Slashdot Mirror
Why Doesn't the Itanium Get the Respect It's Due?
happycorp wonders: "As in recent years the Itanium does well, easily beating x86 processors even at its low clockspeed (1.4Ghz). The supercomputer people are serious about benchmarking (no easily tricked microbenchmarks or reliance on closed-source
commercial apps), so the discrepancy between the performance and perception of this chip is serious.
With a single-CPU Itanium2 system at
around $2000 their price is already reasonable, and the price would come down
(and software would be ported) if the Itanium ever became a mass market chip. Having an affordable chip one step above a Xeon or Opteron in floating-point performance would not be such a bad thing for gaming enthusiasts (or 3D artists). So, the recent
article
on the
Top 500 supercomputers list brings up a question I've been meaning to ask:
Why do we see so many disparaging opinions of the Itanium processor (all those 'Itanic' jokes, etc.)?"
"It seems computing enthusiasts' sentiment is set against this processor, and its likely that it's going to be abandoned sooner or later. We'll be paying for x86 compatibility indefinitely (recall the Xeon has roughly
three times the number of transistors of the ppc970 for example; but we hardly get three times the performance).
These are a couple scores from the top 20, with the total gigaflops divided by the number of processors to obtain a per-processor speed:
rank processor ghz (gflops / #procs) speed #5 ppc970 2.2 (27910 / 4800) 5.81 #7 itanium2 1.4 (19940 / 4096) 4.86 #10 opteron 2.0 (15250 / 5000) 3.05 #20 xeon 3.06 (9819 / 2500) 3.92
Given this, consider what a 2 or 3 Ghz Itanium could do.
(fine print: I am not affiliated with the Itanium or the top500 list in any way)."
These are a couple scores from the top 20, with the total gigaflops divided by the number of processors to obtain a per-processor speed:
rank processor ghz (gflops / #procs) speed #5 ppc970 2.2 (27910 / 4800) 5.81 #7 itanium2 1.4 (19940 / 4096) 4.86 #10 opteron 2.0 (15250 / 5000) 3.05 #20 xeon 3.06 (9819 / 2500) 3.92
Given this, consider what a 2 or 3 Ghz Itanium could do.
(fine print: I am not affiliated with the Itanium or the top500 list in any way)."
The chipmaker has released two new Itaniums for two-processor servers as part of its effort to eliminate price premiums on the chip.
Intel announced on Monday two new Itanium processors for two-processor servers, another step in the company's efforts to eliminate price as a barrier to Itanium acceptance.
The 1.4GHz Itanium 2 with 3MB of cache is designed for servers in clusters. The new chip will provide about 25 percent more performance and cost much less than the initial Itanium optimised for clusters, which came out last year, said Jason Waxman, director of multiprocessor platform marketing at Intel.
The second new chip, a 1.6GHz Itanium 2 with 3MB of cache, is optimised for higher performance in general-use two-processor servers, he said.
Waxman reiterated that Intel is working on several technologies that will eliminate any price premium on Itanium by 2007 and thereby allow its performance advantages to, hopefully, blossom.
"The price/performance balance will be heavily in favour of Itanium," Waxman said.
With the focus on price, the Itanium melodrama is once again reaching a turning point. After several years of delays, the chip family debuted in 2001 to poor reviews and negligible customer acceptance. A second version of the chip that appeared in 2002 dramatically improved performance but failed to spark the market.
Itanium finally began to gain acceptance in 2003 with Madison, a new version of Itanium 2 that substantially improved performance again and lowered the cost. Intel shipped about 100,000 Itaniums in 2003, compared with only around a few thousand for the first two years. Itanium volume is expected to double this year, chief executive Craig Barrett said in February.
But in 2004, Intel announced that it would come out with a version of its Xeon chip that runs both 32- and 64-bit code. Xeon and Pentium chips typically run 32-bit code. Itanium runs 64-bit code, which, among other advantages, lets a computer maker pack far more memory into a computer.
