backtop


Print 20 comment(s) - last by beemercer.. on Jul 18 at 9:09 PM


A diagram of the Itanium 2 "Montecito" die released at Fall IDF 2005
Intel's newest Itanium has some dramatic changes, but enough to stay competitive?

For my twenty-third birthday tomorrow, July 18, 2006, Intel is set to announce its newest addition to the Itanium 2 in three years.  Montecito marks a lot of first for the Itanium line.  For example, Montecito is:
  • Intel's first Itanium on the 90nm process
  • Intel's first dual-core Itanium
  • Intel's first Itanium with Virtualization Technology (VT) and Cache Safe Technology (CST)
  • Intel's first Itanium to feature processor numbers
The transition for Itanium 2 from 130nm to 90nm alone is a major step for the company.  The 130nm process size was already long in the tooth in 2003 when Itanium 2 Fanwood and Madison made an appearance.  Although the rest of Intel's processor manufacturing has for the most part switched to 65nm, Itanium 2 is just starting 90nm with Montecito.

Of course, a die shrink was inevitable to cram those 1.7 billion transistors on Montecito.  The high end model, Itanium 2 9050, utilizes a 24MB L3 cache which is split between the two processor cores on-chip

According to Intel documentation, existing Madison systems can use Montecito as a drop in replacement.  The 400 MHz FSB E8870 chipset, announced in August of 2000, still powers the Itanium line.  Interestingly enough, while some Madison processors were rated to 667 MHz FSB, the new Montecito chips top out at 533 MHz FSB, with a clock frequency that tops out around 1.6 GHz.  Since Intel has moved the entire product line to a number-based naming system, below is a table of the newest Intel chips.

Intel Itanium 2 "Montecito" Processors
Processor
Number
Core
Frequency
Bus
Frequency
L3
Cache
TDP
Price
9050 2 x 1.6 GHz 533 MHz 24MB 104W
$3,692
9040 2 x 1.6 GHz 533 MHz 18MB 104W
$1,980
9030 2 x 1.6 GHz 533 MHz 8MB 104W
$1,552
9020 2 x 1.42 GHz 533 MHz 12MB 104W
$910
9015 2 x 1.4 GHz 400 MHz 12MB 104W
$749
9010 1.6 GHz 533 MHz 6MB 75W
$696

The Itanium 2 9010 will be Intel's only single-core Montecito at launch.  Unlike Intel's Tulsa (scheduled to launch August 27, 2006) Xeon processor, Montecito will not feature shared L3 cache.  This has been a planned feature for Montecito almost as long as there has been a codename for it. 

Like previous Itaniums, Montecito features Hyper-Threading on most of its lineup.  Thus, dual-core processors will actually appear as logical quad-core processors.  The only chips not getting Hyper-Threading with this iteration are the dual-core Montecito 9030 and the single-core Montecito 9010.

Intel made the claim a year ago that Montecito is more than 60% faster than the Madison core, claiming a four-socket (8-core) system was able to achieve a LINPACK score of 45 GFLOPs.  To put that in perspective, a four-socket HP ProLiant BL45p loaded with Opteron 880s (4 processors, 2-cores per processor) has a LINPACK score of around 16 GFLOPs (PDF). Considering the $10B USD cash influx the Itanium line just received, we would expect Montecito to live up to the hype.

Montecito is expected to be replaced by Montvale and Tukwila CPUs, but Intel has not released a for these chips yet.


Comments     Threshold


This article is over a month old, voting and posting comments is disabled

It ever loses in SpecFP
By defter on 7/17/2006 2:38:34 PM , Rating: 2
In the past Itanium has dominated SpecFP because of Intel's good compiler, but now the times have changed:

SpecFP:
3GHz Woodcrest: 3056
http://www.spec.org/osg/cpu2000/results/res2006q3/...
1.6GHz Montecito: 3017
http://www.spec.org/cpu2000/results/res2006q2/cpu2...

