But can it add negative delay, making outputs appear earlier than they
would have without the SSO?
I have always thought so ... based on the model that much of the effect
is a simple addition of coupled noise into other outputs, which may
appear to move their driven edges in either direction ... depending on
whether all outputs switch in the same direction, or some go the other
Reason for asking: the AGP Specification and Design Guide state that SSO
can only add positive delay. See AGP ECR# 40
(ftp://download.intel.com/pc-supp/platform/ecr40.pdf), which says: "SSO
pushout only increases the delay of a buffer, never decreases it;" and:
"While, all other skews are equally likely to reduce setup or hold time,
SSO is different in that it only causes data delay and it cannot cause
the signal to occur earlier." Aren't they wrong?