NVIDIA MCX653106A-HDAT-SP ConnectX-6 VPI Adapter Card HDR/200GbE

The NVIDIA MCX653106A-HDAT-SP is a high-performance, data-center-grade network adapter built for demanding HPC, AI, and storage workloads. This ConnectX-6 VPI (Virtual Protocol Interconnect) card brings together HDR InfiniBand and 200GbE networking on a single PCIe 4.0 x16 card, delivering ultra-low latency, extreme bandwidth, and flexible fabric interoperability. With a tall bracket design and SP packaging indicating a single-pack SKU, this adapter is optimized for dense server deployments, multi-node clusters, and accelerator-rich servers where every microsecond of latency and every gigabit of throughput matters. It’s engineered to scale with your infrastructure—whether you’re building a bare-metal HPC cluster, a GPU-accelerated AI training farm, or a NVMe over Fabrics storage fabric—while simplifying cabling and reducing I/O complexity through intelligent virtualization and protocol offloads.

• The MCX653106A-HDAT-SP anchors its value in a dual-port QSFP56 interface, offering up to 200 Gb/s per port across HDR InfiniBand and Ethernet 200GbE. This means you can deploy ultra-fast InfiniBand HDR links for low-latency HPC workloads and, on the same card, sustain high-throughput 200GbE Ethernet traffic for diverse services. The dual-port design enables network fabric aggregation, link-level redundancy, and the flexibility to segment traffic across separate fabrics or to run multi-protocol workloads side by side, all without swapping hardware. This level of integration reduces the number of separate NICs required per server and simplifies cabling in large racks where space and airflow are at a premium.
• Powered by NVIDIA’s ConnectX-6 VPI technology, the card provides robust RDMA and offload capabilities that dramatically decrease CPU overhead, accelerate data movement, and enable NVMe over Fabrics (NVMe-oF) along with RoCEv2, iSCSI, and other common streamlining protocols. The VPI architecture lets administrators configure the fabric to meet changing production needs—switching between InfiniBand HDR and Ethernet 200GbE on demand, or running multiple protocols on a single physical adapter. For workloads like AI model training, large memory transfers, and distributed data processing, this flexibility translates into smoother scaling, lower end-to-end latency, and more predictable performance across clusters and storage networks.
• From a performance and reliability perspective, this card is designed for data-center environments. The PCIe 4.0 x16 interface provides the bandwidth headroom required by two 200Gbps ports, ensuring that neither port becomes a bottleneck as workloads scale. The tall bracket accommodates full-height server chassis and blade servers often found in HPC and enterprise data centers, while the SP packaging supports scalable deployment across a growing fleet of servers. NVIDIA’s firmware and driver stack—dedicated to ConnectX-6 VPI—delivers consistent performance, compatibility with major Linux distributions and Windows Server ecosystems, and ongoing updates to keep pace with evolving fabric technologies. Thermal design and robust components help it operate reliably under sustained loads typical of HPC clusters and data-center fabrics, reducing maintenance downtime and supporting longer maintenance intervals between interventions.
• In terms of integration and management, the MCX653106A-HDAT-SP is engineered to play well with existing network fabrics and orchestration tools. It supports standard management interfaces, vendor-provided drivers, and firmware upgrade pathways that administrators rely on for consistent performance tuning, fabric provisioning, and firmware security updates. SR-IOV and virtualization-friendly features enable multi-tenant environments, allowing multiple virtual machines or containers to share the same physical NIC without sacrificing throughput or isolation. This is particularly valuable in labs and hyperscale deployments where resource utilization and partitioning are critical. The card’s architecture also reduces the number of separate NICs and switches needed in a data center, leading to simpler cabling, fewer points of failure, and easier scaling as workloads evolve from scientific simulations to real-time analytics and AI inference pipelines.
• Finally, the value proposition for enterprises is clear: a single, flexible, multi-protocol adapter capable of supporting high-speed InfiniBand HDR interconnects and 200GbE Ethernet traffic, with VPI-enabled protocol interoperability, PCIe 4.0 x16 bandwidth, and a form factor suitable for dense rack deployments. The -SP Single Pack packaging makes it straightforward to scale a fleet of servers, while NVIDIA’s ongoing driver and firmware support helps ensure compatibility with current-generation data center software stacks, from distributed training frameworks to storage fabrics and virtualization platforms. Whether you’re constructing a cutting-edge HPC cluster, building a hybrid cloud storage fabric, or accelerating AI workloads with GPU-rich nodes, this card provides the foundational interconnect performance, reliability, and flexibility required for modern data centers.

