backtop

Print E-mail del.icio.us 24 comment(s) - last by Amiga500.. on Jun 19 at 4:15 AM
Intel says Itanium will continue to sail for the long run

Intel this week revealed new details about where it plans to take its server processor business, specifically with the Itanium processor. As many know, Intel took a gamble on the Itanium when it was released several years ago. Intel's use of a then uncommon Explicitly Parallel Instruction Computing (EPIC) architecture left the majority of the industry unsure of Itanium's practicality in an x86-dominated world. Today however, the Itanium family brings in roughly $3.5 billion per annun for Intel.

Diane Bryant, vice president of Intel's enterprise group, revealed several details that indicate Intel will push forward with Itanium development for the foreseeable future.

Currently, Intel's flagship Itanium 2 processor is the Montecito core. Intel announced Montecito last July, marking the company's first dual-core enterprise and mainframe processor. Until now, Montecito ran on a 533MHz front-side bus but will soon make the transition a 667MHz front-side bus processor in Q4 2007, dubbed Montvale. According to Bryant, Montvale will consist of minor updates, improving bus speed but also improving overall stability.

Recent Intel roadmaps indicate that Montvale will consist of roughly 25 percent of Intel's Itanium business in Q4 2007. By Q1 2008, Intel guidance suggests Montvale will take up a whopping 40 percent of all Itanium 2 sales. Despite Intel's desktop processors currently seeing day light at 65nm, Itanium 2 processors will still be on 90nm manufacturing technology. Montvale will also top out at a core speed of 1.66GHz with a total of 24MB of L3 cache.

Intel's next major milestone in the Itanium family will come with the arrival of Tukwila, a quad-core processor due sometime in late 2008. According to Bryant, Tukwila will be roughly twice as fast as Montecito and feature an on-die memory controller.

This will be a turning point for Intel because with Tukwila's need for a discrete memory controller gone, the company will introduce its long waited common system interface (CSI). Tukwila's use of CSI will be a direct response to AMD's HyperTransport technology. Intel previously stated it will not restrict CSI to the Itanium family, but will eventually use the technology as the main transport bus for the Xeon family as well.

Tukwila will also come with even larger caches and a new reliability feature called double device data correction (DDDC). DDDC acts as a failsafe mechanism to protect system memory failures from bringing down a live system. In a hardware failure where a memory chip on a memory module fails, DDDC will be able to mark that chip as unusable without compromising system stability. DDDC differs from traditional ECC and parity technology due to its capability to withstand more than one chip failure.

Bryant went on to reveal details about the future release of an entirely new Itanium architecture code-named Poulson. With Poulson, Bryant claims Intel will introduce even more cores; greater scalability and the introduction of 32nm die fabrication for its enterprise segment. The company will skip 45nm technology altogether for Itanium. Bryant did not give details on a possible release date for Poulson although Intel is expecting to introduce 32nm processors in approximately two years.

Comments     Threshold


This article is over a month old, voting and posting comments is disabled
RISC?
By borowki on 6/18/2007 9:04:04 AM , Rating: 2
Is it correct to classify Itanium's VLIW architecture as RISC?




RE: RISC?
By RamarC on 6/18/07, Rating: -1
RE: RISC?
By masher2 (blog) on 6/18/2007 10:08:27 AM , Rating: 4
Risky, only for those people who mistakenly believed Itanium was intended to supplant Intel's desktop and small-server cpus. Itanium was a glove thrown down to challenge big iron...and in that marketplace, its actually done quite well.

As to the terms themselves, some people have called VLIW an "extension" of superscalar RISC, but in my opinion, there's very little real overlap.


RE: RISC?
By Laitainion on 6/18/2007 10:36:03 AM , Rating: 2
Origionaly, it was eventually.

From wikipedia (http://en.wikipedia.org/wiki/Itanium):

"During development, Intel, HP, and industry analysts were predicting that IA-64 would dominate in servers, workstations, and high-end desktops, and eventually supplant RISC and complex instruction set computer (CISC) architectures for all general-purpose applications."


RE: RISC?
By masher2 (blog) on 6/18/2007 11:03:11 AM , Rating: 2
> "During development..."