Itanium, however, requires completely different software to work well, a factor that has hindered adoption. Part of the appeal of the Opteron chip is that it can handle larger memory loads in 64-bit mode on essentially the same software base.
Lowering the cost of Itanium servers won't eliminate the software issue, but it will begin to create an environment in which greater acceptance could occur, which in turn could cause software developers to gravitate to Itanium. Analysts and PC makers have viewed this theory with various doses of scepticism, but the range of opinion is generally substantially less negative than it was 18 months ago.
Price drops have already had some effect. In 2002, a two-processor Itanium server cost about $18,000 (£9,859). With the new chips, a similarly configured system can sell for less than $8,000, while basic one-processor Itanium servers will go for just more than $2,000.
Some of these price cuts have come as a result of Moore's Law, which predicts that the number of transistors on a chip will double every 18 months. But Intel has also expanded its product line to better suit the economic realities of two-processor servers. The company also designs and partly manufacturers many of the Itanium servers on the market, which cuts independent engineering costs.
To lower the price further, Intel will begin to create products and add features to Itanium so that Itanium servers can be made out of many of the same components as Xeon servers. In 2005 and 2006, Itanium servers will be able to use the same memory or other components of Xeon servers, Waxman said.
In 2005, Intel will also come out with two different chipsets for Montecito, the next major version of the chip. One chipset will wring maximum performance out of the chip, Waxman said, while the other will allow server makers to insert Montecito into their Madison-based servers, thereby cutting down independent design efforts.
By 2007, Intel will
a few reasons.
One, market penetration. Windows *kind of* works on Itaniums. Code has to be compiled specifically for the platform - they're not very good at x86 code through WoW.
The BIOS replacement they use is not functional. It's very difficult to set up disks for use, and if you lose the disk that the BIOS data is kept on, you're screwed. As far as I know, there is no way to make that fault-tolerant short of manually storing the contents of that partition on another drive.
Support for the Itaniums has been terrible. The HP systems are riddled with hardware problems, and their support personnel (at the enterprise level) have no idea how to comprehend that they don't operate quite like any other workstation.
While, the IA64 has always had great floating point performance, there's an awful lot of us out here that don't need fast FPUs -- e.g. code development, database, web serving, network i/o etc. Sure, IA64 is a winner for the teraflop oriented supercomputing community, but for the rest of us, integer performance matters more. And for price/performance, x86 and x86_64 beat ia64.
The Itanium was designed to change the way processors worked. Most processors today are some sort of dymically scheduled behemoth that are capable of detecting instruction collisions on the fly, and reordering instructions for optimal parallelism and thus performance in the light of those collisions. Itanium takes a completely different approach. It is an extremely wide processor that has absolutely no collision detection or reordering. All of the work in this respect is placed on the compiler's shoulders. In theory, a good compiler could make this chip very, very fast, and in reality, as you see, this can be the case. So why did it fail? Intel hyped the hell out of this processor, and then missed their release date by a full two years. That is microprocessor suicide in the land of Moore's law. So, when Intel delivered a chip too late that failed to perform the way they marketed it to, the chip died. In recent years, Itanium has really come around, but it's hard to escape your past in this industry.
Other relevant problems for adoption are tied to this need for a good compiler. Making a compiler as smart as it needs to be for Itanium to live up to its potential is not cheap, and Intel is not known for just giving away such technology. I'm sure the fees to license Intel's compiler are nontrivial, and that does not encourage development. Realistically, Itanium will never become a desktop chip just because of the massive adoption effort that would go into such a switch.
One thing to note, however, is that other chips aren't that far away. You suggest that a 2ghz or 3ghz Itanium would be incredibly fast, and I agree, but I seriously doubt Intel can ramp it that fast. Also, the Opteron specs you show are for 2.0ghz, and I believe Opteron is up around 2.6 or 2.8 ghz nowadays.
