Server NIC, Transceiver and DAC Compatibility: What Actually Fits

Every pluggable module carries an EEPROM that describes itself: vendor name, part number, serial number, supported media and reach, plus vendor-specific bytes. Switches and some network cards read that data at insertion and decide what to do with it. This is why 'a 10G SFP+ short-reach module' is not a specification. The same physical optic, programmed with different identity data, is a different part number and behaves differently depending on what it is plugged into.

Third-party modules solve this by being programmed to match a target platform's expected identity — which is why a compatible module's part number names the platform it was coded for, not just the speed. Our catalog rows say so explicitly: a cable described as compatible with a specific Juniper part is coded for Juniper, and one described as compatible with a specific Mellanox part is coded for that. Read the target, not just the number of gigabits. If you order optics without naming the switch and the card, you are ordering a coin flip.

Platform behaviour varies. Some switches refuse an unrecognised module outright and leave the port down. Some raise a log message and pass traffic anyway. Some have a documented override — the widely-known example is a hidden command on certain Cisco platforms that enables unsupported transceivers — which works, is unsupported, and can complicate a support case if the port later misbehaves. Whether your organisation is willing to run in that state is a policy decision that should be made before the modules arrive, not by an engineer at a maintenance window.

As a rule the switch side is stricter than the host side, but OEM-branded server adapters do sometimes enforce their own lists too. When a design spans a switch from one vendor and a card from another, state both on the request so the coding decision is made once, deliberately.

10GBASE-T deserves its own paragraph because it is the option that keeps the cabling you already own. A hundred metres over Cat6A, in the same RJ-45 patch fields as the rest of the plant, with no module to code and no fibre to clean. That is a genuine advantage in a wiring closet, and it is why quad-port 10GBASE-T cards remain a common answer for access-layer aggregation and for connecting servers that live outside the data centre.

The costs are power and latency. A 10GBASE-T PHY draws materially more power per port than an SFP+ direct-attach connection — the difference is watts per port against fractions of a watt — and the line coding adds latency measured in microseconds where direct-attach adds a fraction of one. In a dense rack, multiplied across every port, that power is a real thermal line. For latency-sensitive east-west traffic it is a real design constraint. For a four-port aggregation card in a closet, neither matters much.

One honest word about supply, because it changes the right answer more often than any specification does. The pluggable inventory we hold in depth is direct-attach and active optical assemblies; discrete transceivers are a much thinner pool and a large share of them quote as back-ordered. Where the distance genuinely allows a direct-attach or active optical cable, it is usually also the faster path to a delivery date. Where it does not, plan the lead time deliberately rather than discovering it after the switch is racked — and check the live availability on the product page rather than assuming, because it moves.

A network card cannot move more traffic than its slot can carry, and the arithmetic is simple enough to do on the quote. A PCIe 3.0 lane carries roughly 985 MB/s after encoding overhead, so an x8 slot delivers about 63Gb/s in each direction; PCIe 4.0 doubles it to roughly 126Gb/s at x8. Set that against line rate: a dual-port 25G card needs 50Gb/s and fits a 3.0 x8 slot with headroom. Four ports of 10G needs 40Gb/s and fits the same slot. A single 200G port needs more than a 3.0 x16 slot can supply and only reaches line rate on PCIe 4.0 x16 — which is exactly why high-rate adapters specify a generation as well as a width.

Then there is the trap that catches experienced people: physical width is not electrical width. A slot that is physically x16 may be wired for x8 or x4, and in 1U and 2U servers the riser card determines what each slot actually gets. The owner's manual riser diagram is the authority — not the look of the connector, and not the slot number. Putting a card that needs eight lanes into a physically-x16, electrically-x4 riser slot produces a card that works, benchmarks badly, and generates a support ticket blaming the driver.

Slot generation matters the same way. A PCIe 4.0 card in a 3.0 slot negotiates down to 3.0 rates. For a dual 25G card that is invisible. For anything at 100G or above it is the difference between line rate and a permanent ceiling. When a design mixes new adapters into an older chassis, check the generation of the specific slot, because on many servers not every slot is the same generation.

Breakout is the cheapest capacity you will ever buy and the easiest to order incorrectly. A high-rate switch port can often be split — a 200G port into two 100G, a 100G port into four 25G — using a cable with one connector at one end and two or four at the other. Two conditions apply. The switch has to support the split on that specific port, which is usually a port-group setting affecting neighbouring ports as well as the one you are configuring, and the cable has to be the breakout part rather than the straight one. The host side sees ordinary ports and needs no special configuration, which is why the mistake is nearly always made on the switch or the purchase order rather than on the server.

Brackets are the most boring cause of a failed installation. Cards ship with a full-height bracket fitted and a low-profile one in the box, or the reverse, or only one of the two. A 1U server needs low profile. If the listing does not say which brackets are included, ask before the card is racked, because the alternative is a card that is physically correct and cannot be secured.

Firmware and driver are a matched pair on high-rate adapters, not two independent updates. A card that links, passes a ping and then falls over under sustained load is far more often a firmware-and-driver mismatch than a cable fault. Update through the server vendor's own update utility where one exists, so the adapter firmware and the operating-system driver move together, and record the versions with the build. This is also the first thing to check when one host in an otherwise identical cluster is the slow one.

Finally, measure the cable path rather than the straight line. Direct-attach lengths are fixed at manufacture, so the run has to be right the first time: up the rack, through the vertical manager, along the tray and back down, plus service loop. A three-metre cable specified from a floor plan is regularly a four-metre cable in the rack. Where the run is genuinely marginal, an active copper or active optical assembly buys the margin that a passive cable cannot.