NVIDIA MCX555A-ECAT 100Gb/s Single-Port QSFP28 InfiniBand Adapter PCIe 3.0 x16

The NVIDIA ConnectX-5 MCX555A-ECAT is a single-port 100Gb/s InfiniBand adapter card in a low-profile PCIe form factor. Leveraging the proven ConnectX-5 architecture, it delivers up to 100Gb/s throughput with sub-microsecond latency and a high message rate. The card supports both InfiniBand (up to EDR) and 100GbE, providing versatile connectivity for high-performance computing, storage, and virtualized environments.

Built with an embedded PCIe switch and advanced RDMA capabilities, the MCX555A-ECAT offloads critical communication tasks from the CPU — enabling higher application performance, lower power consumption, and reduced total cost of ownership. It is fully compatible with PCIe 3.0 x16 slots and supports a wide range of operating systems and acceleration frameworks.

The ConnectX-5 architecture integrates a range of hardware acceleration engines that reduce CPU intervention and improve application scalability:

• MPI Tag Matching & Rendezvous Offload: Offloads message matching and rendezvous protocol processing, dramatically improving MPI performance for HPC clusters.
• Out-of-Order RDMA with Adaptive Routing: Enables efficient use of multiple network paths while maintaining ordered completion semantics, maximizing fabric utilization.
• NVMe-oF Target Offloads: Allows NVMe storage systems to serve remote access with near-zero CPU overhead, ideal for disaggregated storage architectures.
• Dynamically Connected Transport (DCT): Provides extreme scalability for large compute and storage systems by eliminating connection setup overhead.
• ASAP2 Accelerated Switching & Packet Processing: Hardware offload for Open vSwitch (OVS) and overlay network tunneling (VXLAN, NVGRE, GENEVE).
• On-Demand Paging (ODP): Supports virtual memory paging for RDMA operations, simplifying application development.

• High-Performance Computing (HPC): Ideal for supercomputing clusters, MPI-based simulations, and scientific research workloads requiring low latency and high message rates.
• AI & Deep Learning Training: Combined with GPUDirect RDMA, enables fast GPU-to-GPU communication across nodes, accelerating training times.
• NVMe-oF Storage Systems: Deploy as storage targets or initiators in NVMe over Fabrics environments for high-throughput, low-latency block storage access.
• Cloud & Virtualized Data Centers: SR-IOV and virtualization offloads support multi-tenant environments with guaranteed QoS and secure isolation.
• High-Frequency Trading (HFT): Ultra-low latency and hardware timestamping (IEEE 1588v2) meet the demands of financial services applications.

The MCX555A-ECAT is designed for broad compatibility with NVIDIA InfiniBand switches (e.g., Quantum, Spectrum) and third-party 100GbE switches. It supports both passive copper DACs and active optical cables via QSFP28 ports.

Operating Systems & Software Stacks:

• RHEL / CentOS, Ubuntu, Windows Server, FreeBSD, VMware ESXi
• OpenFabrics Enterprise Distribution (OFED) / WinOF-2
• NVIDIA HPC-X, OpenMPI, MVAPICH2, Intel MPI, Platform MPI
• Data Plane Development Kit (DPDK) for kernel bypass

Note: Specifications derived from NVIDIA ConnectX-5 product documentation. For the latest details and firmware support, refer to official NVIDIA release notes.

For OCP or Multi-Host variants, please contact sales. All cards support backward compatibility to lower speeds.

• Superior Application Performance: Hardware offloads for MPI, NVMe-oF, and overlays free CPU cores for business logic.
• Scalable RDMA Fabric: DCT, XRC, and out-of-order RDMA deliver linear scalability for thousands of nodes.
• GPU Acceleration Ready: GPUDirect RDMA enables direct memory access between GPUs and network adapters, eliminating CPU bottlenecks in AI clusters.
• Flexible Deployment: Single QSFP28 port simplifies cabling and is ideal for 100Gb/s leaf-spine architectures.
• Investment Protection: Support for both InfiniBand and Ethernet allows seamless transition between protocols as needs evolve.

Hong Kong Starsurge Group provides complete lifecycle support for NVIDIA ConnectX-5 adapters, including pre-sales configuration assistance, firmware update guidance, and warranty service. Our technical team can help with:

• Compatibility verification with your server and switch infrastructure
• Performance tuning for HPC or storage workloads
• Custom bracket options and bulk packaging requirements
• RMA processing and advanced replacement services

Contact our sales engineers for volume pricing and lead time information.

Electrostatic Discharge (ESD): Always use ESD-safe practices when handling the adapter. Store in anti-static packaging until installation. Cooling Requirements: Ensure adequate airflow in the server chassis to maintain operating temperature within specified range. Firmware Updates: Use NVIDIA official firmware tools (MFT) and verify compatibility with your OS and driver version before updating. Cable Bending: Follow QSFP28 cable bend radius guidelines to avoid signal degradation.

This is a Class A product. In a residential environment it may cause radio interference. Ensure proper shielding and grounding per local regulations.

• ☑ Confirm server has an available PCIe 3.0 x16 (or x8) slot with adequate clearance.
• ☑ Determine port count: single-port (MCX555A-ECAT) vs dual-port (MCX556A-ECAT).
• ☑ Choose cable type: passive copper DAC for short distances (≤5m) or optical for longer runs.
• ☑ Verify operating system and driver support (OFED, Windows, VMware).
• ☑ For GPU clusters, ensure GPUDirect RDMA compatibility with your GPU model and driver version.
• ☑ Check if tall or short bracket is required for your server chassis.

• NVIDIA MCX556A-ECAT – Dual-port 100Gb/s ConnectX-5 adapter
• NVIDIA MCX556A-EDAT – ConnectX-5 Ex with PCIe 4.0 support
• NVIDIA Quantum-2 QM9700 40-port 800Gb/s InfiniBand Switch
• Mellanox QSFP28 passive DAC cables (1m, 2m, 3m)
• NVIDIA Spectrum-4 SN5600 100GbE/400GbE Ethernet switches

• NVIDIA ConnectX-5 InfiniBand Adapter Card User Manual
• RDMA over Converged Ethernet (RoCE) Deployment Guide
• GPUDirect RDMA Best Practices for AI Clusters
• NVMe over Fabrics with ConnectX-5 – Configuration Guide
• OFED Installation and Tuning Guide