Ultimately, Itanium is a great design, but wrapped in a poorly executed initial implementation. It does teach a good lesson that compilers can really help improve chip performance, and down the road, architectures that take this into account may reign supreme. But I wouldn't look to Itanium to do any more than instruct us for the future. She is not a desktop chip.
I'd rather be cycling.
Peter.
Now show me how it compares against a real CPU. ... I wouldn't be surprised if an n GHz MIPS stuffs an n GHz Itanic into the floor.
Guess what? It doesn't. Itanium really does outperform MIPS and if you'd care to look it up yourself, you'd see. Itanium and POWER have been rougly neck and neck in vying for the top performance spot since the Itanium 2 was first released. Each new processor from either vendor bests the other.
As for your disparaging remarks about X86, consider that it offers the highest performance outside of Itanium and POWER on floating point and overall keeps pace on integer code. Topping X86 is, believe it or not, a real feat. Top of the line AMD64 and Intel chips are engineering marvels as far as processors go. MIPS certainly can't touch them.
It may be fashionable to dis X86 but if you look at the numbers and the microarchitecture, you'll be hard pressed to find anything significantly better.
In my case, this is actually a load of bull.
I've never had a significant problem with a CPU that couldn't be pointed at some external issue with the platform - be it cooling, the MB chipset sucking, etc. Things that Intel -does- try to "Do Right" by us on.
That said - I use AMD CPUs pretty much exclusively in my work. Xeon 64bit wasn't there when we needed it (Feb of last year) and Opteron beats the shit out of Xeon on the stuff we do and finally, I can get -real- dual-core Optys -now-, not bolt-ons using the same damned broken shared bus that Intel x86 has been beating on forever.
Price, as a matter of fact, is no longer a consideration. I know I'm paying more for AMD. I'm also getting the product I want and need.
Itanium -could- actually fit into my DC. Everything we do is compiled by us, locally, and it's entirely possible that we'd benefit from it. However, even at 2k/CPU I can't justify the cost when we're looking at driving the CPU number up (our jobs are of the type where the more chunks you can divvy your dataset up into, the faster things get done - so the more processers, the better - it doesn't matter if that Itanium gets my process done in 45hours while it takes that Opteron 55 hours if I can throw twice as many Opterons at the problem and get it done in 30 due to smaller data chunks).
From TFA...
That $2000 buys the processor, alone, and I don't believe I know any gamers that buy processors in lots of 1000.
The entire issue of price and performance is moot, however. The severe restriction to acceptance of the chip for the market is the ia64 architecture (with practically non-existant emulation performance). It is very difficult to find commercial software (even HPC software) available for the system, so you're limited to in-house programs and open-source programs (most of which need to be tweaked and rewritten). The platform isn't deployed in sufficient quantities to create enough demand for commercial vendors to bother supporting it yet, so you end up with the chicken-and-egg problem that most Linux gamers are all too familiar with.
That's a neat trick. Considering the AMD64 architecture was released about 2 years after the first Itanium.
While you're right that Itanium was meant to keep AMD out of the market by essentially creating a new "proprietary" architecture, it's hardly made to be different from amd64 "on purpose" since they didn't really have any clue what amd64 was (since it didn't exist).
Itaniums do run 32bit applications. At least for Linux.
However, if your running a 32bit OS on a 64bit machine, something is not right.
Out of curiosity, I just checked itanic prices at dell. The cheapest configuration for a single (dual capable) 1.5GHz itanic with 2GB RAM and 36GB SCSI HD is over $17K. For comparison, a similarly configured 3.6GHz Xeon (also dual capable, 2GB RAM) is just over 5K.
The article poster is simply trolling. Where the fuck can you get an itanic for $2000? The cpu *alone* costs that much! The article that the moron linked to confirms this: "The 1.4GHz Itanium 2 comes out Monday for $1,172 in 1,000-unit quantities. A 1.6GHz version comes out in May for $2,408 in similar quantities." (last paragraph)
Need I give any more reasons for why it's not popular?
___
If you think big enough, you'll never have to do it.
