Yes, but not by plugging it in. A QSFP28 port on an SN2700-CS2FC is four 25G lanes bonded into one 100G link, and a QSFP28-to-4xSFP28 breakout cable splits those lanes into four independent 25GbE ports. Per-host bandwidth drops to 25G, and the cable and FEC settings decide whether the link comes up.
What Is Actually Inside a 100GbE Port
A QSFP28 cage is not one pipe. It is four parallel lanes, each running 25 Gb/s, and 100GbE is what you get when the switch bonds all four. That is why a 100G port can be carved up at all: the hardware underneath already counts in 25G units. SFP28, the connector on a 25GbE network card, is exactly one of those lanes.
The two connectors are electrically related and mechanically different. QSFP28 uses a wider cage and a 38-pin edge; SFP28 uses the smaller SFP cage and a 20-pin edge. You cannot seat one in the other.
Three Ways to Put 25GbE on a 100GbE Port
The supported path is a breakout assembly: one QSFP28 shell at the switch, four SFP28 shells at the server end, each carrying a single lane. Passive copper breakout is a short-reach part, roughly 2 to 3 metres, enough inside one rack and not enough across an aisle. Active optical breakout goes further, at a higher cost.
A second path exists. A QSFP-to-SFP adapter, usually called a QSA, lets a single SFP28 module sit in a QSFP cage and run at 25G on one lane. Whether it works depends on the switch platform and how the ASIC configures that cage, so confirm it for your exact model.
| Port mode | Lanes used | Switch-end connector | Hosts served | Where it fits |
|---|---|---|---|---|
| 1x 100GbE | 4 | 1x QSFP28 | 1 | Leaf uplinks, storage targets, GPU nodes |
| 2x 50GbE | 2 + 2 | 1x QSFP28 to 2x QSFP28 | 2 | Mid-density racks with 50G adapters |
| 4x 25GbE | 1 each | 1x QSFP28 to 4x SFP28 | 4 | Server access with 25G adapters |
| 1x 40GbE | 4 at 10G | 1x QSFP28 or QSFP+ | 1 | Older 40G gear, cage must stay whole |
The server side does not need to know it is on a split port. A dual-port 25GbE adapter such as the MCX4121A-ACAT or the MCX512A-ACAT presents two independent SFP28 ports, each taking its own leg. Two adapters in two servers fill four legs and consume one cage.
What the Split Actually Costs
Aggregate bandwidth is unchanged: 100G in, 100G out. What changes is the shape. A host that would have received 100G now receives 25G, four hosts each receive 25G, and the cage is fully consumed, leaving no trunk on that port for an uplink.
That catches people building small clusters. A 32-port SN2700 gives you 32 cages. Split eight for servers and you have handed out 32 access ports and left 24 cages for uplinks and storage. Plan the uplink count before the split count.
- Reach. Passive copper breakout tops out near 2 to 3 metres; beyond that you need active optical breakout, at roughly double the cost per port.
- Port groups. On many 100G platforms cages are arranged in pairs, and splitting one can consume or restrict the cage beside it. Read the port map for the exact model.
- No 40G on a split cage. A split cage runs 25G per lane; 40GbE uses all four lanes at 10G each and needs the cage whole.
- Per-host ceiling. 25G is about 3.1 GB/s of line rate, roughly 2.9 GB/s after overhead. Fine for a compute node, tight for an NVMe-oF target serving several initiators.
Where the Link Refuses to Come Up
Three settings account for most dead 25G links on split ports, and none produces an error naming the cause.
FEC. 25GbE as defined in 802.3by expects RS-FEC on the link. If one end is set to the older BASE-R forward error correction and the other to RS-FEC, or one end is left on auto and lands on the wrong one, the port stays down and the log says very little. Set it explicitly on both ends.
Speed. 25G does not negotiate the way 1G and 10G do. The switch sub-interface and the adapter port both need to be set to 25G by hand, and a sub-interface created by a split is sometimes left administratively down, which looks identical to a bad cable.
Coding. Copper cables and optical modules carry a vendor code in their EEPROM, and some adapter and switch combinations decline to link on a code they do not recognise. This is the most common reason a batch of breakout cables works on one switch and fails on another. Test one assembly end to end before buying forty.
When You Should Not Split
If the host really needs 100G, do not split. GPU nodes pulling training data and NVMe-oF initiators both push past 25G under load, and a split turns a bandwidth question into a topology problem. Put an MCX516A-CCAT in the server and run one 100G cable to an unsplit cage; we covered the PCIe side in MCX516A-CCAT versus CDAT.
If your rack is small and will stay at 25G, a 100G switch is more cage than you need today. If there is any chance of 100G uplinks or storage later, buying the 100G box now and splitting it for access is usually cheaper than replacing a 25G-only switch. The denser SN3800-CS2F and the SN3700-CS2FC give you more cages to split; the differences between the Spectrum models are in our SN3700, SN3800 and SN2100 notes.
Splitting is also wrong for two servers connected back to back, where a direct cable beats a switch, and for a fabric where every host already runs 100G. Mixing adapter generations on the same fabric is a separate decision, covered in our ConnectX generation notes.
On a used switch, check three things before committing to a split layout: that the port map allows it, that the installed firmware exposes the split setting, and that one breakout assembly links at 25G on a server you already have. Availability is confirmed per part number against the models listed on zhongxindata.hk, so send the switch and adapter models together.
