---
title: "Error correction on fast links"
question: "How big an error can a fast link fix by itself?"
url: "https://nearfieldnotes.com/videos/error-correction/"
site: "Near Field Notes"
video: "https://nearfieldnotes.com/media/error-correction/full.mp4"
duration: "2:55"
date: "2026-11-17"
note: "https://nearfieldnotes.com/notes/error-correction/"
---

# Error correction on fast links (video)

How big an error can a fast link fix by itself?

2:55, with sound. [Read Note 44](https://nearfieldnotes.com/notes/error-correction/) for the full answer, with figures, conditions and sources.

## Transcript

A 400G Ethernet link is allowed to get 100 million bits wrong every second.

That link is in spec.

PCIe 6 allows 64,000 a second on every lane. Thunderbolt 5, a few hundred.

Here's how links fix their own mistakes, and how big a hit each one can take, from one bit up to a scratch on a CD.

For decades, a link aimed for one wrong bit in a trillion. A CRC caught it, and the packet was sent again.

Then links went to four voltage levels. The levels sit three times closer, and noise crosses them all the time.

With no repair, 19 of every 20 full-size frames would arrive damaged. So the receiver fixes the data itself.

The idea is from 1947. Richard Hamming's weekend jobs at Bell Labs kept dying, because the computer could detect an error but not fix it.

His answer was check bits that overlap.

Each check covers a different group. Flip any one bit, and the checks that fail point straight at it.

Flip it back, and the data is right again.

Every camera packet header still works this way, and so does every DDR5 memory chip.

But fast links don't fail one bit at a time. The equalizer feeds each decision into the next one. In my simulation, after one wrong bit, the next is wrong half the time.

So in 1960, Reed and Solomon built a code that works on whole symbols.

Treat the data as points on a curve, and send a few extra points. A damaged point falls off the curve.

Every two extra symbols fix one bad symbol, anywhere in the block.

And a symbol with one wrong bit costs the same as a symbol with every bit wrong. That's why it handles bursts.

USB4 adds 4 check bytes to every 194. That fixes any two bad bytes.

And a precoder turns a burst of any length into just two wrong bits: one where it starts, one where it ends.

DisplayPort added correction in 2016, along with compression. It's required whenever the picture is compressed.

PCIe 6 can't wait around. Its code is tiny: three groups, each fixing one byte, in under 2 ns.

The groups take turns, so a burst of 16 bits lands one byte in each.

With two bad bytes in one group, a code this small makes a wrong fix a third of the time. So a CRC checks every flit, and a bad one is sent again.

Ethernet goes the other way. 30 check symbols on every 514 fix any 15. That's any burst up to 141 bits, and it takes over 100 ns.

At the limit, a block carries about one bad symbol, against a budget of 15. So 100 million errors a second becomes one lost block every 4 hours.

But the edge is a cliff. 4 times more errors, and blocks fail 100 million times more often.

The record still belongs to the compact disc. Two codes, with the data spread out between them, fix a burst of nearly 4,000 bits. That's a scratch 2.5 mm long.

One bit. Nine bits. 141. Nearly 4,000.

So a healthy link is wrong 100 million times a second, and almost none of it reaches you.

---

Source: https://nearfieldnotes.com/videos/error-correction/
