NVIDIA Mellanox® MCX683105AN-HDAT — ConnectX-6 DE 200Gb/s HDR InfiniBand & 200GbE Single-Port QSFP56 PCIe 4.0 x16 Adapter

Technical Identity
The NVIDIA Mellanox® MCX683105AN-HDAT is a ConnectX-6 DE (Data-center Edition) single-port host channel adapter delivering 200 Gb/s HDR InfiniBand and, on the Ethernet side, 200 GbE over a single QSFP56 cage. The card seats in a PCIe 4.0 x16 slot, providing the full 256 Gb/s of host bandwidth needed to sustain line-rate HDR without back-pressure. Built on the Mellanox ConnectX-6 ASIC, it offloads RDMA verbs, RoCE v1/v2, NVMe-oF, GPUDirect RDMA, SR-IOV, ASAP² Open vSwitch acceleration, and hardware T10-DIF — the standard tool-set for HPC, GPU clusters, and high-throughput storage fabrics. New, factory-sealed, bench-tested by T.E.S IT-SOLUTIONS.

Specifications

OEM: NVIDIA Mellanox
Part Number (MPN): MCX683105AN-HDAT
Product Family: ConnectX-6 DE (Data-center Edition)
ASIC: Mellanox ConnectX-6
InfiniBand Generation: HDR — 200 Gb/s per port
Ethernet Speed: 200 GbE (RoCE v1/v2 capable)
Port Count: Single-Port
Connector: QSFP56 (200G-capable)
Host Interface: PCIe Gen4 x16 (backward-compatible with PCIe 3.0 x16 at reduced bandwidth)
Form Factor: Low-profile PCIe (tall bracket included; low-profile bracket sold separately)
Cooling: Passive heatsink — requires ≥200 LFM server-grade airflow across the card
Condition: New, factory-sealed


What This Is NOT

NOT a switch. This is a host-side PCIe adapter — it terminates one link into one server; it does not forward traffic between multiple hosts.
NOT a dual-port card. One QSFP56 cage only. If you need port redundancy or aggregation in a single slot, look at MCX653106A-HDAT (dual-port VPI) instead.
NOT a ConnectX-7 / NDR (400G) adapter. Maximum line rate is 200 Gb/s; do not deploy this where 400 Gb/s NDR is the target.
NOT a SmartNIC / DPU. No ARM cores, no on-board OS — this is a classic NIC. For programmable data-plane offload use BlueField-2 or BlueField-3.
NOT a Socket Direct card. Single PCIe x16 connector — does not split into 2× PCIe x16 across two CPU sockets.
NOT bundled with QSFP56 transceivers, AOC/DAC cables, or low-profile bracket — purchase these separately as your topology requires.


Compatibility

Drivers: NVIDIA MLNX_OFED for Linux (recommended for production HPC/AI), in-box mlx5_core on upstream Linux kernels ≥4.x, WinOF-2 for Windows Server, nmlx5 for VMware ESXi.
Recommended IB switches: NVIDIA Quantum HDR (e.g., QM8700, QM8790) — full 200 Gb/s end-to-end. Quantum-2 NDR switches (MQM9700 series) downshift to HDR on this card via OSFP→QSFP56 cabling.
Recommended Ethernet switches: NVIDIA Spectrum-2 / Spectrum-3 (SN3000 / SN4000 series) at 200 GbE over QSFP56.
Compatible cabling: QSFP56 DAC up to ~2 m, QSFP56 AOC up to 100 m, QSFP56 transceivers over MMF/SMF. Splitter cables (1× QSFP56 → 2× QSFP56 100G) supported per fabric design.
Server platforms: Any server with a PCIe 4.0 x16 slot — Dell PowerEdge R/C series (Gen4), HPE ProLiant Gen10 Plus / Gen11, Supermicro X12/H12, Lenovo ThinkSystem SR series.
GPUDirect RDMA: Validated with NVIDIA Tesla / A-series / H-series GPUs over PCIe 4.0 root complex.


Use Cases

HPC and AI/ML training clusters using HDR InfiniBand fat-tree or dragonfly topologies.
GPU compute nodes requiring GPUDirect RDMA into NVMe-oF or peer-GPU memory.
NVMe-oF storage initiators / targets at 200 Gb/s line rate.
Virtualized data centers using SR-IOV and ASAP² hardware Open vSwitch.
Software-defined networking with VXLAN / NVGRE / Geneve encapsulation offload.


Frequently Asked Questions
Q: Is the MCX683105AN-HDAT a VPI card — can it run both InfiniBand and Ethernet?
A: Yes. As a ConnectX-6 DE adapter the card supports VPI dual-protocol operation — 200 Gb/s HDR InfiniBand or 200 GbE — selectable per port via the driver. Re-link the port in the desired protocol after configuration.
Q: What PCIe slot is required?
A: PCIe 4.0 x16. The card is backward-compatible with PCIe 3.0 x16, but in a Gen3 slot the host bandwidth is roughly 128 Gb/s — you will not reach full 200 Gb/s HDR line rate.
Q: Which switches pair with this adapter?
A: For HDR InfiniBand, NVIDIA Quantum (QM87xx). For 200 GbE Ethernet, NVIDIA Spectrum-2 / Spectrum-3 (SN3000 / SN4000). The card itself is NOT a switch and does not belong in any switch category — it is the host-side endpoint.
Q: What cables and transceivers does it accept?
A: QSFP56 DAC (passive copper, ≤2 m), QSFP56 AOC (active optical, up to 100 m), and QSFP56 SR4/LR4 transceivers over multimode or single-mode fiber. Use NVIDIA LinkX-branded cables for fully validated link margin.
Q: Does it support GPUDirect RDMA for AI/ML workloads?
A: Yes — GPUDirect RDMA is supported between this adapter and NVIDIA datacenter GPUs sharing the same PCIe 4.0 root complex, eliminating CPU-mediated copies in HPC and AI training pipelines.
Q: Is this new or refurbished?
A: New, factory-sealed. Bench-tested by T.E.S IT-SOLUTIONS for link establishment and PCIe enumeration prior to dispatch.

Why T.E.S IT-SOLUTIONS
Every MCX683105AN-HDAT is bench-tested before dispatch: PCIe enumeration, firmware identification, port link-up at HDR rate against a Quantum reference switch, and cosmetic inspection. Anti-static packaging, EU stock, global B2B shipping, and direct technical consultation for HPC / AI cluster topology design. Volume pricing and matched-batch firmware available on request.