In SpecINT performance is just horrible:
3GHz Woodcrest: 3065
http://www.spec.org/osg/cpu2000/results/res2006q3/...
1.6GHz Montecito: 1474
http://www.spec.org/cpu2000/results/res2006q2/cpu2...
For the comparison, ancient 2GHz Opteron scores 1447
http://www.spec.org/osg/cpu2000/results/res2005q3/...

If Montecito performance is so low in well optimized benchmark, I wonder how low it will be in real life application that aren't nearly that well optimized?




RE: It ever loses in SpecFP
By segagenesis on 7/17/2006 4:26:18 PM , Rating: 2
Warning: I'm speculating here because I dont have access to Itanium hardware as its just too darned expensive and unnecessary where I work.

Could one say that Intel actually competing with themselves now with Itanium 2 vs. Woodcrest? Granted the Itanium is designed and targeted for "big iron" type machines you see in supercomputer type applications. But really with the advent of clustered hardware I question if they are barking up the wrong tree. If you can build a supercomputer with a cluster of Opterons then why not do the same with Woodcrest now?


RE: It ever loses in SpecFP
By masher2 (blog) on 7/17/2006 4:33:43 PM , Rating: 2
The supercomputer cluster market is a little different (both in applications and intended customers) than your traditional big-iron servers. Its relatively easy to write a physics simulation that runs on a loosely-knit cluster of autonomous processors. Doing the same for, say, a SQL database, is considerably more difficult.


RE: It ever loses in SpecFP
By hstewarth on 7/18/2006 12:01:23 AM , Rating: 2
One thing I found looking at current Spec results

http://www.spec.org/osg/cpu2000/results/res2006q3/

Check out the 5160's scores in a lot of places the 4 core 2 cpu system has the same score as 2 core 1 cpu 5160's. Something is wrong here.

Also I notice that most of Itanium2 systems have large amout of cpus - like 128 or 256 cpus.

The woodcrest is incrediable cpu, I can wait to X7DA3 motherboard is available so I can complete mind.


RE: It ever loses in SpecFP
By defter on 7/18/2006 3:57:28 AM , Rating: 2
quote:
Check out the 5160's scores in a lot of places the 4 core 2 cpu system has the same score as 2 core 1 cpu 5160's. Something is wrong here.


It's normal. Normal Spec is a single threaded benchmark. If you want to see multithreaded results, you should look at Spec rate scores.


RE: It ever loses in SpecFP
By IntelUser2000 on 7/18/2006 3:19:18 PM , Rating: 2
quote:
In the past Itanium has dominated SpecFP because of Intel's good compiler, but now the times have changed:

SpecFP:
3GHz Woodcrest: 3056
http://www.spec.org/osg/cpu2000/results/res2006q3/...
1.6GHz Montecito: 3017
http://www.spec.org/cpu2000/results/res2006q2/cpu2...

In SpecINT performance is just horrible:
3GHz Woodcrest: 3065
http://www.spec.org/osg/cpu2000/results/res2006q3/...
1.6GHz Montecito: 1474
http://www.spec.org/cpu2000/results/res2006q2/cpu2...
For the comparison, ancient 2GHz Opteron scores 1447
http://www.spec.org/osg/cpu2000/results/res2005q3/...

If Montecito performance is so low in well optimized benchmark, I wonder how low it will be in real life application that aren't nearly that well optimized?


Montecito with HP chipset and HP compilers will get over 3700 SpecFP and 1700 SpecInt.


BTW, Hyperthreading in Intel's Montecito is SoEMT, while Hyperthreading in Intel's Pentium 4's are SMT.


Horrible pun, and a FSB question
By asliarun on 7/17/2006 2:27:14 PM , Rating: 2
Augh, "The Count of Montecito" is truly a stinker pun :-)

On a serious note, i have a very basic question: People keep saying that the reason why Xeon and now, Woodcrest, does not well beyond 2-way is because the FSB gets saturated. Which is why HT is such a good design and scales better.

