Mellanox/Lenovo ConnectX-5 VPI Adapter Card (MCX556A-ECAT / 01KT927)

The Mellanox ConnectX-5 Virtual Protocol Interconnect (VPI) dual-port adapter card provides an extremely high-bandwidth, low-latency, and agile interconnect solution for high-performance computing (HPC), machine learning/artificial intelligence (AI), enterprise data centers, and advanced cloud storage systems. It supports up to 100 Gb/s EDR InfiniBand or 100 Gb/s Ethernet on each of its dual QSFP28 ports.

This specific part number includes the Lenovo FRU designation (01KT927), indicating it is certified and tested for native compatibility with Lenovo ThinkSystem enterprise servers.

1. Product Identification

• Mellanox OPN (Orderable Part Number): MCX556A-ECAT

• Lenovo FRU (Field Replaceable Unit): 01KT927

• Family: ConnectX®-5 VPI (Virtual Protocol Interconnect)

• Bracket/Form Factor: Low Profile (LP) MD2 design (often shipped with both Full-Height and Low-Profile brackets, but the board itself is low profile).

2. Technical Specifications

Feature	Specification
Interface / Bus Type	PCI Express (PCIe) 3.0 x16 (Fully backwards compatible with PCIe 1.1 / 2.0)
Bus Transfer Rate	8.0 GT/s per lane (total raw bandwidth of up to 128 Gb/s)
Ports	Dual-Port (2x) QSFP28 (Quad Small Form Factor Pluggable 28)
Network Protocols	Dynamic dual-protocol: EDR InfiniBand or 100GbE Ethernet
InfiniBand Speed Rates	SDR, DDR, QDR, FDR, EDR (up to 100 Gb/s)
Ethernet Speed Rates	1 / 10 / 25 / 40 / 50 / 100 Gb/s
Onboard RAM	1GB DDR4 memory
Cabling Support	Direct Attach Copper (DAC), Active Copper Cables (ACC), Active Optical Cables (AOC), or optical transceivers.
Dimensions	2.71 in x 5.6 in (68.90 mm x 142.24 mm) – standard MD2 Low Profile

3. Key Performance & Hardware Offloads

Intelligent Offloads & Accelarators

• Advanced Storage Offloads: Native support for NVMe over Fabrics (NVMe-oF) target/initiator offloads, T10-DIF signature handover, and NVMe emulation, which significantly lowers CPU overhead in high-performance NVMe SAN arrays.

• RoCE v1/v2 & RDMA: Hardware-based RDMA over Converged Ethernet (RoCE) and native InfiniBand Verbs, ensuring sub-0.8 microsecond ($<0.8 \mu\text{s}$) latency with kernel bypass.

• GPUDirect RDMA & PeerDirect®: Direct communication pathway between GPUs and the network interface card, essential for multi-node AI training workloads (e.g., NVIDIA HGX/DGX architectures).

• ASAP² Direct (Accelerated Switch and Packet Processing): Highly efficient vSwitch/vRouter hardware offloading (including OVS/Open vSwitch) for VXLAN, NVGRE, GENEVE, and MPLS virtualization overlays without burning CPU cycles.

• Packet Handling: Capable of up to 215 million packets per second (Mpps).

High-Performance Computing (HPC) Optimization

• Tag Matching and Rendezvous Offloads: Hardware-based MPI (Message Passing Interface) offloads.

• Adaptive Routing: Automatically bypasses network congestion on reliable transport.

• Burst Buffer Offload: Speeds up checkpointing operations for long-running HPC workloads.

4. Power & Environmental Specs

• Power Consumption:

  • Typical (Passive Cables): ~16.2 W

  • Maximum (Passive Cables): ~19.0 W

  • Maximum (with 1.5W Active Cables/Optical): ~22.3 W

  • Maximum (with 2.5W Active Cables/Optical): ~24.6 W

  • Max Power allocated per QSFP28 port: 5W

• Operating Temperature: $0^\circ\text{C}$ to $55^\circ\text{C}$ ($32^\circ\text{F}$ to $131^\circ\text{F}$)

• Regulatory Compliance:

  • Safety: CB / cTUVus / CE

  • EMC: CE / FCC / VCCI / ICES / RCM

  • Environmental: RoHS compliant

5. Protocol and IEEE Standards Supported

• InfiniBand Standards: IBTA v1.3a compliant, Auto-negotiation, Congestion Control.

• Ethernet Standards:

  • 100G: 100GBASE-CR4, 100GBASE-KR4, 100GBASE-SR4

  • 40G/50G: 40GBASE-CR4, 40GBASE-KR4, 40GBASE-SR4, 40GBASE-LR4, 50GBASE-R2, 50GBASE-R4

  • 10G/25G: 25GBASE-R, 20GBASE-KR2, 10GBASE-LR, 10GBASE-SR, 10GBASE-CR, 10GBASE-KR, SGMII, 1000BASE-CX

  • Additional Specifications: IEEE 802.1Q (VLAN), 802.1p (Priority), 802.3ad (LACP), 802.1ax, 802.1qbb (PFC), 802.1qaz (ETS), 802.1qbg, 1588v2 (PTP).