NVIDIA Mellanox MCX346A-FCPN ConnectX-3 Pro VPI Dual-Port QSFP+ OCP 2.0 Mezzanine Card, IB FDR 56Gb/s and 40GbE Ethernet: This is a refurbished OCP 2.0 mezzanine network card with two QSFP+ ports built on the ConnectX-3 Pro VPI controller. Each port can be configured at boot as InfiniBand FDR (56Gb/s) or Ethernet (40GbE). The host interface is PCIe 3.0 x8 routed through the OCP mezzanine connector. This is NOT a standard PCIe add-in card — it requires an OCP 2.0-compatible server slot. This is a refurbished unit.
Engineering Context
VPI (Virtual Protocol Interconnect) means the firmware can present each port as InfiniBand or Ethernet through mlxconfig (LINK_TYPE_P1, LINK_TYPE_P2). Maximum InfiniBand rate is FDR at 56Gb/s (4x14Gb/s NRZ), and maximum Ethernet rate on this Pro generation is 40GbE — the hardware QSFP+ cage is not standards-compliant for 56GbE. Aggregate dual-port FDR is 112Gb/s, so the PCIe 3.0 x8 host interface (64Gb/s) is the bottleneck for fully saturated dual-port traffic — single-port saturation runs at line rate. ConnectX-3 Pro adds VXLAN/NVGRE overlay offload and RoCEv2 with congestion control on top of the ConnectX-3 baseline.
Deployment & Use Cases
OCP-server HPC compute nodes running MPI over InfiniBand FDR. Mixed-protocol racks where some servers are reconfigured between IB and 40GbE workloads via firmware. Storage front-end on OCP servers using NVMe-oF or NFSoRDMA at 40GbE or FDR. Dual-homed cluster nodes with port-level redundancy across two switches.
Technical Specifications
- OEM: NVIDIA Mellanox
- Part Number: MCX346A-FCPN
- Controller: ConnectX-3 Pro VPI
- Ports: Dual QSFP+
- InfiniBand: FDR 56Gb/s, FDR10, QDR, DDR
- Ethernet: 40GbE (no 56GbE, no 100GbE)
- Host Interface: PCIe 3.0 x8 via OCP connector
- Form Factor: OCP 2.0 Mezzanine, no bracket
- RDMA: InfiniBand verbs, RoCEv1, RoCEv2
- Overlay Offload: VXLAN, NVGRE stateless
- Virtualization: SR-IOV up to 127 VFs/port
- Compliance: RoHS R6
- Condition: Refurbished — Tested by T.E.S IT-SOLUTIONS
Compatibility & Constraints
Requires OCP 2.0 mezzanine slot in the host server. WILL NOT fit standard PCIe x8/x16 add-in slots. Not mechanically compatible with OCP 3.0 (NIC 3.0) cages. PCIe 3.0 x8 host bandwidth (64Gb/s) is below the 112Gb/s aggregate dual-port FDR rate — full bidirectional dual-port saturation is bottlenecked by the host bus. Each port is selectable IB or Ethernet independently via firmware. QSFP+ cage accepts FDR-rated DAC, AOC, and SR4/LR4 transceivers; EDR/100G transceivers will not negotiate at 100G on this card. RoCEv2 deployment requires PFC/DCB on the switching fabric.
Why TES
T.E.S IT-SOLUTIONS bench-tests every dual-port VPI OCP NIC for InfiniBand and Ethernet link establishment on both ports, PCIe x8 width through the OCP connector, and firmware reachability before shipment. Visit tes-itsolutions.com for OCP-platform compatibility checks and bulk pricing. Handling time 1–3 business days with international shipping.