Intel makes Power top target of Itanium

[YKhan]

Intel has usually gone with the attitude that the competition doesn't
exist, and never names the competition specifically. Looks like it's
trying to get some buzz going for Itanium now.

"It's a two-horse race. It's Power and Itanium," Gelsinger
said. "It's not particularly surprising they're not pursuing Itanium
when their motivation is Power."

And Talwalkar said of Itanium that "the primary focus and competition
is primarily Power."

Intel names top Itanium target: IBM | CNET News.com
http://news.com.com/Intel+names+top+Ita ... tml?tag=nl

Yousuf Khan

[Nick Maclaren]

In article <1109783833.710297.274530@z14g2000cwz.googlegroups.com>,
YKhan <yjkhan@gmail.com> wrote:

Intel has usually gone with the attitude that the competition doesn't
exist, and never names the competition specifically. Looks like it's
trying to get some buzz going for Itanium now.

"It's a two-horse race. It's Power and Itanium," Gelsinger
said. "It's not particularly surprising they're not pursuing Itanium
when their motivation is Power."

And Talwalkar said of Itanium that "the primary focus and competition
is primarily Power."

I didn't realise that IBM were in THAT much trouble! It is traditional
in such comparisons to name the competitor that you are least afraid
of - which implies some interesting things about what Intel thinks of
SPARC.


Regards,
Nick Maclaren.

[Greg Lindahl]

In article <1109783833.710297.274530@z14g2000cwz.googlegroups.com>,
YKhan <yjkhan@gmail.com> wrote:

Intel has usually gone with the attitude that the competition doesn't
exist, and never names the competition specifically. Looks like it's
trying to get some buzz going for Itanium now.

Itanium marketing has explicitly mentioned clobbering RISC vendors for
a long time. With all the dropouts, only POWER is left. So, nothing
really new.

-- greg

[Guest]

YKhan wrote:

Intel has usually gone with the attitude that the competition doesn't
exist, and never names the competition specifically. Looks like it's
trying to get some buzz going for Itanium now.

"It's a two-horse race. It's Power and Itanium," Gelsinger
said. "It's not particularly surprising they're not pursuing Itanium
when their motivation is Power."

And Talwalkar said of Itanium that "the primary focus and competition
is primarily Power."

Intel names top Itanium target: IBM | CNET News.com

http://news.com.com/Intel+names+top+Ita ... tml?tag=nl

Yousuf Khan

Whats new?
The two-horse race has been said several times before.

[Yousuf Khan]

icer...@spray.se wrote:

Whats new?
The two-horse race has been said several times before.

Never explicitly.

Yousuf Khan

[Nick Maclaren]

In article <42260a19$1@news.meer.net>, Greg Lindahl <lindahl@pbm.com> wrote:

In article <1109783833.710297.274530@z14g2000cwz.googlegroups.com>,
YKhan <yjkhan@gmail.com> wrote:

Intel has usually gone with the attitude that the competition doesn't
exist, and never names the competition specifically. Looks like it's
trying to get some buzz going for Itanium now.

Itanium marketing has explicitly mentioned clobbering RISC vendors for
a long time. With all the dropouts, only POWER is left. So, nothing
really new.

Well, there are an awful lot of SPARC systems in use, Sun is still
selling them, and they perform an awful lot better than their
detractors claim. Whether you like it or not, they are still
hanging in there.

Of course, it DOES make a difference that, for the sort of HPC we
are into, the serial CPU performance is less important than both
the memory bandwidth and the interconnect. This makes a SunFire
F15K deliver a lot more aggregate oomph than its raw MFlops would
indicate.

We shall have to see whether Niagara and Rock give SPARC enough of
a boost to keep it going to 2010.


Regards,
Nick Maclaren.

[Nick Maclaren]

In article <1109793003.791321.139150@o13g2000cwo.googlegroups.com>,
Yousuf Khan <yjkhan@gmail.com> wrote:

icer...@spray.se wrote:
Whats new?
The two-horse race has been said several times before.

Never explicitly.

Boggle. Where have you been? Arcturus? I got sick of the nonsense
of a two-horse race a decade ago - of course, it hasn't always been
the SAME two horses ....


Regards,
Nick Maclaren.

[Christian Bau]

In article <d0570a$dqi$1@gemini.csx.cam.ac.uk>,
nmm1@cus.cam.ac.uk (Nick Maclaren) wrote:

In article <1109793003.791321.139150@o13g2000cwo.googlegroups.com>,
Yousuf Khan <yjkhan@gmail.com> wrote:
icer...@spray.se wrote:
Whats new?
The two-horse race has been said several times before.

Never explicitly.

Boggle. Where have you been? Arcturus? I got sick of the nonsense
of a two-horse race a decade ago - of course, it hasn't always been
the SAME two horses ....

And I didn't realize Itanium was still one of them...

[Guest]

Nick Maclaren wrote:

In article <42260a19$1@news.meer.net>, Greg Lindahl <lindahl@pbm.com
wrote:
In article <1109783833.710297.274530@z14g2000cwz.googlegroups.com>,
YKhan <yjkhan@gmail.com> wrote:

Intel has usually gone with the attitude that the competition
doesn't
exist, and never names the competition specifically. Looks like
it's
trying to get some buzz going for Itanium now.