Those statements were back before Itanium was even named as such, much less released. The target market changed as development proceeded. From the point the design was finalized, Itanium was intended only for the big-iron marketplace.

Long-term though, I wouldn't be surprised if some of those predictions don't ultimately come true. IA-64 benefits from a large cache much more than x64 does. Past the 32nm node, I wouldn't be surprised to find Itanium in low-end servers and some workstations.


RE: RISC?
By shady28 on 6/18/2007 1:03:33 PM , Rating: 5
Intel's major Itanium vendor is HP. Most of HP's midrange and high end servers will ultimately be itanium only - HP has stopped developing the PA-RISC architecture that has carried them through 20+ years of servers. Seeing HP has about 24% of the server market, that's a good thing for Intel.

Having said that, on a technical basis the Itanium isn't all that great. IBM's Power5 / Power6 is a superior performer, now hitting 4Ghz on a quad core chip and both Power5 and Power6 hold top slots in Spec benchmarking.


RE: RISC?
By eilersr on 6/18/2007 1:04:41 PM , Rating: 6
No, Itanium's VLIW is a different category than RISC.
VLIW is the name for a the general approach. Intel's approach in Itanium is EPIC (Explicitly Parallel Instruction Computing), which takes VLIW in a slightly different direction.

Don't take these as absolute definitions, but this is how I see it:

RISC: Simple instructions that are easy to process. Most RISC architectures have a relatively small set of well defined instructions (hence, Reduced Inst. Set). Typically every instruction has the same fixed length (# of bits). Greater # of instructions required to accomplish the same operation than CISC, but are typically executed faster than CISC to make up for it.

CISC: Complex instructions that may take longer to process, but accomplish more work per instruction than RISC. Each instruction can have variable bit length, depending on what needs to be done. CISC architectures tend to have many broadly defined instructions for very specific tasks (e.g., all the various SSE instructions). Most modern CISC archictures (incl. x86) break CISC instructions down to something closer to internal RISC-like operations (e.g. micro-ops) to take advantage of architectural techniques that are easier to employ in a RISC architecture (pipelining, super-scalar, etc.)

VLIW: Can think of it as a number of RISC-like instructions that are strung together to achieve a larger operation. The final instructions are much longer than what is seen in CISC architectures. The compiler must do a lot of work to group the sub-instructions together efficiently.


This is all a BIG SCAM
By LogicallyGenius on 6/18/07, Rating: -1
RE: This is all a BIG SCAM
By mmonnin03 on 6/18/2007 8:37:43 AM , Rating: 2
That will be completely unfeasible.
1) The motherboard will be monstrous, absoluelty huge. Too big by any means.
2) It would need MANY more layers for interconnects.
3) The chip(s) connecting all these CPUs would need hundred of pins connecting to the motherboard or even multiple 'northbridges' thus adding another chip for data to pass through
4) No way you could power all this on 1 motherboard, the current draw would too much for 1 PSU, or else you would have to design something new for multiple PSUs for 1 motherboard to all come up safely at once.
If they want 10-20 CPUs, go to Cray and I bet even they have multiple boards with 2 or so CPUs connected together through fiber. 20 CPUs on 1 board is just retarded.


RE: This is all a BIG SCAM
By encryptkeeper on 6/18/2007 8:42:14 AM , Rating: 2
10 or 20 processors? So I guess you just want motherboard manufacturers head's to explode huh? Jesus Christ, imagine trying to design the chipset for that. I think, in his heart, the original poster may have been intending for the core count to reach this level. Don't worry it'll happen. Let software designers have a chance to redesign code for quad processors and beyond before we start with 20 cores.


RE: This is all a BIG SCAM
By masher2 (blog) on 6/18/2007 12:22:44 PM , Rating: 3
> "10 or 20 processors? Jesus Christ, imagine trying to design the chipset for that.."

There have been 8 cpu mobos out for years now. Plug in quad core processors, and thats 32 effective cores.


RE: This is all a BIG SCAM
By Puddleglum1 on 6/18/2007 3:59:11 PM , Rating: 3
Reading his post, it says
quote:
I think, in his heart, the original poster may have been intending for the core count to reach this level.
So his main point was not how to get that many cores, but whether it was plausible to go beyond 4 cores at present due to software limitations.

