Front-End vs Back-End Networking for an AI Server: When an Intel X710 Is Enough and When You Need ConnectX

A single AI inference server needs only a front-end network: 10GbE for user requests, model downloads, storage and management. An Intel X710 covers it. You need a back-end GPU fabric, with RDMA over NVIDIA ConnectX adapters, only when one model or training job is split across several servers. In a PCIe Gen3 server like the Dell R740xd or HPE DL380 Gen10, the sensible back-end card is 100G, because a Gen3 x16 slot can't feed 200G.

Here is how to tell which network you need, which card fits each, and what to cable them with.

Two Networks, Two Jobs

• Front end (north-south): API requests in, tokens out, model files pulled from a NAS or object store, SSH, monitoring and backups. It is ordinary TCP/IP traffic. Bandwidth matters mostly when loading models.
• Back end (east-west): GPU-to-GPU traffic between servers when a model or training job is split across them. Frameworks such as NCCL exchange activations or gradients on every step, so latency and bandwidth set the speed of the whole job. It runs on RDMA (RoCE v2 over Ethernet, or InfiniBand), ideally with GPUDirect RDMA so the NIC reads GPU memory directly without going through the CPU.

The out-of-band management port (iDRAC or iLO) is a third, separate link, and 1GbE is fine there.

When One Server Needs No Back End at All

If the model fits in the GPUs inside one server, there is no back-end traffic. Two A40s in the same R740xd talk over PCIe, not the network. A single 48GB A40 runs a 70B model at 4-bit; two cards in one chassis give 96GB (what fits in 48GB). Running several independent inference servers behind a load balancer doesn't need a back end either; each server answers its own requests.

For that server, 10GbE is plenty. Generated text is a few kilobytes per second per request. The only big transfer is loading weights: at about 1.1 GB/s, a 40GB model loads from a NAS in roughly 35–40 seconds. Our R740xd GPU builds already have Intel X710 10GbE SFP+ ports on the network daughter card, so they need no extra front-end card.

When You Do Need ConnectX

• A model too big for one server, split with tensor or pipeline parallelism across nodes (vLLM and others support multi-node serving).
• Distributed training or fine-tuning across servers, where gradients are exchanged on every step.
• RDMA storage such as NVMe over Fabrics, or Ceph and parallel file systems tuned for RDMA, feeding several GPU nodes.

NVIDIA's guidance for certified multi-node systems starts at 200 Gb/s per node. In a PCIe Gen3 server you won't get that from one slot: a Gen3 x16 slot carries about 126 Gb/s. That is why a 200G ConnectX-6 tops out near 100G in an R740xd, and why a 100G ConnectX-5 is the sweet spot for these servers.

Which Card for Which Job

RoCE needs a lossless Ethernet setup on the switch (PFC and ECN) to perform well; see lossless Ethernet for RoCE. VPI cards run as InfiniBand or Ethernet per port; here's how to switch them with mlxconfig. To decode any ConnectX part number, see How to Read a ConnectX Part Number.

Slots: Plan the Back End Around the GPUs

In an R740xd, the GPU risers take most of the full-height slots, so plan the back-end card's slot before you buy GPUs. Put the ConnectX card on the same processor as the GPUs it serves where you can, so GPUDirect RDMA traffic doesn't cross between sockets. Since the front end lives on the network daughter card in our builds, the only add-in NIC is the back-end card.

Cables for Each Network

• 10GbE front end: SFP+ DACs to a top-of-rack switch, or 10GBASE-T copper with the X550-T2. Intel 700-series cards check module coding, so order the NIC end coded for Intel; we can program a DAC with different coding on each end.
• 25GbE back end: the MCP2M00-A001 25G SFP28 DAC ($29.00) between ConnectX-4 Lx or ConnectX-5 cards and a switch. To fan one 100G switch port out to four 25G servers, use the MCP7F00-A002R30N breakout ($149.95).
• 100G back end: QSFP28 DACs within the rack; see the 100G ConnectX cable guide.

