PCI-Express Lane To Lane Deskew Problem

["Alan Crawford" <crawford@tality.com>]

I thought I'd raise my concern over what I see as a problem with the 
Lane to lane deskew method proposed in the PCI-Express specification.

Although the specification does not indicate any specific methodology for 
deskewing the lanes, it does mention the use of the COM symbol
as "an unambiguous de-skew mechanism" (page 185 of Rev 1.0).

This would be alright if it wasn't for the following:

The specification indicates that the system shall have a "Total Skew"
of 20nS which equates to 50 UI which in turn after 8B10B encoding
becomes 5 symbol periods.

Thus the lane to lane deskew shall be capable of aligning lanes 
whose data can be up to 5 symbol periods apart. 

This would lead to the minimum time between alignment symbols
to be of the order of 11 symbol periods or more (definitely more). 

   In XAUI the maximum skew is 55UI  (I think) which leads to 
   removing a skew of 5-6 symbol periods, the standard defines that
   alignment symbols shall occur no closer than every 16 symbols
   ensuring that if one lane was skewed by the maximum (plus
   or minus) then the receiver can still correctly align the lanes.

However in PCI-Express the COM symbols can occur 4 symbol
periods apart (two successive SKIP sequences or a SKIP
sequence followed by a TS1/TS2 sequence).
This means that COM symbols from different ordered-sets can
be aligned. Yes, the last COM symbols received would flag a 
mis-aligned link so the lanes can then be re-aligned but this does 
not guarantee that the lanes shall always end up aligned.

Does anyone have any input which can clarify this issue?

Alan


============================================
Alan Crawford         Cadence Design Foundry
Chief Consulting Eng       1 the Alba Campus
(+44) 1506 595 067                Livingston
                              United Kingdom
============================================