NVIDIA Mellanox MCX623106AN-CDAT Server Adapter Technical Solution
June 10, 2026
This technical white paper provides network architects, pre-sales engineers, and operations teams with a comprehensive solution design centered on the NVIDIA Mellanox MCX623106AN-CDAT. We address how to achieve deterministic low latency and maximize server throughput using RDMA/RoCE in modern data center environments.
1. Project Background & Requirements Analysis
Traditional TCP/IP stacks introduce significant overhead—interrupt handling, memory copies, and context switching—that limits application performance, especially for NVMe-oF, distributed databases, and AI training clusters. Key requirements for next-generation infrastructures include:
- Sub-10µs end-to-end latency for storage and real-time analytics
- CPU bypass to free cores for business logic (target: <20% network-related CPU usage)
- Lossless Ethernet fabric supporting PFC and ECN across 100GbE leaf-spine topologies
- In-line encryption without performance degradation (IPsec/TLS)
- Seamless integration with existing automation frameworks (Ansible, Terraform)
2. Overall Network & System Architecture
The reference architecture adopts a two-tier leaf-spine topology with 100GbE ports throughout. Each compute/storage node is equipped with the MCX623106AN-CDAT ConnectX adapter PCIe network card. Key design elements:
- Leaf switches: Configured with 4MB shared buffer, PFC priorities 3 (storage) and 4 (AI traffic), ECN marking threshold at 200KB queue depth
- Spine switches: Non-blocking 12.8Tbps fabric, 3:1 oversubscription ratio (optimized for east-west patterns)
- Host configuration: PCIe 4.0 x16 (or PCIe 5.0 when available), 256MB firmware memory per adapter
- RDMA transport: RoCEv2 with DCQCN congestion control, 4K MTU, inline data up to 256 bytes
The MCX623106AN-CDAT Ethernet adapter card solution supports up to 1,024 virtual functions (VFs) and 256 physical functions (PFs), making it ideal for NFV and containerized workloads.
3. Role of the NVIDIA Mellanox MCX623106AN-CDAT & Key Features
The NVIDIA Mellanox MCX623106AN-CDAT serves as the data plane engine, offloading all network processing from the host CPU. According to the MCX623106AN-CDAT datasheet and official MCX623106AN-CDAT specifications, critical capabilities include:
| Feature | Benefit for RDMA/RoCE |
|---|---|
| Hardware steering (ASAP²) | Direct packets to specific queues – zero CPU intervention |
| On-chip RoCE congestion management | DCQCN implemented in firmware – 100ns response granularity |
| GPUDirect® RDMA | GPU memory access bypassing host memory – 2µs latency reduction |
| VirtIO acceleration | Native performance for VM/container networking |
The MCX623106AN-CDAT Ethernet adapter card also includes secure boot and flash encryption, meeting stringent security policies for financial and government sectors.
4. Deployment & Expansion Recommendations
For brownfield deployments, the adapter is fully MCX623106AN-CDAT compatible with existing 100GbE switches (broadcom, Mellanox, Cisco) after firmware upgrade. Recommended deployment phases:
- Phase 1 (lab validation): Two storage nodes + two compute nodes, enable per-port PFC and ECN, validate with ib_write_bw (target 98Gb/s throughput per port)
- Phase 2 (pilot): One full rack (40 nodes), deploy DCQCN with ECN thresholds: min=150KB, max=500KB
- Phase 3 (scale-out): Multi-rack with spine layer, monitor global PFC pause frames (keep <0.1% of total traffic)
For expansion, the adapter supports link aggregation (LAG) and adaptive routing (AR) when used with NVIDIA Spectrum switches. Mixed-speed environments (100GbE to 25GbE) require buffer reservations on egress ports.
5. Operations, Monitoring & Troubleshooting
The NVIDIA DOCA framework provides comprehensive telemetry. Essential monitoring metrics include:
- Per-queue counters: rdma_out_of_sequence, cnp_sent, cnp_received (via ethtool -S)
- PCIe health: Correctable/uncorrectable errors, link speed negotiation (mlxlink tool)
- Congestion detection: PFC watchdog timers, ECN marked packet ratio (target <5%)
Common troubleshooting: If RoCE performance drops below 80Gb/s, verify PCIe bifurcation (must be x16), disable TSO/GRO, and confirm switch ECN thresholds. For those evaluating MCX623106AN-CDAT price, the solution typically achieves ROI within 9–12 months by consolidating server count.
6. Summary & Value Assessment
The NVIDIA Mellanox MCX623106AN-CDAT delivers a deterministic, low-latency fabric while reducing CPU overhead by up to 75% compared to software-based networking. This solution guide has detailed architecture design, key features, deployment steps, and operational best practices. Organizations seeking MCX623106AN-CDAT for sale can contact authorized distributors for evaluation units. For complete reference, the official MCX623106AN-CDAT datasheet and detailed MCX623106AN-CDAT specifications are available upon request.