To encryptkeeper: As far as I know, only servers are using more than 4 cores right now, and it's common for up-to-date server applications to add new threads for each new process. 20+ cores is feasible right now, which is why Intel(http://anandtech.com/cpuchipsets/intel/showdoc.asp... Sun(http://www.sun.com/processors/UltraSPARC-T1/index.... IBM, ClearSpeed (http://news.zdnet.com/2100-9584_22-5399128.html?ta... and a few other companies are heavily invested into it.

http://en.wikipedia.org/wiki/Manycore_processing_u...

I agree with you that software needs to catch up before we start seeing more than four cores in a system. But there is already a need for more cores, and 10-20 cores on a motherboard is ridiculous in comparison to the benefits of 8-16 cores on single processor.


RE: This is all a BIG SCAM
By encryptkeeper on 6/18/2007 4:24:52 PM , Rating: 3
Alright, at least SOMEONE understood the post.

"But there is already a need for more cores, and 10-20 cores on a motherboard is ridiculous in comparison to the benefits of 8-16 cores on single processor."

I'm going to guess you meant 10-20 processors on a motherboard is ridiculous...


RE: This is all a BIG SCAM
By masher2 (blog) on 6/18/2007 5:11:31 PM , Rating: 2
> "So his main point was not how to get that many cores, but whether it was plausible to go beyond 4 cores at present "

I think one of his subpoints was the difficulty of designing a motherboard around that many processors...hence his quote that I highlit. But thats a side point.

32-way Intel-based servers have been sold for half a decade now. But its not only servers that are using more than four cores. There's a considerable amount of workstation based software (scientific, simulation/modelling, financial-analysis etc) that will easily use 8+ cores.


RE: This is all a BIG SCAM
By FujiT on 6/18/2007 8:51:02 AM , Rating: 2
1. Aren't they already huge?

3. With the quad core due out in 2008, it has CSI. The northbridge is integrated into the CPU. And it's all point to point which means that they could use the extra CSI links to connect the processors. And we're already at 1300+ pins for next year's Nehalem CPUs. Plus it's a serial bus which is supposed to reduce the pin count. Case in point? FB-DIMM

4. They power quad xeons fine. And what they're planning on will be built on the 32nm process, almost a third of the size of the current process.


RE: This is all a BIG SCAM
By colonelclaw on 6/18/2007 9:49:12 AM , Rating: 2
ok this isn't quite what you guys are talking about, but it comes from the same tech genre:
http://www.renderdrive.com/page/109/raybox.htm
the architecture is different but its still a motherboard with 14 cpu sockets on it and it's fairly small and relatively cheap. well, cheap in industry terms :)


RE: This is all a BIG SCAM
By Duraz0rz on 6/18/2007 8:40:15 AM , Rating: 2
Imagine the logistical nightmare of creating such a board with their current tech.


RE: This is all a BIG SCAM
By noirsoft on 6/18/2007 8:47:07 AM , Rating: 1
I'd rather see a single processor with 10-20 cores than the honking huge board with 10-20 processor slots.

Besides, what would make Intel more profit?

1) Buying 10 Intel processors, 10 non-intel motherboards, and making a 10-computer cluster
2) Buying a 10-slot Intel motherboard, 10 intel processors, and making a single supercomputer.

Assuming the only Intel parts are the proc and mobo, obviously, option 2 would be more profit. So, your argument is bogus from a financial/profit perspective as well.


RE: This is all a BIG SCAM
By LogicallyGenius on 6/18/2007 12:16:33 PM , Rating: 2
I wonder how much power we will get if we can put 20 processors that today costs 50$ each ?


RE: This is all a BIG SCAM
By Amiga500 on 6/18/2007 1:06:34 PM , Rating: 3
No, you are mixing up a desktop workstation with HPC "workstations".

A desktop needs several different kinds of processors (GPU/CPU/physics/High Math etc), while a HPC machine is just a number cruncher, no graphics or multimedia necessary.

Parallel coding isn't an issue for HPC as the codes are alreadly highly optimised for such runs.