Your quoting FP performance. The "integer" (aka general purpose) performance isn't nearly as competitive. This is because its a static VLIW machine, and its hard to write a good VLIW compiler. Writing fast FP code is simpler. Then there is the fact that the Itanic is 3x the hardware of the machines your comparing it to. Bigger caches, and all that. Your misunderstanding of clock rate is also simplistic. In order to get the Itanic faster they would have to create a longer pipeline, this would more than likely decrease the IPC and keep the processor from scaling lineraly.
Basically it was pointless. we don't need yet another processor targeted into the same market the POWER64/SPARC64/PARISC and now the X86-64 etc are in.
The whole arch is a mess in my opinion its accually probably worse than the x86, this is evident in how long it took to get the thing out the door. For a processor based on the idea that superscaler wasn't easy and wouldn't perform its beginning to look like the itanic is accually in that boat. Its a dead arch, there are orders of magnitude more x86-64 machines out there even though the itanic had a two year lead. Why should I use itanic when there is a larger software base for PPC/POWER and its multivendor?
POWER is cheaper,faster and more mature and it can barely compete with x86 in the desktop area. ARM has pretty much taken over the smaller chores (cellphones, PDA's MP3 players etc..) and smaller chips like the 8051 clones sit below that.
Give it up, it was stupid, Intel was wrong. My opionion is that itanic was a marking plan to lock up the processor market. If we were all forced to run itaniums back in 96-98 then we would all be buying intel chips for everything. Instead intel had to release the P-Pro to keep ahead of Cyrix/AMD, only they never got far enough ahead to kill AMD to release the pressure and transition everyone to Itanic, where theyhold all kinds of patents and copyrights on the instruction set. Plus they couldn't make the thing work and it slipped for 5 years.
itanium (itanic) is a poor design for anything other than numbercrunching. It is a relic of theoretical supercomputer designs that were popular in the late 1970s. itanic shines on floating-point benchmarks, and is mediocre at best on everything else.
Since the late 1970s, we have had RISC and then superscalar RISC, some now with elements of VLIW. This provides better real-world (general-purpose) performance using substantially less power and fewer transistors than itanic.
Modern RISC processors (including x86 which are RISC internally) can reschedule execution of instructions dynamically (i.e. at run time). itanic can not. It relies on the compiler to schdule the code. It is only possible to schedule code well at compile time for very well-defined problem sets i.e. floating-point maths intensive programs like numerical simulations. NASA currently owns 5% of the world's itanic processors (in a single machine).
itanic was intel's attempt to kill the 64-bit RISC market, putting all of its competitors out of business. Like all great megalomaniacal plans, it has failed. It was a marketing-driven processor, and a failure.
It can't compete with clunky old UltraSPARC IV on server-oriented workloads. Even that market, which isn't big enough to sustain Sun and its processors, is orders of magnitude bigger than the market in which itanic has any relevance.
For big servers nowadays, you have a choice between Opteron and POWER.
In science and engineering, you're often better with something like Opteron, POWER or something fancy from Cray, NEC or Fujitsu. itanic runs hot and consumes too much electricity.
Has anyone ever seen one? I haven't. There was one at a show once on the Red Hat stand, but they wouldn't let me performance test it... and they wouldn't even let me see it because it had over-heated.
itanic is about the most expensive turkey in computing history.
Stick Men
The Pentium Pro was not really a flop. The PPro was faster and better then the Pentium. Operating systems such as Linux, FreeBSD, OS/2 and WindowsNT all worked very well on the PPro. The problem was that the PPro was optomized for 32bit opperations. Software that used 16bit opperations took a performance hit (25% slower IIRC). This was a big problem for Microsoft as Windows 95 still used a lot of 16bit code (as did 98 and ME). At the time Microsoft had no plans for a consumer based 32bit operating system. Intel was basically forced to redesign the PPro for better 16bit performance and the Pentium 2 was born. In the process some of the more server oriented features of the PPro were stripped such as SMP and 36bit addressing.
