Server NIC selection goes wrong in a predictable way: a buyer falls for a headline rate, buys the card, and then discovers the switch fabric, the PCIe slot or the fiber plant disagrees. This guide inverts the process. It walks the five decisions in the order they actually constrain each other, using the tiers in our server network card range as reference points.
Decision 1: the fabric you already own
A NIC cannot outrun its switch. If your top-of-rack switches are 10G SFP+, then 10G cards are the correct purchase today regardless of what the server could theoretically use, because the link negotiates to the slower end. If the fabric is being refreshed in the same project, the question opens up — and the industry answer for new general-purpose fabrics is 25G per server port, because the per-port premium over 10G is now small while the bandwidth headroom is 2.5×. Reserve 100G per server for the nodes that demonstrably move that much data: storage nodes, backup targets, HPC and AI nodes, and virtualisation hosts aggregating many tenants.
Decision 2: the slot budget
Check the slot before the card. A 100G adapter needs a PCIe x16 slot, and for full line rate on both ports it needs PCIe 4.0 electrical width — a mechanical x16 slot wired at x8, or a Gen 3.0 slot, will quietly cap your expensive card. The PCIe card page maps our range by slot width and generation. Also check bracket height: low-profile chassis need low-profile brackets, a detail that has delayed more installs than any silicon shortage.
Decision 3: media and optics
Copper or fiber is decided by distance and plant. Within a rack, DAC assemblies are the cheapest 10G/25G interconnect ever made. Between racks, or wherever isolation matters, you are on optics — which means the card choice becomes a cage choice (SFP+, SFP28, QSFP28) and the module budget joins the card budget. Total the link, not the line items: card plus two modules plus cabling is the real unit price of an optical port.
Decision 4: the features behind the rate
Two cards with the same rate can be different products. Virtualisation offloads (SR-IOV, VMDQ), tunnel offloads (VXLAN/NVGRE), RDMA support and jumbo-frame ceilings separate data-centre adapters from desktop-grade parts in server clothing. Inline appliances add a harder requirement — bypass behaviour — which is a hardware property, not a feature flag: if the server is a security appliance in the path of traffic, you need a bypass NIC, not a fast regular one.
Decision 5: the platform list
Finally, the unglamorous check that prevents returns: operating system and driver support for your exact stack, and — for domestic-platform projects — CPU adaptation and ecosystem verification. Our flagship models publish their OS lists on their product pages; the FF-1102E, for example, documents support across mainstream and domestic Linux distributions because its G810-2Q controller was designed for those platforms.
The short version
- Fabric is 10G → buy 10G cards, spend the savings on optics quality.
- Fabric refresh → 25G per server is the default; 100G for proven heavy nodes.
- Security appliance in-path → bypass hardware, rate second.
- Always: verify slot width/generation, bracket, OS list, and price the whole link.
If you send us your switch model, server model and node role, we will return the shortlist with the link priced as a whole. That is the normal way these are bought, and it is what the inquiry form is for.
Is 25G worth it over 10G for ordinary virtualisation hosts?
In a fabric refresh, generally yes: the incremental cost per port is modest and the headroom covers the host’s whole service life. In a 10G fabric, no — the switch is the ceiling.
Can one server use both 25G and 100G cards?
Yes, if the slot budget allows: a common pattern is 2×25G for production traffic and a 100G card for storage or backup networks.
The mistake audit: five ways this decision goes wrong
- Buying the card before the fabric: a 25G card in a 10G fabric is a 10G card with a receipt.
- Counting mechanical slots instead of electrical lanes: an x16-shaped slot wired x8 caps a 100G card at half its promise.
- Pricing the card and forgetting the link: optics and cabling can exceed the adapter’s price at the higher tiers.
- Ignoring the OS list until install day: the finest controller is worthless without a driver for your exact stack.
- Specifying rate and forgetting role: storage, backup and security nodes need features — RDMA, offloads, bypass — that the rate figure does not describe.
Run those five against any shortlist and most bad purchases die on paper, where they are cheap. What survives is usually a short list of two or three defensible options, separated by price and platform fit rather than by correctness — which is exactly where a selection process should end. If a vendor will not discuss the five with you, that silence is itself selection data.
Should every server in a refresh get the same NIC?
Standardising on one or two models pays in drivers, spares and sanity. Deviate only where a role proves it needs more — storage and security nodes are the usual exceptions.
Total cost of the port, worked honestly
Compare two candidate designs the way a CFO would force you to: cost per connected server port, fully loaded. Design A keeps the 10G fabric: card, two SFP+ modules (or one DAC per server in-rack), no switch changes. Design B refreshes to 25G: card at a modest premium, SFP28 optics or DAC, and the switch uplift amortised across every port it serves — the number buyers forget to divide. In many real refreshes Design B’s per-port delta lands surprisingly small once the switch is divided by its ports and the design’s service life; in others, where the switch uplift serves few servers, Design A wins by a distance. The method matters more than either answer: any NIC recommendation that quotes the card alone is quoting the smallest third of the decision.
Add the line nobody prices until it bites: spares. One spare card and a handful of optics per model in the fleet is the professional ratio; a fleet standardised on one card model carries one spares shelf, while a fleet of five models carries five. Standardisation is a cost feature, and it belongs in the comparison table beside the rate.
Postscript for the undecided: if two tiers both survive the audit, choose the one whose optics and cabling your team already stocks and understands. Operational familiarity is a real specification — it just never appears on the datasheet. A network the team knows deeply at 10G will outperform, in lived reliability, a faster network they are still learning to read.