket Direct kit: an adapter card (low-profile PCB, 167.65 x 68.90 mm, two QSFP56 cages) plus a PCIe Auxiliary Connection Card (129.30 x 59.00 mm), joined by two 35 cm Cabline CA-II Plus harnesses, one white and one black. Retention clips and tall plus short brackets for both boards are included. Adapter IC MT28908A0-XCCF-HVM. Each QSFP56 port is Virtual Protocol Interconnect: InfiniBand (SDR/DDR/QDR/FDR/EDR/HDR100/HDR) or Ethernet (1/10/25/40/50/100/200 GbE), selected per port in firmware.
Engineering Context
Socket Direct on this part is a genuine 32-lane design. The ConnectX-6 32-lane PCIe bus splits into two 16-lane buses: one reaches the host through the adapter's x16 edge connector, the other through the Auxiliary PCIe Connection Card. The halves are not electrically identical - the edge connector runs PCIe Gen 3.0 or Gen 4.0 (8.0 / 16.0 GT/s) while the auxiliary path is PCIe Gen 3.0 only (8.0 GT/s). Seating the two boards in slots owned by two different CPUs gives each socket its own root-complex path to the NIC instead of crossing the QPI/UPI link. HDR InfiniBand is four lanes at 53.125 Gb/s with 64b/66b encoding for 200 Gb/s effective; ConnectX-6 is IBTA v1.4 compliant, sub-600 ns latency, up to 200 million messages per second. Adapter power is 31.4 W typical with passive cables plus 3.0 W typical / 4.0 W max for the auxiliary card, 5 W per cage; airflow is 700 LFM at 55 deg C heatsink-to-port and 1050 LFM with 4.7 W active cables. Operating 0 to 55 deg C.
Deployment & Use Cases
Dual-socket HPC and AI nodes where both CPUs need direct fabric access: MPI over NVIDIA HPC-X, OpenMPI or MPICH on HDR InfiniBand; NCCL collectives with GPUDirect RDMA and GPUDirect Storage; NVMe-over-Fabrics targets and initiators pinned per NUMA node; Lustre and Spectrum Scale (GPFS) clients; 200GbE RoCEv2 leaf-spine uplinks; dual-rail designs with each port on a different leaf switch.
Technical Specifications
Brand / family: NVIDIA Mellanox ConnectX-6 VPI
MPN / OPN: MCX654106A-HCAT
NVIDIA P/N: 900-9X6B4-0058-DT0
Product type: PCIe InfiniBand/Ethernet adapter kit (HCA / NIC)
Ports: 2 x QSFP56, copper and optical
Per-port speed: HDR 200Gb/s InfiniBand or 200GbE Ethernet
InfiniBand: SDR/DDR/QDR/FDR/EDR/HDR100/HDR, IBTA v1.4
Ethernet: 1/10/25/40/50/100/200 GbE
Host interface: Socket Direct 2 x PCIe x16 - edge Gen 3.0/4.0 @ 8.0/16.0 GT/s, auxiliary Gen 3.0 @ 8.0 GT/s
Boards / harnesses: adapter 167.65 x 68.90 mm, auxiliary 129.30 x 59.00 mm, 2 x 35 cm Cabline CA-II Plus
Power / airflow: 31.4 W adapter + 3.0 W auxiliary typical; 700 LFM passive at 55 deg C heatsink-to-port
Offloads: RDMA, RoCEv2, GPUDirect RDMA and Storage, NVMe-oF, MPI tag matching, T10-DIF, SR-IOV, ASAP2
Condition: New - factory-sealed
Compatibility & Parent Systems
InfiniBand switches: NVIDIA Quantum HDR QM8700 (managed) and QM8790 (externally managed), 40 x HDR 200Gb/s QSFP56. Down-negotiates to EDR 100Gb/s against Switch-IB 2 SB7800 / SB7890.
Ethernet switches: Spectrum-2 SN3700 (32 x 200GbE QSFP56) and Spectrum-3 SN4600 (64 x 200GbE QSFP56) at 200GbE; SN3700C, SN3800, SN4600C and SN2700 at 100GbE.