Techncal Details of NVIDIA MCX653106A-HDAT-SP

• Model: MCX653106A-HDAT-SP
• Product family: ConnectX-6 VPI
• Interconnects: HDR InfiniBand and 200GbE
• Ports: Dual-Port QSFP56
• PCIe interface: PCIe 4.0 x16
• Form factor: Tall Bracket
• Packaging: SP (Single Pack)
• Protocol support: Virtual Protocol Interconnect (VPI) enabling multi-fabric interoperability
• Performance: Up to ~200 Gb/s per port (dual-port); up to ~400 Gb/s aggregate bandwidth across both ports, depending on fabric and configuration
• Drivers and OS support: NVIDIA ConnectX-6 VPI drivers for Linux and Windows Server; firmware updates supported
• Ideal deployments: Data centers, high-performance computing clusters, AI/ML training and inference farms, NVMe-oF storage fabrics, and virtualization-enabled environments
• Management: SR-IOV capable, virtualization-friendly features, and vendor management tooling for fabric provisioning and monitoring

How to install NVIDIA MCX653106A-HDAT-SP

• Step 1: Prepare your server and environment. Power down completely, unplug the power cords, and discharge any static electricity. Verify you have an available PCIe 4.0 x16 slot with enough space for a full-height, tall-bracket adapter, and ensure the chassis airflow won’t be blocked by the card or cables.
• Step 2: Install the adapter. Insert the ConnectX-6 VPI card into the PCIe slot, applying even pressure until the connector seats securely. Align the card’s bracket with the mounting holes and fasten the bracket to the chassis with a screw to keep the card stable during operation and cable routing.
• Step 3: Connect the network fabric. Attach QSFP56 cables to the card’s HDR InfiniBand and/or 200GbE ports as required by your fabric design. Route cables neatly to minimize strain, interference, and airflow obstruction. If you’re deploying both InfiniBand and Ethernet fabrics, label cables to prevent cross-connection issues during maintenance.
• Step 4: Power on and install drivers. Boot the system and install the NVIDIA ConnectX-6 VPI drivers from the official NVIDIA repository or your operating system’s package manager. Follow the on-screen prompts to complete the driver and firmware installation. If firmware updates are suggested, apply them to ensure compatibility with your fabric firmware and switches.
• Step 5: Validate the installation. After rebooting, verify that the interfaces are up using appropriate network tools (for example, ip link show, ifconfig, ethtool, or ibstat). Run basic throughput tests and latency measurements to confirm that the expected bandwidth is available and that RDMA paths are operating correctly. Document the results for performance baselines and future comparisons.

Frequently asked questions

• Q: What networks does this card support?
  A: The NVIDIA MCX653106A-HDAT-SP supports HDR InfiniBand and 200GbE networking via dual QSFP56 ports, enabling ultra-low-latency HPC interconnects and high-throughput Ethernet workloads on a single adapter.
• Q: Which PCIe slot is required?
  A: A PCIe 4.0 x16 slot is recommended to maximize bandwidth across both ports; using a lower-speed slot may work but with reduced performance.
• Q: Can it be used in virtualized environments?
  A: Yes. ConnectX-6 VPI cards provide virtualization features and SR-IOV support, allowing multi-tenant networking with strong isolation and high performance.
• Q: What does -SP mean?
  A: -SP indicates a Single Pack SKU, designed for scalable deployment in data-center environments.
• Q: What cabling is required?
  A: You’ll need QSFP56 cables compatible with HDR InfiniBand and 200GbE to achieve the card’s full performance potential.