Itanium marketing has explicitly mentioned clobbering RISC vendors
for
a long time. With all the dropouts, only POWER is left. So, nothing
really new.

Well, there are an awful lot of SPARC systems in use, Sun is still
selling them, and they perform an awful lot better than their
detractors claim. Whether you like it or not, they are still
hanging in there.

Of course, it DOES make a difference that, for the sort of HPC we
are into, the serial CPU performance is less important than both
the memory bandwidth and the interconnect. This makes a SunFire
F15K deliver a lot more aggregate oomph than its raw MFlops would
indicate.

Can you share the numbers, please? For example, would 72-socket
(144-core) US-IV based SunFire15K outperform 16-socket (32 cores) p570?
32-socket p595?
How it faires against mid-range (64P) Altix3K?
All for your sort of HPC, of course.

We shall have to see whether Niagara and Rock give SPARC enough of
a boost to keep it going to 2010.

Niagara is strictly low-end - the segment Intel/Itanium already gave
up. Rock is too far off to make an educated guess.

Regards,
Nick Maclaren.

[Guest]

Oh, sorry, mixed SunFire15K with SunFire25K ):
Aren't US-III Cu-based SunFire15K computers too expensive for HPC?

[Nick Maclaren]

In article <1110109738.580076.95890@o13g2000cwo.googlegroups.com>,
<already5chosen@yahoo.com> wrote:

Can you share the numbers, please? For example, would 72-socket
(144-core) US-IV based SunFire15K outperform 16-socket (32 cores) p570?
32-socket p595?
How it faires against mid-range (64P) Altix3K?
All for your sort of HPC, of course.

I am afraid not. I have not tried those, and the information I have
on some machines is NDA. What I can tell you that ISN'T under NDA,
is that a Sun F15K beats the balls off an IBM P670 (Regatta). This
is because IBM made a complete pig's ear of the memory management on
the POWER4, which I believe they have fixed in the POWER5.

No, I don't know why and how, and IBM aren't talking, but the difference
between the claims and the delivery jumps up and hits you in the teeth
as soon as you start to run actual code. No SMP system is immune
from scalability problems, but I am extremely impressed by the work
of SGI and Sun. Note that I haven't used either a POWER5 system or
a HP Itanium SuperDome, and so am not commenting on them (either way).

The Altix is very clearly the technical leader in large-scale SMP,
but I can't give you figures. What impressed me so much with Alan
Charlesworth's memory management for the F15K was how good it is,
given the large jump from any previous practical experience.

We shall have to see whether Niagara and Rock give SPARC enough of
a boost to keep it going to 2010.

Niagara is strictly low-end - the segment Intel/Itanium already gave
up. Rock is too far off to make an educated guess.

It's not QUITE like that, but I take your point.


Regards,
Nick Maclaren.

[Nick Maclaren]

In article <d0evq2$r45$1@gemini.csx.cam.ac.uk>,
Nick Maclaren <nmm1@cus.cam.ac.uk> wrote:

In article <1110109738.580076.95890@o13g2000cwo.googlegroups.com>,
already5chosen@yahoo.com> wrote:

Can you share the numbers, please? For example, would 72-socket
(144-core) US-IV based SunFire15K outperform 16-socket (32 cores) p570?
32-socket p595?
How it faires against mid-range (64P) Altix3K?
All for your sort of HPC, of course.

I am afraid not. I have not tried those, and the information I have
on some machines is NDA. What I can tell you that ISN'T under NDA,
is that a Sun F15K beats the balls off an IBM P670 (Regatta). This
is because IBM made a complete pig's ear of the memory management on
the POWER4, which I believe they have fixed in the POWER5.

Perhaps I should clarify this. Precisely the converse was true of
single-CPU performance, including memory access, where the POWER4
beat the balls off everything else. The press was solid with
statements that everything else (except perhaps the Itanium) was
doomed, because the idiots looked solely at the single-CPU Spec
figures.

However, when you start to run N copies of an almost identical
processes in parallel (as is normal for HPC), things degraded badly.
Systems like the Altix and F15K degrade when there is a conflict,
but scale well where there isn't. IBM never claimed that the 670
(actually, I mean 690) would scale linearly but, in the event,
things were a lot worse than expected (especially with respect to
latency, for some reason).

STREAMS figures published by John McCalpin showed this very clearly,
as they also show that the POWER5 has fixed at least the obvious
issues. I can't tell you what the problem was, but I do know that
it wasn't expected. I really can't explain the latency effect.

The effect was that a F15K couldn't even approach a 670/690, CPU
for CPU, but it scales up to c. 100 CPUs (with memory corresponding
to 72), and a 670/690 tops out at 32. There is actually more to this,
because tuning for AIX 5L is considerably harder than for Solaris 9,
especially in this area, and I was therefore referring to the sort
of performance the "ordinary" HPC program would see.

I cannot say what would happen on a single-application HPC system,
where AIX could be tuned for that application. This is a large-
page issue, incidentally - see the AIX documentation for the details.

Until and unless I compare a F25K and 595 myself, I can't say what
they will do (and then it is unlikely I will be able to post).


Regards,
Nick Maclaren.