NVIDIA Mellanox® MCX653105A-EFAT ConnectX-6 VPI HDR100/EDR Single-Port QSFP56 PCIe 4.0 Network Adapter (Refurbished)
Technical Identity
The NVIDIA® Mellanox® MCX653105A-EFAT is a single-port ConnectX-6 VPI (Virtual Protocol Interconnect) HCA built on the Mellanox ConnectX-6 ASIC. The card delivers up to 100 Gb/s on a single QSFP56 port operating as HDR100 InfiniBand, EDR InfiniBand, or 100 Gigabit Ethernet — protocol selected per port via mlxconfig. The host bus is PCIe 4.0 x16 (16 GT/s, backward-compatible with PCIe 3.0 x16 at half bandwidth). Form factor is standard half-height, half-length (HHHL) with a tall bracket. Sold here in refurbished condition, checked (see Inspection and checks).
Specifications
Manufacturer documentation: NVIDIA ConnectX-6 VPI adapter card user manual: introduction. T.E.S IT-SOLUTIONS is independent of NVIDIA.
What This Is NOT
NOT a 200 Gb/s HDR adapter. The "E" suffix is HDR100/EDR — single-lane 100 Gb/s max. For full HDR (200 Gb/s) use an MCX653105A-HDAT or MCX653106A-HDAT.
NOT a Dual-port adapter. Single QSFP56 cage only. Dual-port equivalent is MCX653106A-EFAT.
NOT a Multi-Host / Socket-Direct variant. Standard single PCIe host. Socket-Direct variants carry the "S" suffix (e.g. MCX653105A-EFA-SP).
NOT an OCP 3.0 / OCP 2.0 mezzanine card. This is a standard PCIe slot card.
NOT a ConnectX-6 Dx, ConnectX-6 Lx, or ConnectX-7 card. Different ASIC, different feature set, different drivers — do not cross-flash firmware.
NOT factory-new. Refurbished, tested. New units available on request.
NOT bundled with cables, transceivers, or rail kits — sold as the adapter only.
Compatibility & Use Cases
Host server: any server with a free PCIe 4.0 x16 (or PCIe 3.0 x16) slot. PCIe 3.0 hosts will run the card at PCIe 3.0 speeds — sufficient for 100 Gb/s single-port operation.
InfiniBand switches: NVIDIA Quantum HDR / Quantum-2 NDR (downshifts to HDR100), Mellanox Switch-IB 2 EDR (interop at EDR). Cabling: QSFP56 passive copper (MCP1650-H…E26), QSFP56 active optical (MFA1A00-H…), or QSFP56→2×QSFP56 splitter (MFS1S50-H… or MCP7H50-H…) when fed from a Quantum HDR 200G port in 2×HDR100 split mode.
Ethernet switches: NVIDIA Spectrum / Spectrum-2 / Spectrum-3 100GbE ports, or any 100GbE QSFP28/QSFP56-compatible switch. Use QSFP56 or QSFP28 100G DAC / AOC / optics — 100GBASE-SR4 / LR4 / CWDM4 transceivers all supported on QSFP56 cages.
OS & drivers: NVIDIA MLNX_OFED on Linux (RHEL/CentOS/Rocky/Ubuntu/SLES), WinOF-2 on Windows Server, native nmlx5_core on VMware ESXi 6.7+. inbox drivers in modern kernels.
Typical workloads: HPC clusters running MPI / RDMA, AI training and inference fabrics (single-port edge nodes), NVMe-oF storage initiators and targets, virtualized data centers using SR-IOV, software-defined networking (SDN / NFV) with ASAP² hardware-accelerated Open vSwitch.
Frequently Asked Questions (FAQ)
What is the MCX653105A-EFAT?
The MCX653105A-EFAT is a single-port ConnectX-6 VPI host channel adapter. It supports up to 100 Gb/s on a QSFP56 port as HDR100 InfiniBand, EDR InfiniBand, or 100 Gigabit Ethernet. Host interface is PCIe 4.0 x16.
Does the MCX653105A-EFAT support full HDR 200 Gb/s?
MCX653105A-EFAT: No. The "E" in -EFAT denotes HDR100 / EDR / 100GbE — single-lane 100 Gb/s maximum. For full HDR (200 Gb/s) use MCX653105A-HDAT (single-port) or MCX653106A-HDAT (dual-port).
What connector does the MCX653105A-EFAT use?
The MCX653105A-EFAT uses a QSFP56 cage. It is backward-compatible with QSFP28 (100GbE / EDR) modules and cables. It will not fit OSFP, SFP+, SFP28, or QSFP-DD form factors.
What switches and cables work with the MCX653105A-EFAT?
MCX653105A-EFAT: InfiniBand head-end: NVIDIA Quantum HDR or Quantum-2 NDR switches. Ethernet head-end: NVIDIA Spectrum / Spectrum-2 / Spectrum-3 or any 100GbE QSFP28/QSFP56 switch. Cabling: QSFP56 DAC, QSFP56 AOC, QSFP56→2×QSFP56 HDR splitter cables, or 100G QSFP28 DAC/AOC/optics. Protocol: VPI dual-protocol — selectable per port.
How do I switch the MCX653105A-EFAT between Ethernet and InfiniBand?
MCX653105A-EFAT: This is a VPI (Virtual Protocol Interconnect) adapter. Each port is software-selectable via mlxconfig -d <pci-id> set LINK_TYPE_P1=<ETH|IB|AUTO> followed by a host reboot or firmware reset. Drivers: NVIDIA MLNX_OFED on Linux, WinOF-2 on Windows, nmlx5_core on VMware ESXi.
Will the MCX653105A-EFAT run in a PCIe 3.0 server?
MCX653105A-EFAT: Yes. The card is PCIe 4.0 x16 but fully backward-compatible with PCIe 3.0 x16 slots. PCIe 3.0 x16 provides ~126 Gb/s of usable bandwidth — enough to saturate a single 100 Gb/s port.
What workloads does the MCX653105A-EFAT accelerate?
MCX653105A-EFAT: HPC clusters with MPI / RDMA, AI / ML training and inference fabrics, NVMe-oF storage initiators and targets, virtualized data centers using SR-IOV, software-defined networking (SDN / NFV) with hardware-accelerated Open vSwitch (ASAP²), and GPUDirect RDMA / GPUDirect Storage with NVIDIA GPUs.
What condition is supplied by T.E.S IT-SOLUTIONS?
MCX653105A-EFAT: Refurbished. Each unit is checked by T.E.S IT-SOLUTIONS (see Inspection and checks). Shipped from Haifa, Israel in anti-static packaging with global B2B express options. Free pre-sale technical consultation for HPC / AI / storage fabric design.
Why buy from T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS is a Haifa-based NVIDIA Mellanox specialist serving HPC, AI, and enterprise data-center procurement teams since 2011. Multi-currency B2B invoicing with worldwide express options, and direct technical contact with our engineering team — not a call-center. See tes-itsolutions.com.