NVIDIA Mellanox MCX454A-FCAT ConnectX-4 VPI Adapter, Dual-Port QSFP28, FDR InfiniBand (56Gb/s) and 40/56GbE Ethernet, PCIe 3.0 x8 (P/N 900-9X426-0054-ST0): This is a Virtual Protocol Interconnect (VPI) adapter, meaning each port can be independently configured for InfiniBand or Ethernet operation, not a single-protocol NIC. Maximum InfiniBand rate on this SKU is FDR (56Gb/s); maximum Ethernet rate is 56GbE per port. This SKU is the FDR/56GbE bin and is NOT EDR-capable. Ships with a tall (full-height) PCIe bracket. This is a refurbished unit.
Engineering Context
ConnectX-4 VPI integrates IB and Ethernet on the same silicon, allowing per-port protocol selection via the mlxconfig tool. The card supports RDMA (RoCE v1/v2 over Ethernet, native RDMA over IB), GPUDirect for NVIDIA GPUs, hardware-offloaded transport (T10-DIF/SIG), SR-IOV (up to 128 VFs), and overlay-network offload for VXLAN, NVGRE, and GENEVE. PCIe 3.0 x8 yields ~64Gb/s bidirectional host bandwidth, which is the binding constraint when both QSFP28 ports run at 56Gb/s simultaneously. The -FCAT bin specifies the FDR/56GbE max-rate variant; faster -CCAT (EDR/100GbE) and -DCAT (EDR/40GbE) bins exist as separate SKUs.
Deployment & Use Cases
HPC compute and storage clusters with FDR InfiniBand fabric. AI/ML training nodes using GPUDirect over InfiniBand. Hybrid IB/Ethernet head-node where one port joins the IB compute fabric and the other carries 40/56GbE management or storage traffic. NFS-RDMA or NVMe-oF (RoCE) initiator/target adapter at 40/56GbE. Refresh of FDR InfiniBand cluster nodes where matching the existing fabric speed avoids costly rip-and-replace.
Technical Specifications
OEM: NVIDIA Mellanox
Part Number: MCX454A-FCAT (NVIDIA P/N 900-9X426-0054-ST0)
Series: ConnectX-4 VPI
Network Ports: 2 × QSFP28
Maximum InfiniBand Rate: FDR (56Gb/s) per port
Maximum Ethernet Rate: 56GbE per port (40/56GbE)
Protocol Selection: Per-port via mlxconfig (VPI)
Host Interface: PCI Express 3.0 x8
RDMA: Native IB RDMA + RoCE v1/v2
SR-IOV: Up to 128 virtual functions
GPUDirect: Supported
Bracket: Tall (full-height)
Condition: Refurbished — Tested by T.E.S IT-SOLUTIONS
Compatibility & Hard Constraints
This is the FDR/56GbE bin — it will NOT operate at EDR (100Gb/s IB) or 100GbE; for EDR/100G use a -CCAT or -ECAT bin. Requires a PCIe 3.0 x8 (or wider) slot; PCIe 3.0 x8 is the host-side bandwidth ceiling when both ports are saturated. VPI per-port protocol must be set with mlxconfig BEFORE the host driver loads in the target mode — changing protocol requires reboot. Ships only with a tall bracket; low-profile bracket is not included. RoCE deployments require DCB/PFC-configured Ethernet switching. InfiniBand deployments require an FDR-or-faster IB switch and Subnet Manager.
Why Choose T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS validates every ConnectX-4 VPI on a live PCIe 3.0 host: PCIe link width verification, both ports brought up in IB and Ethernet modes, firmware-level health check. Visit www.tes-itsolutions.com. Standard handling time 1–3 business days.