What every lane in a wide link carries

What the lanes divide between them

A x16 link is not a faster lane, it is sixteen lanes carrying the same conversation at once. Packet data is striped across them byte by byte, and every lane in the bundle carries the same rate, the same clock embedded in its own pair. This page reads the bundle the way one reads an ensemble: each lane has the same instrument, and the width is the number of players.

The striping rule is simpler than it sounds. Bytes of a packet are distributed across the lanes in round-robin order, so lane zero takes bytes 0, 16, 32 on a x16 link, lane one takes 1, 17, 33, and so on. Each lane is a full-duplex channel, two differential pairs, and each runs at the generation's rate on its own. Nothing in the arrangement makes one lane the leader: the parallelism is in the data path, not in a hierarchy of lanes.

The analogy to an ensemble is not a courtesy. A string band works because the roles are divided and each instrument is audible on its own; the Irish desk that publishes the Fifth String journal spends whole pages on what the banjo's fifth string contributes that the other four do not. The lane bundle asks the same question in reverse: in a x16, every string is identical, and the width rather than the voicing is what carries the tune.

How do the lanes stay together?

Because a skew budget is enforced at the receiver. Signals sent together do not arrive together: traces differ in length, vias differ in count, and the fastest and slowest lane of a bundle arrive spread across a measurable window. The specification gives the receiver a deskew budget, a fixed number of symbol times within which it must realign the lanes back into one byte stream. The receiver buffers the early arrivals and waits for the late ones, which is why the adder of a few centimeters of mismatched routing is a design rule, not an exotic failure.

Each lane also carries its own clock. There is no reference distributed on the data pairs: clocking is embedded in the serial stream itself, recovered per lane, and the aggregate clocking of the link is the union of sixteen independent recoveries held within the deskew window. The mechanism underneath is described on the page about the PIPE interface, and the rates each lane runs are listed in the link speed table.

What happens when a lane goes bad?

The bundle narrows rather than fails. The negotiation that sets the width can be re-run by the link itself when a lane proves unreliable, and the result is a link that trains to x8 or x4 instead of x16, still carrying the same striped stream across fewer strings. From software this is visible only as the negotiated width in the status register, which is why a narrow link is the most common surprise on the bench: the socket said x16, the training said otherwise.

That behavior is the design's quietest advantage over a parallel bus. A parallel interface with one bad line loses the whole word; a striped bundle loses one lane and renegotiates. The price is that the deskew and the recovery machinery exist at all, and that the reader of a link must always ask what the width actually negotiated to rather than what the silkscreen claims.

Why sixteen and not more?

Because wider is not free. Every added lane costs two pairs of routing, two layers of board complexity and a deskew window that must still close. The standard historically defined wider widths, x12 and x32 among them, which the reference literature describes as virtually never used; the ecosystem settled on x16 as the widest slot in common practice, with wider effective bandwidth delivered by faster generations rather than by wider bundles. The trade is visible in the rate ladder: when the standard wants more throughput it raises the per-lane rate, not the lane count.

The width of a link is the count of identical voices, and the skill in reading it is hearing the ensemble rather than the loudest string. When the number on the slot and the number in the register disagree, the register is the truth, and the gap between them is the number of lanes that did not make the cut.