Optics

FEC explained: why your 25G or 100G link won't come up (and what RS-FEC costs)

Light levels fine, link still down or full of errors? It's often FEC. Which link needs which FEC, why a mismatch kills the link, and how to read the counters.

5 min readBy Yvan Watchman

Your 25G or 100G link stays down, even though both optics are recognised and Rx power sits comfortably inside the datasheet range. Or it comes up, but you're seeing CRC errors you can't explain. There's a good chance the FEC setting is wrong: it differs between the ends, or is off on a link that needs it. This post covers which link expects which FEC, why a mismatch takes the link down, how to fix it in five steps, and how to use the FEC counters as an early warning.

What FEC does: pre-FEC and post-FEC

Forward error correction adds parity on the transmit side so the receiver can repair bit errors itself, with nothing retransmitted. That lets the raw link be far noisier than older Ethernet would tolerate. A 100GBASE-SR4 link is specified at a BER of 5×10⁻⁵ before correction, the pre-FEC BER. After RS-FEC you're left with a post-FEC BER of 10⁻¹² or better, which is what Ethernet links are expected to deliver.

At 25G and 100G you'll meet two flavours:

  • BASE-R FEC (Clause 74) — also called FC-FEC or Fire code. Lightweight: 32 parity bits per 2080-bit block, correcting bursts of up to 11 bits.
  • RS-FEC (Clauses 91 and 108) — Reed-Solomon RS(528,514) over 10-bit symbols, correcting up to 7 bad symbols per codeword. Clause 91 covers 100G, Clause 108 covers 25G. With PAM4 at 50G per lane (Clause 134 for 50G) you get the heavier RS(544,514), which corrects up to 15 symbols.
LinkFEC per standard or MSA
25G DAC, CA-25G-N (typically up to 3 m)no FEC needed
25G DAC, CA-25G-S (typically up to 3 m)BASE-R FEC or RS-FEC
25G DAC, CA-25G-L (typically up to 5 m)RS-FEC
25GBASE-SRRS-FEC (some SR optics run without, check the datasheet)
100GBASE-SR4 and -CR4RS-FEC, Clause 91
100G CWDM4 and PSM4 (MSA)RS-FEC mandatory on the host
100GBASE-LR4no FEC required

LR4 is the exception that bites most often. The standard doesn't rely on FEC, but platforms handle it differently. Even within one vendor it varies: Juniper's general FEC documentation has RS-FEC off by default for LR4, while many QFX switches turn Clause 91 on by default for every 100G port. Connect two vendors over LR4 and you've got yourself a mismatch. With PAM4 optics, part of the FEC work often happens inside the module; there, the datasheet rules, and don't switch host FEC off unless the datasheet says so.

Both ends have to do exactly the same thing. RS-FEC against BASE-R FEC, or RS-FEC against off, doesn't give you a slow link, it gives you no link. Auto-negotiation can agree FEC on copper, but optics don't auto-negotiate, so you're relying on each switch's defaults, and those differ by vendor. Rule out the basics first with our SFP checklist; if those check out, work through this:

  • Read the active FEC mode on both ends, not just the config. On NX-OS, show interface ethernet1/1 shows both the admin and the oper FEC state; on Linux, use ethtool --show-fec eth0.
  • Look up which FEC your optic or DAC expects in its datasheet.
  • Set both ends to the same fixed value: both RS-FEC, both BASE-R FEC, or both off. Juniper calls them fec91, fec74 and none, Cisco Catalyst cl91 (100G) or cl108 (25G), cl74 and off, ethtool rs, baser and off.
  • Check breakouts separately: many Juniper QFX switches (including QFX5120 and QFX5200) default to Clause 74 on a 4×25G breakout, while a 25GBASE-SR at the far end expects RS-FEC.
  • Check the FEC counters as soon as the link is up.

FEC counters: your early warning

FEC has a cliff. As long as the pre-FEC BER stays within what the code can handle, every error is repaired and the link looks flawless. Past that point it goes quickly: uncorrectable codewords, lost frames, CRC errors. So a dirty connector or an ageing optic shows up in the FEC counters before it shows up in your traffic.

Two counters matter:

  • Corrected codewords — what FEC repaired. On an RS-FEC link this usually climbs, and that's fine. Watch the trend.
  • Uncorrectable codewords — what FEC couldn't repair. Arista's own MIB documentation calls these a sign of an unhealthy link.

Reading them, read-only:

  • Juniper — show interfaces et-0/0/0 extensive | match FEC shows FEC Corrected and Uncorrected Errors with their rates.
  • Arista — show interfaces Ethernet1/1 phy detail shows pre-FEC BER and the codeword counters among other things; check the fields for your EOS release.
  • Linux — ethtool -I --show-fec eth0 shows corrected and uncorrectable blocks, provided your driver reports them.

Record those values when the link is new, just like Rx power from DOM/DDM. If the corrected rate climbs noticeably later on, clean the connectors before you start swapping optics.

What FEC costs, and help with your design

FEC costs latency: RS-FEC adds roughly 100 ns (100G) to 250 ns (25G) per hop, BASE-R FEC less. For office traffic that's nothing; for trading, RDMA or HPC it adds up. The details, and why a short DAC without FEC wins there, are in Latency: why a DAC beats optics. So don't switch FEC off because you can, only when the link demonstrably copes without it.

For your next link, browse 25G SFP28, 100G QSFP28 or DAC cables. Designing a fabric with two switch vendors, breakouts or 400G? That's where the FEC choice gets tight, and 400G, CWDM/DWDM and MPO are available on quote. Send us your design: we'll think it through with you and send a quote. If you'd like the links installed and tested as well, have a look at our services.

Yvan Watchman

About the author

Yvan Watchman

Founder of FiberNxt and Xyphen IT

Yvan Watchman is the founder of FiberNxt and Xyphen IT. He has been building and running networks and data centre infrastructure in and around Amsterdam for years.

LinkedIn profile

Related products and services

Frequently asked questions

Does FEC have to match on both ends of a link?

Yes. Both ends must run the same FEC, or both none. If one side runs RS-FEC and the other runs BASE-R FEC or nothing, the link stays down, even when Rx power looks perfectly healthy.

Does 100GBASE-LR4 need FEC?

Not according to the IEEE standard: LR4 is specified without FEC. CWDM4, PSM4 and SR4 are built around RS-FEC. Whether your switch enables RS-FEC on an LR4 port depends on the platform, so set both ends explicitly.

Are corrected FEC errors a problem?

Not by themselves. On an RS-FEC link a rising count of corrected codewords is normal. Watch the trend: if that count or the pre-FEC BER suddenly climbs, the link is degrading. Uncorrectable codewords mean you are actually losing frames.

Further reading