At the moment, we've a 74 CPU and 160 CPU (Itanium II) clusters here. I'd happily double that if I could get the licenses to use it.

(Although, for all its concentration on number crunching, some maybe interested to know that per CPU, an Itanium II @ 1.6 GHz is about 15% slower than an Opteron @ 2.4 GHz on the same job)


RE: This is all a BIG SCAM
By offkey on 6/18/2007 4:09:06 PM , Rating: 1
quote:

(Although, for all its concentration on number crunching, some maybe interested to know that per CPU, an Itanium II @ 1.6 GHz is about 15% slower than an Opteron @ 2.4 GHz on the same job)


Unless you are crunching int numbers, opteron is no way to surpass itanium. The latest spec benchmark shows:

n4250QE (S4985) AMD Opteron 2222SE: 14.3 15.2
HP Integrity rx6600 (1.6GHz/24MB Dual-Core Intel Itanium 2): 17.3 18.1

Dual core to Dual core, 1.6G (Itanium) vs 3.0G (Opteron)


RE: This is all a BIG SCAM
By zpdixon on 6/19/2007 12:45:51 AM , Rating: 2
quote:

n4250QE (S4985) AMD Opteron 2222SE: 14.3 15.2
HP Integrity rx6600 (1.6GHz/24MB Dual-Core Intel Itanium 2): 17.3 18.1
Dual core to Dual core, 1.6G (Itanium) vs 3.0G (Opteron)


This benchmark shows Opteron as being much more attractive than Itanium. They show a $3692 Itanium CPU [1] (104W) as being only ~20% faster than a $873 Opteron CPU (119W), both in roughly the same power envelop.

IOW, the Itanium is 423% more expensive but only scores 20% better in this benchmark.
IOW, if you are a potential customer with XXX dollars to spend, you'll get about 4x more processing power by buying Opteron servers instead of Itanium servers.
IOW, if you are a potential customer in need of XXX amount of processing power, Opteron will take you to that level for 1/4th the cost of an Itanium solution.

No wonder why Itanium has been such a failure. (As correctly pointed out by a previous poster, Intel designed the IA-64 architecture to eventually replace i386. This has never happened.)

[1] http://www.intel.com/intel/finance/pricelist/proce...
[2] http://amd.com/pricing


RE: This is all a BIG SCAM
By mars777 on 6/19/2007 12:48:39 AM , Rating: 2
So you said that an Itanium is faster at FPU calculations than an Opteron... wow that was promiscuous.

Now try to count 128 Itaniums killing themselves for bandwidth in a cluster environment vs 128 Opterons running together on HT and optics. (numbers are socket not core)

All that in a more mixed parallel environment and not simply/exclusively FPU calculations.

Well most of the time they will be on par, but if the calculations become more "mixed type" Opterons will surely win.

Itaniums are like a group of Elephants, Opterons like a group of Rhinoceros, but we all know that Elephants dye faster when there is no food. Yes they have bigger stomachs (cache) but they have to share it but they lack a good way of doing it (CSI).

The next itanium could be much better when they bring on the CSI and memory controller.

THe point is - SPEC doesn't mean much if considering big clusters.


RE: This is all a BIG SCAM
By Amiga500 on 6/19/2007 4:15:33 AM , Rating: 2
Unless you are crunching int numbers, opteron is no way to surpass itanium. The latest spec benchmark shows:

n4250QE (S4985) AMD Opteron 2222SE: 14.3 15.2
HP Integrity rx6600 (1.6GHz/24MB Dual-Core Intel Itanium 2): 17.3 18.1

Dual core to Dual core, 1.6G (Itanium) vs 3.0G (Opteron)

Its a 2004 vintage Itanium II (the older cluster is the Madison cores, dunno what the newer one - which I've compared is) versus a 2004 Opteron 250.

And as subsequently pointed out - in terms of bang-for-buck, the AMD is way out ahead.

Hopefully I can get a comparison with some Woodcrests sometime soon.


"So if you want to save the planet, feel free to drive your Hummer. Just avoid the drive thru line at McDonalds." -- Michael Asher

DailyTech Poll
Which web browser do you use on your primary personal machine? 






44 Comments









botimage
Copyright 2009 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki