Advanced Micro Devices Inc (AMD)

[chipguy]

Another core?

Probably not. My guess is they went for the holy grail of
OOO execution processors but couldn't get it working in
time. The holy grail being dynamic multi-threading or DMT.

Imagine a hyperthreaded processor but instead of having to
run two separate threads explicitly coded by programmers
the CPU runs along one thread until it hits a conditional
branch that it knows from its history tables is hard to
predict. So instead of making a likely poor guess it starts
a second thread - one thread goes down one path and the
other thread goes down the other. When the branch condition
is resolved the incorrect path thread is collapsed and its
processor state is reloaded with the correct thread's state
so the whole process can start again.

This is fiendishly complicated because neither thread can
update memory or do anything else permanent until the right
path is known. That means a whole bunch of operations have
to be accumulated in two separate groups and then either
discarded or quickly emptied out to memory. It is easy to
imagine DMT soaking up tens of millions of transistors.

If done right it could be a big speed up (it isn't uncommon
for OOO processors to execute but then discard the results
of up to 3 or 4 instructions for every 10 that are retained in
integer programs because of execution down wrong paths).
But one subtle little logic error or one conceptual corner
case unconsidered and your processor goes off to lala land
and it would be very difficult figuring out why.

HT wasn't ready in time for Willamette's release and in
terms of complexity DMT is like HT squared. If Intel has
attempted DMT in Prescott then it's hardly surprising if
it is still very much be a work in progress.