10 Gigabit Ethernet for Edge AI: Why 2.5GbE Is No Longer Enough

10 Gigabit Ethernet (10GbE) is Ethernet running at 10 Gbit/s, ten times Gigabit and four times 2.5GbE, over copper twisted pair (10GBASE-T), fibre or direct-attach cable in SFP+ ports. At the edge it matters because modern AI systems handle several 4K video streams, LiDAR or radar feeds and high-rate sensor telemetry at once, and together those inputs begin to saturate a 2.5GbE link well before the compute is being stretched.

When the network becomes the limit, more GPU performance stops turning into more useful work: the system looks under-used on paper while the application feels slow. This guide sets out the Ethernet speeds side by side, compares 10GBASE-T with SFP+, works through the bandwidth that real edge workloads need, and explains what a computer needs to run 10GbE at full rate. It keeps the original argument of this article, why 2.5GbE is a transitional step for edge AI and what 10GbE with RDMA changes, and ends with a migration plan, a checklist and examples from our range.

The table compares the Ethernet speeds used in industrial and edge computers. Reach figures are the IEEE limits for the stated cable; power figures are typical per port and vary with the controller and the transceiver.

Our FAQ on what 10GbE networking is gives the short definition. The rest of this guide is about when 10GbE earns its place at the edge and how to build it into an industrial system.

The workload profile at the edge has moved on. A single node now routinely handles several 4K video streams, LiDAR or radar feeds, high-frequency telemetry from industrial sensors, and an inference stage that needs all of it presented cleanly and in order. Each of those inputs on its own is manageable. Together, they begin to saturate a 2.5GbE link well before the compute side is being stretched.

Latency matters as much as throughput

Throughput gets most of the attention, but latency is often the real constraint in edge AI. Machine vision lines, robotics and closed-loop control systems are sensitive not to average bandwidth but to the tail of the latency distribution: the occasional slow frame, the jittery packet, the queue that backs up under load. Intermediate Ethernet speeds help with throughput but do not address the underlying architectural issue, which is CPU involvement in every stage of the data path.

Edge AI is becoming multi-node

The single-box inference engine is no longer the whole picture. Edge deployments increasingly look like small clusters: a capture node, a pre-processing node, an inference node, and often a local storage or gateway node, all expected to talk to each other in real time. Once the architecture looks like that, networking becomes infrastructure rather than a peripheral, and the choice of link speed has to be made with that in mind.

2.5GbE has a legitimate role. It is a sensible upgrade over copper, it reuses existing Cat5e cabling, and it gives a meaningful lift over 1GbE for gateway and aggregation tasks. For a lot of general-purpose industrial workloads it is the right answer; our FAQ on whether 2.5GbE is fast enough for edge AI sets out where it is enough.

None of this makes 2.5GbE the wrong choice for every project. It does make it a weak choice for projects being built seriously for edge AI over the next three to five years.

10GbE gives edge systems enough headroom to carry high-resolution video, large AI datasets and multi-sensor telemetry together without the link becoming the constraint. In practice the inference stage stops waiting on data, and the system can run closer to its designed performance.

Compressed video is lighter than people expect; uncompressed machine vision and bulk data movement are heavier. The table shows typical loads and the link each needs. Usable throughput is below the line rate, about 0.94 Gbit/s on 1GbE, 2.35 Gbit/s on 2.5GbE and 9.4 Gbit/s on 10GbE, and it is wise to plan for no more than about 70% sustained use.

Compressed surveillance video rarely needs 10GbE at the camera, as our guide to PoE NVR systems shows with its storage examples; the codec trade-off is in our FAQ on H.264 vs H.265. It is uncompressed GigE Vision cameras, aggregation of many streams, multi-node traffic and data offload that push systems to 10GbE.

Both have their place. 10GBASE-T runs over RJ45 and twisted-pair copper, so it works with structured cabling and negotiates down to slower devices. An SFP+ port takes a plug-in module: fibre for long or electrically noisy runs, or a direct-attach copper cable for short links inside a cabinet. Our FAQs on SFP+ vs RJ45 for 10GbE and SFP cover the basics.

For fanless edge computers, heat is the deciding factor more often than anyone expects: a 10GBASE-T PHY can draw several watts per port that the chassis has to lose without a fan. SFP+ tends to be the more practical choice for industrial deployments, with longer reach over fibre, immunity to interference and the flexibility to mix copper and fibre. RJ45 10GbE remains a reasonable option for shorter indoor runs where cabling is already in place. Our FAQ on single-mode vs multimode fibre helps choose the optics.

Do I need Cat6a for 10GbE?

For 100 m over copper, yes: 10GBASE-T is specified for Cat6a. Cat6 supports 10GBASE-T up to about 55 m, depending on the installation and alien crosstalk, and Cat5e is not specified for 10GbE at all, although short runs sometimes work. Shielded cable helps in electrically noisy plant, provided it is earthed correctly. Where the run is long, crosses buildings or passes heavy machinery, fibre in an SFP+ port avoids the question.

A 10GbE link is only as fast as the slowest part of the path: the computer’s PCIe lanes, the network controller, the switch and the software stack all have to keep up.

Our guide to PCIe generations and bandwidth gives the lane figures, and our FAQ on network interface cards covers adding ports with a card. Where the computer has no onboard 10GbE, a PCIe card adds two or four ports.

10GbE on its own is a meaningful upgrade. It becomes more interesting when paired with RDMA and RoCEv2, which is where much of the practical edge AI benefit comes from. Remote Direct Memory Access lets one system read from or write to another system’s memory without routing the data through the CPU at either end. RoCEv2, RDMA over Converged Ethernet version 2, brings that capability onto a routed Ethernet network, which is what most industrial infrastructure is.

In a single-node system the benefit is real but modest. In a multi-node edge cluster it is often the difference between a system that meets its latency budget and one that quietly misses it. RoCEv2 needs network controllers that support it and switches configured for it, usually with priority flow control or explicit congestion notification and QoS for the RDMA traffic class; a SmartNIC can take on more of that work.

These are the edge workloads where we see 10GbE specified most often.

Our last-mile fibre verification case study shows a small industrial PC working with multi-gigabit links in the field.

The network around a 10GbE edge computer matters almost as much as the computer itself. Three points are worth settling early: RDMA compatibility across switches and network cards, because RoCEv2 benefits disappear quickly on infrastructure that does not support it cleanly; quality-of-service configuration for latency-sensitive traffic, especially in mixed OT and IT networks; and headroom for bandwidth growth, so the network is not re-designed every time a camera or sensor is added. A managed switch with 10G uplinks is usually the starting point, and where traffic crosses subnets our guide to Layer 2 vs Layer 3 switches explains where routing belongs.

Industrial realities still apply

A faster link does not remove the usual industrial requirements. Thermal performance, vibration and shock tolerance, wide-temperature operation and mechanical fit in real enclosures still matter. 10GbE does not change the list of things an industrial computer has to do; it raises the expectation of what it has to carry while doing them. Where traffic must also arrive on a schedule, Time-Sensitive Networking adds bounded latency, and in vehicles single pair Ethernet is bringing multi-gigabit links to sensors over one pair.

For most systems, a move to 10GbE is not a rip-and-replace programme. Our FAQ on when to upgrade from 2.5GbE to 10GbE lists the triggers; a reasonable sequence looks like this.

Done in that order, the transition is an engineering exercise, not a disruption. Looking further ahead, 10GbE is becoming the practical baseline for serious edge AI work, with 25GbE already in use for higher-density systems, and edge computers are moving towards tighter integration of compute, storage and networking.

Examples from our catalogue. Figures are from each product’s page.

We configure, image and test the computers we supply, with the network cards fitted and the drivers installed. Antaira designs and builds its switches; we supply and specify them, pre-configuring IP addressing and VLANs where you ask. Browse our edge AI computers, industrial GPU computers, Ethernet and PoE cards and rack mount industrial switches for more 10GbE options.