Fujitsu CA05950-1822 — ConnectX-5 VPI EDR/100GbE Dual QSFP28 PCIe 3.0 Adapter (New)
Technical Identity
The Fujitsu CA05950-1822 is a dual-port VPI (Virtual Protocol Interconnect) host channel adapter built on the NVIDIA Mellanox ConnectX-5 ASIC. Each of the two QSFP28 ports operates independently in InfiniBand EDR (100 Gb/s) or 100 Gb/s Ethernet mode, configurable per port via firmware (mlxconfig). The host interface is PCIe 3.0 x16. CA05950-1822 is the Fujitsu PRIMERGY OEM order number and is hardware-identical to the generic NVIDIA Mellanox MCX556A-ECAT — same silicon, same firmware images, same MLNX_OFED drivers. Condition: New.
Engineering Context
ConnectX-5 VPI delivers native RDMA (true IB verbs for InfiniBand, RoCE v2 for Ethernet), GPUDirect RDMA for direct NIC-to-GPU data paths, SR-IOV with up to 127 virtual functions per port, hardware-accelerated overlay networks (VXLAN, NVGRE, Geneve), and stateless offloads (TCP/UDP/IP checksum, LSO, LRO). PCIe 3.0 x16 provides approximately 128 Gb/s of usable host-bus bandwidth — sufficient for single-port 100 Gb/s line rate, while dual-port simultaneous line rate (2 × 100 Gb/s = 200 Gb/s) is host-bus constrained. The card uses standard NVIDIA Mellanox firmware, so the Fujitsu OEM order number does not lock it to PRIMERGY chassis.
Deployment & Use Cases
HPC clusters running MPI workloads over EDR InfiniBand, where dual-port redundancy or aggregate throughput is required. Mixed fabric deployments where one port is wired to an InfiniBand spine (e.g., Switch-IB 2 / Quantum down-rated to EDR) and the other to a 100GbE leaf (e.g., Spectrum SN2700) — VPI eliminates the need for a second NIC. Virtualized cloud hosts requiring RoCE v2 with SR-IOV for tenant isolation. Fujitsu PRIMERGY upgrade or break/fix where the original OEM order number must be preserved on the BOM. GPUDirect RDMA pipelines for AI training where the card sits adjacent to NVIDIA GPUs on the same PCIe root complex.
Specifications
OEM Brand: Fujitsu (PRIMERGY)
OEM Part Number: CA05950-1822
NVIDIA Mellanox Equivalent: MCX556A-ECAT
ASIC: ConnectX-5
Protocol: VPI (InfiniBand EDR + 100GbE Ethernet, per-port selectable)
Per-Port Speed: 100 Gb/s (EDR InfiniBand) or 100 Gb/s (Ethernet)
Ports: 2 × QSFP28
Host Bus: PCIe 3.0 x16
RDMA: Native IB verbs + RoCE v2
SR-IOV: Up to 127 VFs per port
Offloads: VXLAN / NVGRE / Geneve, stateless TCP/UDP, GPUDirect RDMA
Form Factor: Full-height, half-length (FHHL), tall bracket
Drivers: MLNX_OFED (Linux), WinOF-2 (Windows), nmlx5 (ESXi)
Condition: New
What This Is NOT
Not ConnectX-6 — EDR maxes at 100 Gb/s per port; HDR 200 Gb/s requires ConnectX-6 silicon.
Not a 200GbE single-link adapter — each port is 100 Gb/s. Aggregate is two independent 100 Gb/s links, not one 200 Gb/s pipe.
Not an Lx (Ethernet-only) variant — this is a full VPI dual-protocol adapter.
Not a 40GbE-class card — QSFP+ (40GbE) modules are accepted via backward compatibility at 40 Gb/s, but the card is classified as a 100GbE / EDR device.
Not OCP form factor — this is a standard PCIe slot card (FHHL tall bracket).
Not a transceiver, cable, or switch — this is a host-side network interface card.
Compatibility & Hard Constraints
Each port is independently configurable for InfiniBand EDR or 100GbE Ethernet via mlxconfig — protocol selection is firmware-driven and not auto-negotiated on the wire. PCIe 3.0 x16 is required for full bandwidth; PCIe 2.0 slots reduce throughput by approximately 50%, and PCIe x8 (mechanical or electrically narrowed) slots bottleneck dual-port line rate. QSFP28 ports accept both 100GbE QSFP28 and EDR InfiniBand QSFP28 optics and DACs (e.g., NVIDIA Mellanox MMA1B00-C100D, MCP1600-Exxx, MMA1L10-CR). Backward compatible with 40GbE QSFP+ optics at 40 Gb/s. In InfiniBand mode the card pairs with Quantum HDR and Switch-IB 2 EDR fabrics (Quantum down-rates to EDR on the link). In Ethernet mode the card pairs with Spectrum and Spectrum-2 100GbE switches. Despite the Fujitsu PRIMERGY OEM order number, the card is hardware-identical to MCX556A-ECAT and operates in any vendor server platform — Dell EMC PowerEdge, HPE ProLiant, Lenovo ThinkSystem, Supermicro, or whitebox.
FAQ
Q1: What is the per-port bandwidth of the Fujitsu CA05950-1822?
Each of the two QSFP28 ports runs at 100 Gb/s — either EDR InfiniBand or 100 Gb/s Ethernet, selectable per port via mlxconfig firmware tools. Aggregate raw bandwidth across both ports is 200 Gb/s, constrained by the PCIe 3.0 x16 host bus (approximately 128 Gb/s usable).
Q2: Can the Fujitsu CA05950-1822 be used in non-Fujitsu servers?
Yes. CA05950-1822 is the Fujitsu PRIMERGY OEM order number, but the card is hardware-identical to the generic NVIDIA Mellanox MCX556A-ECAT. It uses standard Mellanox firmware and MLNX_OFED drivers and operates in any server with a PCIe 3.0 x16 slot — Dell, HPE, Lenovo, Supermicro, or whitebox.
Q3: What condition is the CA05950-1822 from T.E.S IT-SOLUTIONS?
The CA05950-1822 ships as new from T.E.S IT-SOLUTIONS (tes-itsolutions.com). Every unit is bench-validated for PCIe link training, dual-port initialization, VPI mode switching (InfiniBand ⇄ Ethernet), and RDMA functionality before shipment.
Why Choose T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS (tes-itsolutions.com) is a specialist supplier of NVIDIA Mellanox and OEM-equivalent data-center networking hardware. Every ConnectX adapter is bench-validated for PCIe link training, port initialization, VPI mode switching, and RDMA functionality before it leaves the warehouse. Technical consultation is available on fabric design, OEM cross-equivalence (Fujitsu ⇄ Mellanox ⇄ Dell ⇄ HPE part numbers), firmware compatibility, and Quantum / Spectrum / BlueField ecosystem integration. Global shipping with 1–3 business day handling.