NVIDIA Mellanox MCX653105A-HDAT - ConnectX-6 VPI adapter card: single-port QSFP56, HDR 200Gb/s InfiniBand or 200GbE Ethernet, PCIe 3.0/4.0 x16, tall bracket. Condition: Refurbished, tested by T.E.S IT-SOLUTIONS.
Technical Identity
The MCX653105A-HDAT is a single-port NVIDIA ConnectX-6 VPI host channel adapter. NVIDIA's ordering description reads "ConnectX-6 InfiniBand/Ethernet adapter card, HDR IB (200Gb/s) and 200GbE, single-port QSFP56, PCIe4.0 x16, tall bracket". The board is a PCIe half-height half-length card, 167.65 x 68.90 mm, built on the MT28908A0-XCCF-HVM device, shipped with the tall bracket fitted and a short bracket in the box. The QSFP56 cage is Virtual Protocol Interconnect: InfiniBand (SDR/DDR/QDR/FDR/EDR/HDR100/HDR) or Ethernet (1/10/25/40/50/100/200 GbE), chosen in firmware. A server host adapter - not a switch, not a BlueField DPU, not a Socket Direct kit.
Engineering Context
HDR InfiniBand is four lanes at 53.125 Gb/s line rate with 64b/66b encoding, giving 200 Gb/s effective on one QSFP56 port; in Ethernet mode the same four lanes carry 200GBASE-CR4, KR4 or SR4. ConnectX-6 is IBTA v1.4 compliant with sub-600 ns latency and up to 200 million messages per second. The host slot matters: NVIDIA states a ConnectX-6 in a single PCIe x16 slot supports up to 100Gb/s in a PCIe Gen 3.0 slot, or up to 200Gb/s in a PCIe Gen 4.0 slot - HDR line rate therefore needs Gen 4. Typical board power is 19.3 W with passive cables, with up to 5 W to the QSFP56 module. Airflow is 350 LFM at 55 deg C heatsink-to-port or 250 LFM at 35 deg C port-to-heatsink with passive cables, rising to 500 LFM at 55 deg C with NVIDIA 4.7 W active cables. Operating 0 to 55 deg C, altitude to 3050 m. Hardware offloads: RDMA and RoCEv2, GPUDirect RDMA and PeerDirect, MPI tag matching and rendezvous, adaptive routing, NVMe-over-Fabrics, erasure coding, T10-DIF, SR-IOV and ASAP2 Open vSwitch offload with VXLAN/NVGRE.
Deployment & Use Cases
HDR InfiniBand compute nodes in AI training clusters running NCCL, Megatron-LM, NeMo or PyTorch DDP with GPUDirect RDMA; MPI workloads over NVIDIA HPC-X, OpenMPI or MPICH; Lustre and Spectrum Scale (GPFS) clients; NVMe-over-Fabrics targets and initiators; 200GbE RoCEv2 uplinks in Spectrum fabrics; sites that need one adapter SKU re-purposable between InfiniBand and Ethernet.
Technical Specifications
Brand / family: NVIDIA Mellanox ConnectX-6 VPI
MPN / OPN: MCX653105A-HDAT
Product type: PCIe InfiniBand/Ethernet adapter (HCA / NIC)
Ports: 1 x QSFP56, copper and optical
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: PCIe Gen 3.0/4.0 SERDES @ 8.0/16.0 GT/s, x16 lanes
Form factor: PCIe half-height half-length, 167.65 x 68.90 mm, tall bracket fitted
Power: 19.3 W typical passive; 5 W to the QSFP56 module
Airflow: 350 LFM passive at 55 deg C heatsink-to-port; 500 LFM with 4.7 W active cables
Offloads: RDMA, RoCEv2, GPUDirect RDMA, NVMe-oF, MPI tag matching, T10-DIF, SR-IOV, ASAP2
Condition: Refurbished - tested by T.E.S IT-SOLUTIONS
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 (36 x EDR QSFP28).
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 slot: one PCIe x16 slot; Gen 4.0 for 200Gb/s line rate, Gen 3.0 limits the port to about 100Gb/s. Typical hosts: Dell PowerEdge R750/R7525, HPE ProLiant DL380 Gen10 Plus, Lenovo ThinkSystem SR650 V2, Supermicro GPU nodes.
Drivers: NVIDIA MLNX_OFED (Linux), WinOF-2 (Windows), VMware ESXi and FreeBSD builds; in-box mlx5_core on current kernels.
Port mode: mlxconfig -d <device> set LINK_TYPE_P1=<IB or ETH>, then reboot. One protocol at a time.
What This Is NOT
NOT a Socket Direct card. One PCIe x16 edge connector, no auxiliary-card path. NVIDIA's single-port HDR Socket Direct kit is MCX654105A-HCAT; the single-slot 2x8 variant is MCX653105A-EFAT.
NOT Multi-Host. One PCIe path, one host, one OS image.
NOT a dual-port card. Dual-port HDR is MCX653106A-HDAT.
NOT full speed in a Gen 3 slot. PCIe Gen 3.0 x16 caps the port near 100Gb/s. The 100Gb/s-max sibling is MCX653105A-ECAT.
NOT NDR capable. Maximum is HDR 200Gb/s; NDR 400Gb/s needs ConnectX-7 with OSFP.
NOT an OSFP or QSFP-DD card, and NOT the liquid-cooled MCX653105A-HDAL.
NOT supplied with transceivers, DAC/AOC cables, auxiliary cards or risers.
Frequently Asked Questions
Q. What switch models is the MCX653105A-HDAT 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. Cabling is QSFP56 throughout.
Q. Does the MCX653105A-HDAT reach 200Gb/s in a PCIe Gen 3 slot?
No. NVIDIA states a ConnectX-6 card in a single PCIe x16 slot supports up to 100Gb/s in a Gen 3.0 slot and up to 200Gb/s in a Gen 4.0 slot. Full HDR or 200GbE line rate needs PCIe Gen 4.0 x16.
Q. How do I switch the MCX653105A-HDAT between InfiniBand and Ethernet?
Run mlxconfig -d <device> set LINK_TYPE_P1=<IB or ETH> and reboot. VPI is a per-port setting and the port carries one protocol at a time.
Q. Is the MCX653105A-HDAT a Socket Direct adapter?
No. It is a conventional single-slot PCIe x16 card. NVIDIA's dual-slot Socket Direct kit for single-port HDR is MCX654105A-HCAT, and the single-slot 2x8 Socket Direct variant is MCX653105A-EFAT.
Q. What airflow does the MCX653105A-HDAT need?
350 LFM at 55 deg C heatsink-to-port, or 250 LFM at 35 deg C port-to-heatsink, with passive cables; 500 LFM at 55 deg C with NVIDIA 4.7 W active cables. Typical board power is 19.3 W.
Q. What condition is supplied by T.E.S IT-SOLUTIONS?
Refurbished. Each MCX653105A-HDAT is bench-tested for PCIe x16 lane training, QSFP56 link-up, firmware integrity and thermal sensors, 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. Every ConnectX adapter is powered up and verified before dispatch - PCIe lane width and generation, port link training, firmware revision and EEPROM data. Handling 1-3 business days, international shipping, plus consultation on HDR fabric design and cabling.