Cabling: QSFP56 - NVIDIA LinkX MCP1650-H passive DAC, MFS1S00-H AOC, MMA1T00-HS multimode transceiver, MCP7H50-H HDR-to-2x-HDR100 splitters.
Host requirement, read before ordering: two free PCIe x16 slots in the same chassis, ideally one mapped to each CPU, both within reach of the 35 cm harnesses. Typical hosts are dual-socket 2U servers of the Dell PowerEdge R740/R750, HPE ProLiant DL380 Gen10, Lenovo ThinkSystem SR650 and Supermicro 2U GPU class. Check the block diagram for slot-to-CPU mapping first.
Operating systems: Linux with NVIDIA MLNX_OFED. NVIDIA states the ConnectX-6 Socket Direct 2x PCIe x16 cards - MCX654105A-HCAT, MCX654106A-HCAT, MCX654106A-ECAT - are not supported in Windows or WinOF-2.
Port mode: mlxconfig -d <device> set LINK_TYPE_P1=<IB or ETH> LINK_TYPE_P2=<IB or ETH>, then reboot. Each port serves one protocol at a time.
What This Is NOT
NOT a single-slot adapter. The MCX654106A-HCAT occupies two PCIe x16 slots; with only one free x16 slot the Socket Direct topology cannot be built.
NOT Multi-Host. NVIDIA states ConnectX-6 Socket Direct cards do not connect to two independent hosts; both PCIe paths serve one OS image.
NOT PCIe Gen 4 on both halves. The auxiliary path is Gen 3.0 only at 8.0 GT/s.
NOT supported under Windows or WinOF-2. Linux MLNX_OFED only.
NOT NDR capable. Maximum is HDR 200Gb/s; NDR 400Gb/s requires ConnectX-7 with OSFP cages.
NOT an OSFP or QSFP-DD card, and NOT supplied with transceivers, DAC/AOC cables or PCIe risers.
Frequently Asked Questions
Q. What does the MCX654106A-HCAT Socket Direct kit require from the host server?
Two free PCIe x16 slots in the same chassis, ideally one mapped to each CPU of a dual-socket server, both within reach of the two 35 cm Cabline harnesses. The adapter card goes in one slot, the PCIe Auxiliary Connection Card in the other. A single-slot or single-socket chassis cannot realise the Socket Direct topology. Confirm slot-to-CPU mapping before ordering.
Q. Is the MCX654106A-HCAT supported on Windows Server?
No. NVIDIA states the ConnectX-6 Socket Direct 2x PCIe x16 cards - MCX654105A-HCAT, MCX654106A-HCAT and MCX654106A-ECAT - are not supported in Windows or WinOF-2. Plan on Linux with MLNX_OFED.
Q. What switch models is the MCX654106A-HCAT compatible with?
In InfiniBand mode, NVIDIA Quantum HDR QM8700 and QM8790 (40 x HDR 200Gb/s QSFP56) and EDR Switch-IB 2 SB7800/SB7890 at 100Gb/s. In Ethernet mode, Spectrum-2 SN3700 and Spectrum-3 SN4600 at 200GbE, SN3700C/SN3800/SN4600C/SN2700 at 100GbE.
Q. Is the auxiliary PCIe path on the MCX654106A-HCAT Gen 3 or Gen 4?
Gen 3 only. NVIDIA specifies PCIe Gen 3.0/4.0 at 8.0/16.0 GT/s through the x16 edge connector, but PCIe Gen 3.0 at 8.0 GT/s through the PCIe Auxiliary Connection Card.
Q. What is supplied by T.E.S IT-SOLUTIONS?
New, factory-sealed kits with the adapter card, PCIe Auxiliary Connection Card, both 35 cm Cabline harnesses and brackets, from EU stock. Transceivers and QSFP56 cables 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 kits are checked for PCIe enumeration on both x16 buses, link training on both ports and harness continuity before dispatch. Handling 1-3 business days, international shipping, plus consultation on HDR fabric design and slot mapping.