Frequently Asked Questions
Can a 25GbE card connect to a 100GbE switch?
Yes, where the switch supports breakout. A 100GbE switch with QSFP28 cages can present a 25GbE link on a compatible port, and the 25GbE card negotiates down to it. Both the switch model and the port configuration matter, so we confirm the combination against both part numbers before you order.
What optics do I need for a 25GbE port on a 100GbE switch?
Normally a 25GbE SFP28 optic on the server side, and whatever the switch expects for that port when it presents 25GbE through a breakout cage. The exact requirement depends on the switch’s breakout modes, so tell us the switch model and the port and we will confirm the optic types that work.
Will a 100GbE switch be slower for a 25GbE server?
The link runs at 25GbE, so the server sees 25GbE end to end. What changes is the port economics: breakout can consume more switch ports per server than a native 25GbE port would, and the switch may carry a higher per-port price. For a small number of servers that is a fine trade.
Should I buy a 25GbE switch instead?
If you have several servers to connect at 25GbE, a dedicated 25GbE switch removes the breakout complexity and usually costs less per port. If it is one or two servers, down-negotiating on the 100GbE switch you already have is the cheaper and simpler path. Send us the server count and we will recommend the option that costs less overall.
Does down-negotiation need any special configuration?
Usually the switch needs the port configured for the breakout mode, and some platforms disable it by default. We will confirm the configuration for the switch model. On the adapter side, the card negotiates the rate automatically once the link comes up.
How do you verify a used 25GbE or 100GbE adapter?
We clean, re-flash and bench-test on our own test line, then confirm the part number, port count, negotiated link rate and firmware level, and we read the error counters so a card that links but errors does not get listed. Ask for photos of the exact unit and its firmware level before you commit.
Can I plug an SFP28 module straight into a QSFP28 port?
No. The cages differ in size and pin count. You need a breakout assembly or a QSFP-to-SFP adapter, and adapter support varies by platform.
Does a breakout cable add latency?
No meaningful amount. Passive copper breakout is a wiring fan-out with no active component, and optical breakout adds conversion delay in nanoseconds.
Can I split a 100G port into 2x50G instead of 4x25G?
Yes, on most 100G platforms, using two lanes per leg. It needs 50G-capable adapters on the server side, which are less common in used stock than 25G parts.
Do all four legs need the same cable type?
Within one assembly, yes: a breakout cable carries four legs of the same type and length. The servers on those legs can use different adapters.
Will a 40GbE adapter work on a split port?
No. A 40G adapter uses all four lanes at 10G each, so the cage has to stay whole. Splitting removes 40G mode entirely.
Send the part number and quantity and we will check availability. WhatsApp +86 13068616914 · zhongxindata.hk
Yes, in most cases — the mechanism is breakout
- A 100GbE switch with QSFP28 cages can present a 25GbE link on a compatible port
- The card negotiates down to the rate the port offers, so no special card is required
- Both sides must support the path, which is why the switch model matters as much as the card
Cabling and optics that work for a down-negotiated link
- 25GbE optics in SFP28 are the normal choice for a 25GbE server port
- When the switch presents 25GbE through a breakout cage, the optic type changes accordingly
- Confirm the breakout mode and the optic the switch expects for that port before ordering
When a dedicated 25GbE switch is the better answer
- If many servers need 25GbE, a 25GbE switch avoids breakout complexity and per-port cost
- Breakout on a high-radix switch consumes more ports per server, which affects the topology
- For a handful of servers, down-negotiating is usually the cheaper and simpler route
Related reading
- MCX512A 25GbE adapter — the 25GbE side of the question
- Cabling guide — optic and cable selection
- QSFP28 vs QSFP-DD vs OSFP — connector detail
- Network adapters — category hub
Frequently Asked Questions
Can a 25GbE card connect to a 100GbE switch?
Yes, where the switch supports breakout. A 100GbE switch with QSFP28 cages can present a 25GbE link on a compatible port, and the 25GbE card negotiates down to it. Both the switch model and the port configuration matter, so we confirm the combination against both part numbers before you order.
What optics do I need for a 25GbE port on a 100GbE switch?
Normally a 25GbE SFP28 optic on the server side, and whatever the switch expects for that port when it presents 25GbE through a breakout cage. The exact requirement depends on the switch’s breakout modes, so tell us the switch model and the port and we will confirm the optic types that work.
Will a 100GbE switch be slower for a 25GbE server?
The link runs at 25GbE, so the server sees 25GbE end to end. What changes is the port economics: breakout can consume more switch ports per server than a native 25GbE port would, and the switch may carry a higher per-port price. For a small number of servers that is a fine trade.
Should I buy a 25GbE switch instead?
If you have several servers to connect at 25GbE, a dedicated 25GbE switch removes the breakout complexity and usually costs less per port. If it is one or two servers, down-negotiating on the 100GbE switch you already have is the cheaper and simpler path. Send us the server count and we will recommend the option that costs less overall.
Does down-negotiation need any special configuration?
Usually the switch needs the port configured for the breakout mode, and some platforms disable it by default. We will confirm the configuration for the switch model. On the adapter side, the card negotiates the rate automatically once the link comes up.
How do you verify a used 25GbE or 100GbE adapter?
We clean, re-flash and bench-test on our own test line, then confirm the part number, port count, negotiated link rate and firmware level, and we read the error counters so a card that links but errors does not get listed. Ask for photos of the exact unit and its firmware level before you commit.