Why is it that the Itanium architecture can scale up to 512 way configurations, which is way beyond even HT's ability, even though it uses FSB just the Xeons??




RE: Horrible pun, and a FSB question
By agnot on 7/17/2006 9:41:56 PM , Rating: 2
That's the thing, it doesn't just use FSB like Xeons. I assume you are referring to SGI Altix? Altix scaling to 512 processors has nothing to do with Itanium architecture and everything to do with system design.

Only two processors share a single FSB to connect to the SHUB (think of it as the northbridge). The SHUBs are then interconnected with a cache-coherent network to give the appearance of a single memory image. A processor can access memory on a remote node as if it was local memory, the only difference is that it takes longer (the more hops away the remote memory is, the longer it takes).

Altix is not SMP, it is NUMA. It is a distributed-shared memory architecture, where memory is physically distributed across many machines, but logically shared as a single memory pool.


RE: Horrible pun, and a FSB question
By Sunbird on 7/18/2006 3:21:36 AM , Rating: 2
They should have rather used something like "Gambling with the Montecito" or similar. The Montecito being the casino in the TV show Las Vegas.


RE: Horrible pun, and a FSB question
By Knish on 7/18/2006 6:50:20 AM , Rating: 2
Alternative titles:

Let's Smoke Up a Montecito (Montecristo cigars are world renowned)
Montecito: Aged Twelve Years (Montecristo also takes 12 years to produce a single bottle)
Montecito: Who cares? (Montecello is the birthplace of Thomas Jeffson. I am sure you don't care about that either)
Montecito: That war didn't matter! (Montenegro fought a civil war to become an indepdent nation from Yugoslavia, and then just got absorbed back up into Serbia anyway)

All of my ideas after that are progressively worse. I suppose Montecristo was the same thing that came to my mind too.


*sigh*
By jones377 on 7/17/2006 3:18:52 PM , Rating: 2
Montecity STILL does not have a shared L3 cache...




RE: *sigh*
Alpha
By Neophyte1980 on 7/17/2006 7:18:47 PM , Rating: 2
I just wonder how things would have gone if HP/Compaq hadn't murdered the Alpha series. I don't think this line of computing has recovered since then, allowing x86 to catch up.




RE: Alpha
By jkresh on 7/18/2006 12:15:05 AM , Rating: 1
Remember amd bought a lot of DEC's technology and the athlon64 has a fair amount of alpha in it (and hyptertransport was almost directly taken.)


Die Size
By beemercer on 7/18/2006 9:09:38 PM , Rating: 3
does anyone know the size of the die for the 9050?




DIE
By Alphafox78 on 7/17/06, Rating: 0
RE: DIE
By irev210 on 7/17/2006 10:50:00 PM , Rating: 2
quote:
I say Intel should reduce the cost of the core 2 duo by $10 Billion and kill the Itanium... how many of these things do they expect to sell?



Anything that you build, you learn on.

the 10B in research im sure will make its way into Conroe, etc.

Its not like they spent that $ on manufacturing.


Happy Birthday
By Visual on 7/18/2006 1:39:52 AM , Rating: 2
I wish you health, wealth and all the usual stuff.
And hey, I didn't know you're just a kid ;)
A year and some younger than me anyway, and I don't consider myself particularly grown up yet :)




By ElFenix on 7/18/2006 4:39:10 PM , Rating: 2
damn compaq for buying DEC, damn intel for burying it because intel suffers from 'not invented here' syndrome.




at least the prices are reasonable
By Samus on 7/17/2006 3:30:08 PM , Rating: 1
two of the dual-core itty's are under $1000, thats a dramatic change from when it was introduced (and was complete crap)

too bad the motherboards for $500




"It looks like the iPhone 4 might be their Vista, and I'm okay with that." -- Microsoft COO Kevin Turner

Related Articles
16MB of L3 Cache: Intel's "Tulsa"
May 28, 2006, 7:12 PM













botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki