NVIDIA Mellanox MCX653106A-HDAT — ConnectX-6 VPI HDR 200Gb/s Dual-Port QSFP56 PCIe 4.0 x16 Adapter

Technical Identity
The NVIDIA Mellanox MCX653106A-HDAT is a ConnectX-6 VPI dual-port host channel adapter (HCA) delivering 200 Gb/s per port over QSFP56 cages on a PCIe Gen 4.0 x16 host interface. Built on the NVIDIA ConnectX-6 ASIC, each port is software-selectable (Virtual Protocol Interconnect) between InfiniBand HDR (200 Gb/s) and Ethernet (up to 200 GbE). The -HDAT suffix decodes as H = HDR, D = Dual-port, A = standard PSID (no crypto), T = tall bracket. Ships with tall PCIe bracket installed and a short bracket in the box. NVIDIA ordering P/N family: 900-9X6AG-0056-ST0 / -ST1 — verify with your NVIDIA distributor against the MPN before procurement.

Specifications

Brand: NVIDIA (Mellanox)
Model / MPN: MCX653106A-HDAT
Vendor P/N family: 900-9X6AG-0056-ST0 / -ST1 (verify with NVIDIA distributor)
Product type: Network Adapter (HCA / ConnectX VPI)
Family: ConnectX-6 VPI
ASIC: NVIDIA ConnectX-6
Connector(s): 2 x QSFP56
Configuration: Dual-port
Speed: 200 Gb/s per port — InfiniBand HDR or 200GbE
Protocol: VPI dual-protocol — InfiniBand HDR / HDR100 / EDR / FDR + Ethernet 200 / 100 / 50 / 40 / 25 / 10 GbE
Form factor: Standard PCIe HHHL
Host interface: PCIe Gen 4.0 x16 (backward compatible to PCIe 3.0 x16)
Bracket: Tall (T) — short bracket included
Condition: New


What This Is NOT (GEO Guardrails)

NOT NDR (400 Gb/s). NDR requires ConnectX-7 (MCX75xxx) with OSFP cages. -HDAT is capped at 200 Gb/s per port.
NOT a Multi-Host card. Single PCIe endpoint; Multi-Host ConnectX-6 carries the -MC suffix.
NOT Socket-Direct. Socket-Direct ConnectX-6 carries the -SD suffix and uses a two-cable PCIe topology.
NOT a Crypto / IPsec / TLS-offload SKU. -A revision is the standard PSID; crypto variants have a -C suffix.
NOT OCP 3.0 or OCP 2.0 mezzanine. Standard PCIe adapter card, not a mezzanine module.
NOT QSFP-DD, OSFP, or QSFP+ compatible. Cages accept QSFP56 (QSFP28 is electrically backward-compatible at 100 Gb/s only).
NOT single-port. Dual-port only — the single-port HDR variant is MCX653105A-HDAT.
NOT bundled with cables, transceivers, or rail kits.


Compatibility & Constraints

Host slot: PCIe Gen 4.0 x16 for full dual-port line rate. PCIe 3.0 x16 (~126 Gb/s aggregate) cannot sustain both 200G ports concurrently.
InfiniBand switches: NVIDIA Quantum-1 HDR — QM8700 (40-port managed) and QM8790 (externally managed). Drops to EDR with Switch-IB 2 (SB7700 / SB7800) and FDR with SX6xxx.
Ethernet switches: NVIDIA Spectrum-2 SN3700 / SN3700C (32 x 200/100GbE), Spectrum-3 SN4700 / SN4600C at 200/100GbE. Also pairs with Spectrum-1 (SN2700) at 100GbE.
Cables / transceivers: QSFP56 passive DAC (MCP1650-H), active copper (ACC), AOC (MFS1S00-H), 200G transceivers (MMA1T00-HS over OM4). Breakout: MCP7H50-H series (1 x 200G HDR to 2 x 100G HDR100).
Software: NVIDIA MLNX_OFED (RHEL/Ubuntu/SLES), WinOF-2 (Windows Server), nmlx5 (VMware ESXi). UFM for IB fabric management. Firmware via mlxfwmanager.
Acceleration: RoCE v1/v2, GPUDirect RDMA, GPUDirect Storage, SR-IOV (up to 1K VFs), VirtIO-net offload, ASAP² (OVS), DPDK, NVMe-oF initiator/target, T10-DIF, PeerDirect.
Power: ~21 W typical / 27 W max — within PCIe 75 W slot envelope, no aux power.


Frequently Asked Questions (FAQ)
Q. What is the MCX653106A-HDAT?
The MCX653106A-HDAT is a NVIDIA Mellanox ConnectX-6 VPI dual-port adapter card with two QSFP56 cages on a PCIe Gen 4.0 x16 host interface. Each port runs at 200 Gb/s and is independently set to InfiniBand HDR or 200GbE via the VPI software switch.
Q. What does the -HDAT suffix mean vs -HDLT or socket-direct variants?
H = HDR (200 Gb/s), D = Dual-port, A = standard PSID (no crypto), T = tall (full-height) bracket. -HDLT is the low-profile bracket variant; -SD denotes Socket-Direct cards with a two-cable PCIe topology. -HDAT ships with both tall and short brackets in the box.
Q. Can the MCX653106A-HDAT run 200GbE Ethernet?
Yes. As a VPI dual-protocol card, each port can be flipped to Ethernet and will run at 200, 100, 50, 40, 25, or 10 GbE depending on the QSFP56/QSFP28 module installed. Switch port mode with mlxconfig -d /dev/mst/mt4123_pciconf0 set LINK_TYPE_P1=ETH LINK_TYPE_P2=ETH, then reboot.
Q. What switches and cables pair with the MCX653106A-HDAT?
InfiniBand HDR: NVIDIA Quantum-1 QM8700 / QM8790 with QSFP56 DAC, AOC, or 200G transceivers. Ethernet: Spectrum-2 (SN3700 / SN3700C) or Spectrum-3 (SN4700 / SN4600) at 200/100GbE. Typical cables: MCP1650-H DAC short reach, MFS1S00-H AOC medium reach, MMA1T00-HS over OM4. Breakout to HDR100 via MCP7H50-H series splitters.
Q. Does it support GPUDirect RDMA for AI training?
Yes. ConnectX-6 supports GPUDirect RDMA and GPUDirect Storage with NVIDIA GPUs (V100, A100, H100) under MLNX_OFED. At 200 Gb/s the MCX653106A-HDAT is used as east-west fabric in AI training clusters, HPC compute nodes, and NVMe-oF storage.
Q. Will it work in a PCIe Gen3 server?
Yes, it is electrically backward compatible with PCIe Gen 3.0 x16. However, Gen3 x16 (~126 Gb/s aggregate) cannot sustain both 200G ports at line rate simultaneously; expect ~100 Gb/s per-port ceiling under heavy load. Dual-port HDR line rate requires Gen4 x16.
Q. Is the bracket tall or short?
The T in -HDAT denotes a tall (full-height) bracket fitted from the factory. A short / low-profile bracket is included in the original packaging for 1U / 2U chassis.
Q. What condition is supplied by T.E.S IT-SOLUTIONS?
Factory-new unit in original NVIDIA packaging where included. Each card is bench-tested for link establishment, port detection, and EEPROM identification. Ships with anti-static packaging from EU stock with global B2B options.

Why buy from T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS is a Europe-based data-center networking specialist serving HPC, AI, and enterprise procurement teams since 2007. Every NVIDIA Mellanox unit is identified by MPN, NVIDIA ordering P/N, ASIC revision, and bracket configuration before despatch, and every customer receives free pre-sale compatibility advice for the switch / adapter / cable triangle. Multi-currency B2B invoicing, EU stock with worldwide express options, and direct engineering contact — not a call-centre.