The intel 386 cpu was a 32 bit platform.
The joke was derived from this...
A fine is a tax you pay for doing wrong and a tax is a fine you pay for doing all right.
The intel 386 cpu was a 32 bit platform.
The joke was derived from this...
A fine is a tax you pay for doing wrong and a tax is a fine you pay for doing all right.
It's not popular because previous generations were Hot, Expensive, and Hard to Program. Adoption was slowed mainly by #2, as it was a significant investment for many research groups to even put one on the floor for testing purposes. The Hard to Program meant that you had wonky versions of Linux, or HP-UX with new compilers, as your OS options, which just increased the resistance.
I ran an Itanium-2 cluster (and had briefly an Itanium-1 loaner) and if the compilers were stable, the first generation easily outpaced the UltraSparc-III systems of the era, and the Itanium-2s kept pace or outpaced the Alphas. However, their power-draw was unbelievable, you had to rebuild the GNU toolchain from RedHat stock, then install whatever that week's version of Intel's C/Fortran compiler was, and on any code that was integer bound, they were barely faster than the late Xeons.
Ah, but for floating point, nothing touched them, except possibly the late Power systems that we couldn't afford. A clean 64-bit architecture, and four floats per clock cycle, meant that for a certain class of problems, there were the cost-effective solution; US-III prices, and twice (at least) the performance. They were even pretty easy to port our codes to, as their pure 64-bit environment (8-byte Integers and pointers), meant that generally modifications to run on Alphas and Crays would run on the Itaniums as well. We would have preferred a next generation of Alpha, but the Itaniums work quite well for what they do.
Discouragingly, now for most (but not all) of the simulations I do, AMD-64 or even EM64T systems are cost-effective, and just as fast. With this new price reduction, I may have to look into adding IA-64s back into the mix, though they're going to have trouble competing against IBM's linux-based Power-5 systems, which offer similar prices, and a more well-known and supported architecture.
the more accurate the calculations became, the more the concepts tended to vanish into thin air. R. S. Mulliken
Sure, an Itanium will run all your existing 32-bit stuff...in compatibility mode, which means you get performance akin to a 300MHz Pentium-II on your $2000 CPU. Remind me again why I'm supposed to buy Itanium?
But to return to seriousness again for a moment, the Itanium isn't pitched at mainstream anymore, and it's debatable whether it ever was. It's an entirely new ISA -- and a very good one at that -- and software developers just didn't see a good reason to jump on it when cheap x86 CPU's were selling like hotcakes.
Intel would've loved to have forced the entire industry to move to IA64 years ago. If it had done so before the Athlon XP ever hit the scene, it's possible the chip giant could have pulled it off. However, with the advent of the Athlon XP (and MP's as well), if Intel abandoned x86, AMD would be there to pick up the pieces, giving customers the option of (a) continued use of their paid-for apps and paid-for OS's on a cheap, fast, x86 chip or (b) loss of all practical use of your 32-bit apps and OS's, total rewrites and recompilations of all core software bits, all on a $2,000 CPU. It's quite clear why Intel didn't try to do such a stupid thing.
So, on the one hand, we can thank AMD for giving us cheap, fast CPU's that run pretty much whatever you want these days. On the other hand, we can thank AMD for keeping us stuck on x86 to begin with, for without AMD we'd almost certainly all be on IA64 today. But, since I like competition, I can say I'm extremely glad things turned out the way they did. IA64 would've been the death-knell for AMD and any other kind of competition, and Intel would be milking us for all we're worth today if it could.
In the end they will lay their freedom at our feet and say to us, Make us your slaves, but feed us. - Fyodor Dostoyevsky
In addition, real-world performance sucks without tons of cache and memory bandwidth. In fact, the original Itanium's entire bus and cache subsystem were redesigned for the release of Itanium 2, doubling the bus width, increasing the L2 cache size and tweaking the latency on the L3.
