NVIDIA Mellanox® MCB191A-FCAT — Connect-IB® Single-Port FDR InfiniBand HCA

Technical Identity
The NVIDIA Mellanox® MCB191A-FCAT is a Connect-IB® Host Channel Adapter (HCA) — a single-port FDR 56 Gb/s InfiniBand PCIe 3.0 x16 card built on the Connect-IB ASIC. Connect-IB is the dedicated InfiniBand-only HCA generation that sits between ConnectX-3 and ConnectX-4 in the Mellanox lineage. It was engineered for HPC fabrics that demand sub-microsecond MPI latency and over 100 million messages per second of small-message throughput, and remains a workhorse for FDR backbones in scientific computing.

Specifications

OEM: NVIDIA Mellanox
Part Number (MPN): MCB191A-FCAT
Product Family: Connect-IB® (InfiniBand-only HCA)
Form Factor: PCIe HHHL card, tall standard bracket
Host Bus: PCI Express 3.0 x16 (8 GT/s, ~128 Gb/s host bandwidth)
Ports: 1× QSFP+ (single-port)
Protocol: InfiniBand only (no Ethernet, no RoCE)
InfiniBand Generation: FDR (56 Gb/s, 14 Gb/s × 4 lanes 64b/66b)
Auto-negotiation: Falls back to QDR (40 Gb/s) and DDR for legacy switch interop
Connector: QSFP+
Latency: ~0.7 µs MPI ping-pong (typical, end-to-end through FDR switch)
Message Rate: >100 million messages/second per port
Condition: New, factory-sealed


What This Is NOT

NOT a ConnectX adapter. Connect-IB is its own ASIC generation. It does not share the ConnectX-3 / ConnectX-4 / ConnectX-5 / ConnectX-6 / ConnectX-7 firmware images, PSID, or full feature set. Do not confuse with MCX-series cards.
NOT a VPI dual-protocol card. Connect-IB is InfiniBand-only — it cannot be set to Ethernet link mode and does not support RoCE.
NOT dual-port. MCB191A-FCAT carries a single QSFP+ port. The dual-port equivalent in the same family is MCB192A-FCAT.
NOT bundled with cables, transceivers, low-profile bracket, or a switch.
NOT an Ethernet NIC. If you need 56GbE / 40GbE / 25GbE / 100GbE Ethernet, use a ConnectX-3 Pro EN, ConnectX-4 Lx, ConnectX-5 EN, or later generation.


Capabilities & Features

Native InfiniBand verbs with RDMA read / write / send-receive and atomic operations (64-bit and 128-bit).
Dynamically Connected Transport (DCT) — introduced with Connect-IB to scale Reliable Connection semantics on very large fabrics without exhausting QP memory.
CORE-Direct® collective communication offload (MPI all-reduce, all-gather, barrier, reduce-scatter).
GPUDirect® RDMA for direct peer-to-peer DMA between NVIDIA GPUs and the HCA, bypassing host memory.
Hardware-based congestion control, adaptive routing support, and end-to-end QoS / traffic class isolation.
SR-IOV virtualization for multi-tenant InfiniBand workloads.
Signature handover (T10-DIF/DIX) for storage data-integrity offload.


Compatibility

Drivers: NVIDIA MLNX_OFED (Linux, recommended for production HPC), in-box mlx5_core / IB core stack in modern upstream Linux kernels, WinOF (Windows Server), MLNX-NATIVE-ESX driver bundle for VMware ESXi.
Cabling: FDR-rated QSFP+ DAC copper, FDR QSFP+ AOC, and FDR QSFP+ transceivers (representative part numbers: MC2207126-004, MC2207130-00A, MC220731V-0xx, MFA1A00-Exxx). Auto-negotiates to QDR/DDR with older cabling.
Compatible switches: NVIDIA Mellanox SwitchX®-2 family (SX6036, SX6018, SX6512, SX6710 IB ports) and FDR-era IB director chassis. Interoperable at negotiated lower rates with QDR (e.g. SwitchX) and earlier DDR fabrics.
Software stack: OpenMPI, Intel MPI, NVIDIA HPC-X, MVAPICH2, and IBM Platform MPI run unmodified over Connect-IB; Lustre, GPFS / Spectrum Scale, and BeeGFS use native IB verbs.
Host slot: Requires a free PCIe 3.0 x16 mechanical / x16 electrical slot. Operates in x8 electrical at reduced peak host throughput.


Deployment & Use Cases

FDR InfiniBand HPC clusters — CFD, molecular dynamics, weather and climate simulation, seismic processing.
MPI-bound scientific workloads needing >100M msgs/sec and sub-microsecond latency.
Legacy AI training fabrics retrofitting GPU nodes onto existing FDR backbones via GPUDirect RDMA.
InfiniBand-attached parallel file systems (Lustre, GPFS / Spectrum Scale, BeeGFS) over native IB verbs.
Replacement / spare card for production Connect-IB fleets where same-ASIC firmware and PSID parity is required.
HPC lab and university clusters consolidating FDR fabric without forklifting to HDR/NDR.


Frequently Asked Questions
Q: Is the MCB191A-FCAT a ConnectX-3 card?
A: No. MCB191A-FCAT is a Connect-IB HCA, a distinct ASIC generation that sits between ConnectX-3 and ConnectX-4 in the Mellanox HCA lineage. It uses the mlx5 driver path, not mlx4, and is InfiniBand-only — it cannot be switched to Ethernet link mode.
Q: Can MCB191A-FCAT run RoCE or Ethernet?
A: No. Connect-IB is an InfiniBand-only HCA. There is no Ethernet link mode and no RoCE support. For 56GbE Ethernet or RoCE workloads, use a ConnectX-3 Pro EN, ConnectX-4 Lx, or a later ConnectX generation.
Q: Is MCB191A-FCAT a single-port or dual-port card?
A: Single-port. MCB191A-FCAT exposes one QSFP+ FDR port on a PCIe 3.0 x16 host interface. The dual-port equivalent in the same family is MCB192A-FCAT.
Q: What driver stack should I install?
A: NVIDIA MLNX_OFED is recommended for production HPC and tuned MPI workloads. Modern upstream Linux kernels include in-box mlx5_core and InfiniBand core drivers that enumerate Connect-IB out of the box. WinOF on Windows and MLNX-NATIVE-ESX on VMware ESXi are also supported.
Q: Which switches and cables are compatible with this card?
A: Any FDR-capable InfiniBand switch — Mellanox SwitchX-2 SX6036, SX6018, SX6512, SX6710 IB ports, and FDR-era IB director chassis. The card also auto-negotiates down to QDR and DDR for legacy interop. Use FDR-rated QSFP+ DAC copper, QSFP+ AOC, or QSFP+ transceiver-plus-fiber cabling.
Q: Is this unit new or refurbished?
A: New, factory-sealed in original NVIDIA Mellanox packaging where included. Every unit is bench-tested by T.E.S IT-SOLUTIONS before dispatch.

Why T.E.S IT-SOLUTIONS
Every MCB191A-FCAT we ship is bench-tested by T.E.S IT-SOLUTIONS: PCIe enumeration on a reference host, InfiniBand link establishment to a known-good FDR switch, port LED and SerDes validation, full firmware and PSID readback, and cosmetic inspection. Anti-static packaging, EU stock for fast B2B dispatch, and pre-sales technical consultation for HPC fabric design and Connect-IB ↔ ConnectX interoperability planning.