What Is Forward Error Correction?_
FEC encodes extra parity bits into the data stream before transmission, allowing the receiver to reconstruct corrupted bits up to a certain threshold. This reduces the need for retransmission in high-speed networks where bit errors are more common. It is essential for maintaining link integrity over long fiber runs or noisy copper connections.
Technical Details
FEC is typically implemented at the physical layer (PHY) in Ethernet and InfiniBand transceivers, using algorithms like Reed-Solomon or Fire Code. The overhead is defined by the relevant standard (e.g., RS-FEC for 400GbE adds about 5.8% overhead). In GPU networking, FEC is enabled on all scale-out links (InfiniBand/Ethernet) to handle signal degradation from long cables or optical transceivers. NVLink copper spine inside the rack generally does not use FEC because the short, shielded connections have lower error rates.
How Leviathan Systems Works with Forward Error Correction
During commissioning, we verify FEC is enabled on all switch-to-GPU fiber links per the cluster design, as misconfigured FEC can cause link flaps or degraded throughput. We also check that FEC settings match between the NIC and switch ports to avoid negotiation failures.
Related service
Network Testing & Commissioning →