Linus Torvalds (torvalds@linux-foundation.org) on 2/7/09 wrote:
---------------------------
[..]
>>1) IPF QPI and Xeon QPI don't use the same coherency
>>protocol (Xeon uses a simplified snoopy scheme suitable
>>for 2s/4s, IPF uses a more complex scalable distributed
>>directory system). Where do you think most of the "high
>>level" complexity is.
>
>If true (and I have no real reason to doubt it), that's
>just crazy, but I guess it meshes well with the whole
>insane notion of Intel just traditionally not letting x86
>grow up. I had really thought they got over that.

The scalability stuff in Tukwila takes up a significant
portion of the die area and power budget. Putting even
a fraction of that overhead in the same die that Intel
sells as desktop and laptop processors is just plain
nuts.

What about Beckton (Nehalem generation Xeon MP)?
Well the fact is x86 server sales of 4 sockets and up is
only worth about $1B per quarter, has flat sales growth
and AMD takes a bigger share of it than any other x86
segment. Putting the scalability stuff in Beckton would
probably cost it a couple of cores and maybe a speed
bin. Beckton will compete with future AMD processors
far more than IPF and extra cores and clock speed are
more valuable than scaling well to beyond 4 sockets
(which AMD server processors don't).

The truth is IPF is now more valuable to Intel than Xeon
MP. In the past five years IPF has severely curtailed x86's
presence in the mid and high end server segments. IPF
sales already exceed the total value of x86 server sales
of 4 sockets and up by about a third and is growing much
faster.

That's why IPF QPI gets distributed directory and x86 QPI
doesn't. Sorry, no craziness or conspiracy theory here.
It's just Intel doing what the server market wants. :-)