NVIDIA ConnectX-6 InfiniBand Adapter MCX653105A-HDAT – 200Gb/s RDMA Hardware Encryption

Engineered for demanding HPC, AI, and hyperscale cloud infrastructures, the NVIDIA® ConnectX®-6 MCX653105A-HDAT smart adapter card delivers 200Gb/s bandwidth on a single QSFP56 port with In-Network Computing acceleration. Offloading critical data movement and security tasks from the CPU, it dramatically improves efficiency, scalability, and security — from deep neural network training to real-time data analytics and NVMe-oF storage.

As a core component of the NVIDIA Quantum InfiniBand platform, ConnectX-6 enables end-to-end RDMA, hardware-based reliable transport, and advanced congestion control. The MCX653105A-HDAT single-port model offers a cost-effective yet powerful entry into 200Gb/s HDR InfiniBand and 200GbE networks. It integrates block-level XTS-AES encryption, NVMe over Fabrics (NVMe-oF) offloads, and GPUDirect RDMA acceleration — making it the ideal choice for GPU-accelerated clusters, software-defined storage, and high-frequency trading environments where low latency and security are paramount.

NVIDIA ConnectX-6 extends Remote Direct Memory Access (RDMA) beyond conventional limits. By implementing hardware offloads for MPI tag matching, out-of-order RDMA supporting Adaptive Routing, and Dynamically Connected Transport (DCT), it ensures efficient scaling across thousands of nodes. The adapter's In-Network Memory capability enables registration-free RDMA memory access, reducing software overhead. Combined with PCIe Gen 4.0, data moves directly between memory and network, freeing CPU cycles for application logic. Additionally, hardware-based reliable transport and end-to-end packet-level flow control guarantee data integrity even under extreme load.

With support for RoCE (RDMA over Converged Ethernet), overlay network tunneling offloads, and intelligent interrupt coalescence, ConnectX-6 provides a unified smart fabric for both InfiniBand and Ethernet environments.

• High Performance Computing (HPC): Large-scale simulations, weather modeling, computational chemistry requiring deterministic low latency.
• AI & Machine Learning: Accelerate distributed training of deep neural networks with GPUDirect RDMA and 200Gb/s throughput.
• NVMe-oF Storage Arrays: Build high-performance NVMe/TCP or NVMe/RDMA storage targets with hardware offloads, reducing CPU load by up to 30%.
• Hyperscale Cloud & NFV: Efficient service chaining, OVS offload (ASAP²), and SR-IOV for up to 1K virtual functions.
• Financial Services & Trading: Ultra-low latency market data distribution and algorithmic trading infrastructure.

ConnectX-6 MCX653105A-HDAT is compatible with a wide range of servers, switches, and OS environments. It supports InfiniBand switches up to 200Gb/s (HDR) and Ethernet switches up to 200Gb/s with auto-negotiation and all FEC modes. The adapter works across x86, Power, Arm, GPU, and FPGA-based platforms.

Note: For cold plate variants for liquid-cooled Intel Server System D50TNP, please contact Starsurge for customized ordering.

Hong Kong Starsurge Group provides end-to-end support: from compatibility checks, firmware customization, to on-site deployment guidance. We offer dedicated technical account managers for data center upgrades and proof-of-concept (PoC) testing. All adapters are shipped with anti-static packaging and optional installation kits.
Support offerings include: 24/5 engineering support ticketing system, advanced replacement for business-critical environments, driver & software stack assistance (OFED, WinOF-2, DPDK), and remote troubleshooting.

• Confirm server mechanical clearance: standard height PCIe bracket included; low-profile bracket also provided as accessory.
• For liquid-cooled platforms (Intel D50TNP), verify cold plate option availability before ordering — standard air-cooled bracket is default.
• Please confirm driver compatibility with your Linux distribution version — NVIDIA MLNX_OFED recommended for optimal performance.
• Not publicly specified: Exact power consumption at full 200Gb/s load — refer to NVIDIA user manual or contact Starsurge for typical values (approx 13-16W).
• FIPS certification is hardware-capable but may require specific firmware version — notify sales team if FIPS compliance is mandatory for your deployment.
• Single-port adapter cannot be used for port redundancy; for high-availability designs, consider dual-port MCX653106A-HDAT.

• ☑ PCIe slot availability: x16 mechanical (electrical x16/x8/x4 supported)
• ☑ Required port speed: 200Gb/s or lower; cable type (QSFP56 passive DAC or active optics)
• ☑ OS driver version: Check MLNX_OFED or WinOF-2 compatibility with your kernel
• ☑ Encryption requirement: Standard AES-XTS or FIPS mode (special firmware)
• ☑ Cooling and bracket: Standard air-cooled or cold plate option for liquid cooling
• ☑ Quantity and lead time: Stock confirmation with Starsurge sales team
• ☑ Intended application: Single-port sufficient, or need dual-port redundancy?

• ▸ NVIDIA ConnectX-6 User Manual (Firmware & Configuration Guide)
• ▸ RDMA over Converged Ethernet (RoCE) Deployment Best Practices
• ▸ NVMe-oF with ConnectX-6: Performance Tuning Guide
• ▸ GPUDirect RDMA for AI Clusters – Technical Overview
• ▸ Block-level Encryption Setup for FIPS Environments