[Beowulf] itanium vs. x86-64
diep at xs4all.nl
Tue Feb 10 19:11:54 PST 2009
On Feb 10, 2009, at 11:12 PM, NiftyOMPI Mitch wrote:
> On Tue, Feb 10, 2009 at 7:39 AM, Håkon Bugge <hbugge at platform.com>
>> There are several MPI2007 SPEC submission using Itanium and
>> x86-64. Here is
>> a comparison between two submissions using 128 MPI processes,
>> comparing an
>> SGI Altix 4700 using Itanium 2 Processors 9040 1.6GHz and the
>> fastest x86-64
> Almost apropos to this thread.
I find in such articles the claim that intel blew a lot of money at
My point is SUBSTANTIATE what amount of money intel has thrown at it.
What we do know is:
a) it is a very cheap core, so yields will be very good. Probably
80% or so.
So the profit on each cpu produced is huge.
b) it gets produced at older process technology. So it in no way
interrupts any x86 production line,
making the production of the cpu really cheap.
c) intel must have realize at an early stage that they can't clock
the cpu as high as other future cpu's can,
in short they realized really soon, better than anyone else,
that it would be impossible to improve its performance
factor 2.0 each 18 months.
d) INTELS BUSINESS CASE IS TO CLOCK CPU'S HIGHER THAN COMPETITORS
From 'a' to 'd' i deduce that at an early stage, intel must have
concluded itanium is a dead end for them.
The logical decision companies take then is to size down projects and
see whether you can still get money out of the project.
That is exactly what intel has been doing.
Now the question is: how much in the 90s did Intel put into the project?
My guess is initially a lot, but in the 21th century really little.
On the rule that industry uses (not Moore's rule though): "doubling
of hardware performance each 18 months".
Let's calculate further from alpha.
Alpha is having first 633Mhz cpu's in 1997, i played against 2 of
them and ran at 1 of them.
Impressive was the 21264c. Let's use that as it's not too long ago.
21264c for those who didn't pay attention is
eating the first itanium2 alive, other than that at spec of course
compilers and a bigger cache do the job to be on paper
faster. In reality alpha is a monster for scientists.
If we extrapolate its performance of 1.25Ghz back in 2001
If some itanium2 sees the daylight in 2010:
9 years * 12 = 108 months.
108 / 18 = 6 doublings of speed.
2^6 = 32
Itanium will clock 2x higher in 2010 i suppose. 1.25 ==> 2.4Ghz (say
a slight improvement in ipc compensates):
So in 2010 the itanium to keep up with industry HPC rule needs to
have an effective speed of a 16 core 2.5ghz 21264c
SMT gives too little to compensate for a factor 2, you get that for
free simply if your application can use it,
besides that with so many cores you ain't gonna get 2.5Ghz. Not even
close. Maybe 2Ghz in 45 nm.
In short that is a 16 core Itanium 2.xGhz in 2010 we need to get for
Intel to keep up with market demand.
Well, i assume it is unlikely itanium2 suddenly is 16 cores in 2010.
Itanium core needs too much L3 cache for testsets, and as intel loves
benchmarks, i consider this unlikely.
> T o m M i t c h e l l
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit
More information about the Beowulf