No surprise, the Itanium 2 performs much better than the original Itanium, but it's name was alreay soiled by the mediocre preformance of the original.
In addition, all that cache and high-performance bus architecture means the Itanium can't be priced competitively with anything, due to the hulking die size. I2 NEEDS a minimum of 6MB L3 cache to shine, and 9MB doesn't hurt. Note that the "affordable" I2 chips have a paltry 1MB L2 cache, and perform like crap.
Man is the animal that laughs.
And occasionally whores for Karma.
If not I think you better go watch The Life of Brian and come back to remove the egg off your face.
My particular love for this joke is irony of looking into a room of Monty Python fans and they are all sitting there chuckling and repeating that very same line.
Yes, yes we are all individuals!
Classic!
I think the biggest thing that doomed the Micro Channel Architecture (MCA) was the fact that IBM did not bother to license the technology at very low cost.
If IBM had done a proper job of licensing MCA at a low cost then not only would MCA have replaced the old ISA bus, but alternative bus connection architectures like EISA, VL-Bus, PCI, AGP and PCI Express would have never happened! This is because we know now that MCA could be easily expanded all the way to 64-bit bus connections and support very fast bus speeds. Indeed, if you look at PCI and PCI Express connectors they strongly resemble the old MCA connectors in physical design.
Itaniums do run 32bit applications. At least for Linux.
Yeah, and the speed sucks ass. Even Intel recommends against running any 32bit code on the Itaniums.
I think the reason it's not doing as well as they hoped, and compared to the benchmarks, is because, in real-world performance, they're not that great. We've got 13 itanium boxes in house, all of them cost huge $ (we can't get the cheaper low-power ones, we need speed above all else) and their usage levels over the last few months has bottomed out.
Before tools were available for the amd64 line, the IA64s were used pretty heavily by our engineers. Especially when near the end of a chip design, because they needed to be able to do things like have a single process malloc over 40Gig of ram.. yes, I said 40Gig! Seeing that in top on a linux box is something. The speed compared to SPARC boxes is a huge jump. However, now that the tools are all coming out with amd64 versions, they're all moving there. We can get boxes there with huge ram, and they get 64bit, and they're seeing better performance. This may be at least partly due to some steps in their batch runs being 32bit.. which can't really be changed right now, so the ia64 suffers heavily then.
So who cares what the benchmarks say.. all that matters is how the apps the engineers need to run perform. Today, amd64 kicks the ia64's ass.
- My favorite error message: xscreensaver, running on an old Sparc 5 w/ 8bit color: bsod: Couldn't allocate color Blue
You know, I wish people would get this right. The Itanium was not *started* by Intel. HP created and began the work on the ISA as a successor to the PA-RISC line, and when they realized that they didn't have the design resources or manufacturing capacity to build the chip, they partenered with Intel. Since then, HP has been the big partner with Intel- first to get silicon, first to market. HP always intended to go forward with the VLIW setup, and depreciate PA-RISC. The Compaq merger just gave them another architecture, and they decided to kill it- Alpha (RIP).
So, what's absolutely hilarious is that HP isn't the one getting fucked, or making the wrong decision- they actually pawned the risk onto Intel, and now Intel is the one in trouble if this arch doesn't fly.
Very strange stuff.
Fluid simulation, finite-element stress analysis, nuclear warhead testing simulation, galaxy formation...all of these types of calculations are fairly easy to make parallel and will keep the Itanium pipes running quite efficiently with Intel's VLIW compiler.
That's all well and good if you're a government researcher with access to an Itanium-based supercomputer. But the rest of us don't spend our time testing nuclear warheads; we're just reading email, surfing the web, working on office documents, etc. These types of tasks haven't been shown to perform very well on an Itanium system; in fact, they don't seem to work that well on Intel's lower-end CPUs using the Netburst architecture either, because of the high cost of context switches. For most people, what we need is a CPU with multiple cores, fast context switching, and a nice-sized cache so we can have lots of web browser windows open along with our music player, email client, word processor, etc.