The Analog Front End (AFE)
Digital logic works with what the front end hands it. It cannot add signal that is no longer there.
By the time a stream reaches a logic block, whatever the cable did to it has already happened. A downstream FIFO can absorb lane skew, and a decision-feedback stage can work with what is still above the noise floor, but neither restores what has fallen below it. Whatever the analog front end (AFE) does not deliver, the logic cannot undo. Four circuits do that work here — three on the receive side, one on the transmit side — and all four were designed in-house.
At these rates, a glitch on a transition is one of the errors that shows up first. A common approach removes the glitch at half rate, buying decision time at the cost of a demultiplexed, duplicated path. Our glitch remover works on the full-rate stream instead. It is the one circuit on this page covered by a granted patent.
A cable and a board strip the high-frequency content out of a signal. A two-stage continuous-time equalizer, with the pole and zero of its transfer function set independently, provides up to 15 dB of high-frequency boost (peaking) against cable loss, at the frequency stated in the datasheet, shaped to the channel rather than applied flat. Linear matters here: the equalizer amplifies what the channel left behind, rather than reconstructing edges the channel destroyed.
On the transmit side the timing has to be generated rather than recovered, and it comes from an on-chip PLL locked to an external reference. A negative-impedance oscillator reaches lock with two delay stages where a conventional ring oscillator typically needs four, which shortens lock time. On a test line, each unit costs the link one lock, and lock time is test time.
None of these four circuits was licensed. There is no third-party PHY IP in the signal path, no per-unit royalty on the interface, and no vendor to wait for when a customer asks why a number is what it is. That last one is the part that is hard to buy. A specification can be matched from a datasheet. The silicon behavior behind it has to be characterized over corners and process splits, and that work cannot be shortened.