English 
| Model Number: | |
|---|---|
| Chipset: | |
| Host Interface: | |
| Network Ports: | |
| Data Rate: | |
| InfiniBand Speed: | |
| Item Specification Model Number MCX556A-ECAT Chipset NVIDIA Mellanox ConnectX-5 Host Interface PCIe 3.0 x16 Network Ports 2 × QSFP28 Data Rate 1/10/25/40/50/100GbE InfiniBand Speed EDR 100Gb/s Supported Protocols: | |
| Quantity: | |
MCX556A-ECAT
Mellanox
The MCX556A-ECAT is a dual-port Virtual Protocol Interconnect (VPI) adapter built on the NVIDIA Mellanox ConnectX-5 controller. It supports both EDR InfiniBand and 100GbE Ethernet protocols on the same hardware, with two QSFP28 connectors and a PCIe 3.0 ×16 host interface. This adapter is designed for data center and compute clusters where workloads require flexible protocol support alongside consistent throughput and low-latency data movement.
Industrial clients deploying AI training clusters, high-performance computing environments, distributed storage fabrics, and virtualized server infrastructure can use this adapter to consolidate network and storage traffic on a single card, reducing the number of peripheral slots required per server.
Parameter | Value |
Part Number | MCX556A-ECAT |
Controller | NVIDIA Mellanox ConnectX-5 |
Host Interface | PCIe 3.0 ×16, 8.0GT/s per lane |
Ports | 2 × QSFP28 |
InfiniBand Speed | SDR, DDR, QDR, FDR10, FDR, EDR (100Gb/s) |
Ethernet Speed | 1/10/25/40/50/100GbE |
Protocol Support | InfiniBand (IBTA v1.3), Ethernet, RoCE v1/v2 |
Virtualization | SR-IOV |
Storage Acceleration | NVMe-oF offload |
Typical Power (Passive Cable) | 14.2W |
Maximum Power (Passive Cable) | 17.0W |
Operating Temperature | 0°C to 55°C |
Form Factor | PCIe low-profile, 68.90mm × 142.24mm |
Compliance | RoHS, CB safety certification |
The adapter implements Remote Direct Memory Access (RDMA) over both InfiniBand and Ethernet (RoCE) fabrics. RDMA enables data to transfer directly between application memory across nodes without intermediate CPU processing, which lowers CPU utilization and reduces end-to-end latency for message-passing workloads. RoCE v2 extends this capability over routable Layer 3 networks, supporting deployment across standard Ethernet topologies.
Built-in NVMe over Fabrics (NVMe-oF) acceleration offloads storage protocol processing from the host CPU to the adapter hardware. This offload mechanism maintains storage access latency levels comparable to direct-attached storage even when storage arrays are accessed across the network. Additional offload features include TCP/UDP checksum calculation, TSO, LRO, and stateless offloads for common network protocols.
Single Root I/O Virtualization (SR-IOV) support allows one physical adapter to present multiple virtual functions to guest operating systems. Each virtual function receives dedicated bandwidth and hardware resources, enabling consistent network performance across virtualized tenants in cloud and multi-tenant server deployments.
The dual-protocol design of the MCX556A-ECAT fits infrastructure teams that operate mixed-fabric environments or plan to migrate between InfiniBand and Ethernet. Common use cases include AI and GPU training clusters using InfiniBand for inter-node communication, HPC clusters running MPI-based workloads, scale-out storage clusters using NVMe-oF or iSER, and enterprise data centers consolidating multiple network workloads onto a single adapter.
Commercial procurement teams can standardize on this adapter model across server fleets with different fabric requirements, reducing spare part inventory and simplifying driver management.
The MCX556A-ECAT works with Linux, Windows Server, VMware ESXi, and FreeBSD operating systems. It is compatible with PCIe 3.0 slots and operates at reduced lane widths when installed in x8 or x4 slots. The adapter supports standard QSFP28 optical transceivers and direct attach copper cables for both Ethernet and InfiniBand operation.
Yes. The VPI architecture allows each port to be configured independently for either InfiniBand or Ethernet operation, enabling mixed-protocol deployments on the same adapter.
The adapter uses a PCIe 3.0 ×16 edge connector. It can operate in x8 or x4 slots with corresponding reductions in maximum available bandwidth.
Yes. SR-IOV support enables the adapter to be partitioned into multiple virtual functions for direct assignment to virtual machines in hypervisor environments.