FAQ

Do I need a ConnectX or InfiniBand card for a single AI inference server?
No. If the model fits in the GPUs inside one server, GPU traffic stays on that server's PCIe bus and never touches the network. You only need a front-end card for requests, model downloads and management, and 10GbE is plenty. An Intel X710-DA2 ($49.95 at Resilient Tec) covers it, and Resilient Tec's R740xd GPU builds already include X710 10GbE ports.

When does an AI server need a back-end network?
When one model or one training job is split across more than one server. Then GPUs exchange data across the network on every step, which needs high bandwidth, low latency and RDMA (RoCE or InfiniBand). That is what ConnectX adapters are built for. Distributed training, multi-node fine-tuning and models too large for one server all fall into this group.

Can the Intel X710 do RoCE or RDMA?
No. The Intel X710 and X550 are standard Ethernet adapters without RDMA. They are excellent front-end cards, but GPU-to-GPU traffic between servers needs an RDMA-capable adapter such as an NVIDIA ConnectX card.

Which ConnectX card is best for a PCIe Gen3 server like the R740xd or DL380 Gen10?
A 100G card. A PCIe 3.0 x16 slot carries about 126 Gb/s, so the ConnectX-5 MCX556A-ECAT ($244.95) runs one port at full 100G EDR InfiniBand or 100GbE. A 200G ConnectX-6 tops out near the same speed in a Gen3 slot. For a two-node lab on a budget, the 25GbE ConnectX-4 Lx MCX4121A-ACAT ($54.95) supports RoCE.

Is 10GbE fast enough to load large models?
Yes for inference. 10GbE moves roughly 1.1 GB/s, so a 40GB model loads from a NAS in about 35–40 seconds, once per restart. Tokens returned to users are a few kilobytes per second per request. Move the front end to 25GbE if you reload models often or pull large datasets.

Ready to Buy?

Network adapters and cables for AI servers, refurbished and tested, with a lifetime warranty and free US shipping. Prices checked October 30, 2026.

• Intel X710-DA2 dual-port 10GbE SFP+ — $49.95, in stock at Resilient Tec; front end, SFP+
• Intel X710-DA4 quad-port 10GbE SFP+ — $84.95, in stock at Resilient Tec; front end with separate storage network
• Intel X550-T2 dual-port 10GBASE-T (RJ45) — $94.95, in stock at Resilient Tec; front end on RJ45 copper
• ConnectX-4 Lx MCX4121A-ACAT, 2× 25GbE SFP28 — $54.95, in stock at Resilient Tec; budget RoCE back end
• ConnectX-5 EN MCX512A-ACAT, 2× 25GbE SFP28 — $149.95, in stock at Resilient Tec; 25G back end
• ConnectX-5 VPI MCX556A-ECAT, 2× 100G EDR/100GbE — $244.95, in stock at Resilient Tec; 100G back end for Gen3 servers
• ConnectX-6 VPI MCX653105A-HDAT, 1× 200G HDR — $619.95, in stock at Resilient Tec; 200G for Gen4 servers
• MCP2M00-A001 25G SFP28 DAC, 1m (Mellanox compatible) — $29.00, in stock at Resilient Tec; 25G back-end links
• MCP7F00-A002R30N 100G to 4× 25G breakout DAC, 2m — $149.95, in stock at Resilient Tec; one 100G port to four 25G servers

Planning a multi-node build? Send us the server count and switch and we'll match cards and cables. POs and net terms available.

You Might Also Need

• The complete parts list for an R740xd or DL380 Gen10 AI server — network alongside power, memory and storage
• Why your 200G ConnectX-6 tops out near 100G — PCIe Gen3 explained
• InfiniBand vs Ethernet for AI clusters — choosing the back-end fabric
• Network adapters — Intel, ConnectX and Cisco UCS cards
• Server Parts & Upgrades Guide — every server part in one place

Sources

• NVIDIA — NVIDIA-Certified Systems Configuration Guide (system memory, storage and network sizing)
• NVIDIA — GPUDirect RDMA documentation
• Dell — PowerEdge R740 and R740xd Technical Guide (memory, power supplies, GPU and BOSS options)