ConnectX-6 Dx vs ConnectX-7 vs Intel E810: Choosing a Server Network Adapter

By All Chip Supply · Published Oct 20, 2026

A server network adapter (NIC) decides how fast data moves between servers, storage and the rest of the network. For 25GbE to 400Gb/s, three families come up in most data-center builds: NVIDIA ConnectX-6 Dx, NVIDIA ConnectX-7 and Intel Ethernet 800 Series (E810). This guide compares them in practical terms and explains the part-number details that cause the most ordering mistakes.

The short version

• NVIDIA ConnectX-6 Dx: a mature, widely deployed Ethernet adapter for 25GbE, 50GbE and 100GbE, on PCIe 4.0. A common choice for general servers, storage and virtualization.
• NVIDIA ConnectX-7: the newer generation, up to 400Gb/s, on PCIe 5.0. Available in Ethernet and InfiniBand variants. Common in AI and HPC clusters and in servers that need 200Gb/s or 400Gb/s per port.
• Intel E810: Intel's 100GbE-class Ethernet adapter on PCIe 4.0, available in 25GbE and 100GbE configurations. A frequent choice in Intel-based platforms and for telecom and edge workloads.

All three are well supported by mainstream Linux distributions and hypervisors, but always check the driver and firmware support for your exact operating system version.

Speed and port configuration

Start with the port speed your switches support and the number of ports per server.

• 25GbE (SFP28) remains a common access speed for general-purpose servers.
• 100GbE (QSFP28) is typical for storage, virtualization hosts and leaf-spine uplinks.
• 200GbE and 400GbE (QSFP56, QSFP112 or OSFP, depending on the card) are used where AI training, HPC or very dense storage need more bandwidth.

A dual-port card does not always deliver full line rate on both ports at the same time. Total throughput is limited by the PCIe slot, so check the card's PCIe generation and lane width against what you need.

PCIe generation and slot

• ConnectX-6 Dx and E810 cards are PCIe 4.0 devices.
• ConnectX-7 cards are PCIe 5.0 devices and need a PCIe 5.0 x16 slot for full 400Gb/s throughput. They work in PCIe 4.0 slots, but at reduced maximum bandwidth.

Current server platforms such as AMD EPYC 9005 and Intel Xeon 6 provide PCIe 5.0 lanes. See our EPYC 9005 vs Xeon 6 guide for platform choices.

Form factor: PCIe card or OCP 3.0

Many adapters come in two physical formats:

• Standard PCIe cards, either full-height or low-profile bracket. Check which bracket your chassis needs; some part numbers ship with only one.
• OCP 3.0 cards, a mezzanine format used by many modern servers from major vendors. An OCP 3.0 card does not fit a PCIe slot, and vice versa.

OCP 3.0 cards also differ in how they latch into the server (for example pull-tab or internal-lock versions). Match the variant your server documentation lists.

Ethernet, InfiniBand or VPI

ConnectX-7 is sold as Ethernet-only, InfiniBand-only and VPI (Virtual Protocol Interconnect) versions that can run either protocol. InfiniBand networks need InfiniBand switches and cables. If your cluster is Ethernet-based, order an Ethernet part. This is one of the most common and expensive ordering errors, so confirm it on every quote.

Offloads and security features

All three families offer hardware offloads that reduce CPU load: RDMA over Converged Ethernet (RoCE), SR-IOV for virtual machines, and stateless offloads. Some models add hardware crypto offload (for example inline IPsec or TLS) and secure boot features. Crypto-enabled and non-crypto versions often have different part numbers, so state what you need. The exact feature set depends on the part number; check the product brief.

Reading part numbers

NVIDIA (formerly Mellanox) part numbers start with MCX and encode the family, port count, speed and options. For example, adapters in the ConnectX-6 Dx range start with MCX623, and ConnectX-7 adapters start with MCX75. The suffix identifies details such as crypto, secure boot and bracket. Intel E810 adapters are identified by names such as E810-CQDA2 (dual-port 100GbE QSFP28) or E810-XXVDA2 (dual-port 25GbE SFP28), plus the OCP versions.

Two cards with nearly identical names can differ in crypto support, bracket height or protocol. Quote by the full part number from the manufacturer's product brief.

OEM-branded adapters

Server vendors often sell their own versions of these adapters with vendor-specific firmware. Those cards may be required for full support in that vendor's servers, and generic retail cards may be required elsewhere. If you run a vendor's servers under a support contract, check which part numbers the vendor lists before ordering.

Cables and optics

The adapter is only half of the link. Plan cables and transceivers at the same time:

• DAC (direct-attach copper) cables are low-cost and suitable for short runs within a rack.
• AOC (active optical cables) cover longer distances inside a data hall.
• Optical transceivers with fiber are used for longer runs and between rooms.

Some switches and adapters only accept coded or qualified optics. Confirm compatibility for both ends of the link.

A checklist before you request a quote

• Port speed and number of ports per server
• Switch model or port type at the other end
• PCIe generation of the slot, or OCP 3.0 version
• Bracket: full-height or low-profile
• Protocol: Ethernet, InfiniBand or VPI
• Crypto and secure boot requirements
• Cables or optics needed, with lengths
• Quantity, condition (new or refurbished), delivery country and date

How All Chip Supply can help

All Chip Supply quotes NVIDIA ConnectX-6 Dx and ConnectX-7 adapters, Intel E810 adapters and 100G and 400G optics for integrators and data-center buyers worldwide. Each quote confirms the exact part number, condition and lead time in writing. Some high-performance networking products are subject to export regulations for certain destinations, and we confirm compliance before any order. Browse networking products or request a quote.

Specifications vary by part number and firmware. Always confirm details against the manufacturer's current product brief before ordering.