StarDestroyer.Net BBS

Starglider wrote:Damnit I'm still waiting for some proper K10 benchmarks so I can get on and buy a new workstation (choice between Intel and AMD depends on what performance on the new chips will actually look like after they've finished tweaking the silicon).

The Kernel wrote:At this point the likelyhood of Barcelona being a performance winner seems pretty low.

Starglider wrote:

The Kernel wrote:At this point the likelyhood of Barcelona being a performance winner seems pretty low.

Sadly this is true, but I have been a very satisfied AMD customer for several years (x86-64 was extremely helpful for what I've been doing, without AMD taking the initiative Intel would have kept trying to push their Itanium crap and there still might not be a cost-effective and easy to deploy 64-bit option by now) so I'd like to give them the chance to surprise me.

Starglider wrote:Sadly this is true, but I have been a very satisfied AMD customer for several years... so I'd like to give them the chance to surprise me.

Loyalty to a chip vendor seems rather silly to me. Just buy whatever the best price performer is out there

And btw, there is nothing wrong with Itanium in principle.

It is an excellent 64-bit architecture

and the only real problem is the lack of full speed backwards compatibility with x86.

phongn wrote:Damnit, Starglider, you beat me to the anti-IA64 essay

To be fair, though, in some scientific workloads it does very well.

Starglider wrote:I've written actual essays on the subject many times. At this point I've euthanised so many Intel-fanboy-Itanium-wank threads that I'm just summarising. I was really excited about Itanium too circa 2000 (it helped that I was seriously studying microprocessor design for the first time), but it turned out to be a pile of junk while AMD came out of nowhere and did the impossible; made x86 actually not suck.

In a very few scientific workloads it does a little better than POWER 5. This advantage is likely to be removed by POWER 6, which is in all respects a nicer platform anyway. However the trend is very much towards heterogeneous multicore designs that integrate massively parallel stream-optimised cores with highly serial branch-optimised cores on the same die.

This will far outperform Itanium on all the things Itanium is actually good for, and it sucks as a control processor. Two further facts (beyond the wild optimism with regard to compilers) have been working against it; steadily increasing transistor counts make all that complex prediction & out-of-order logic that Itanium was so keen to throw out not so expensive after all (in fact a major reason Core has such good IPC is even better prediction & out-of-order logic than the P4), while the overheads unique to the Itanium design proved to be much worse clock scaling barriers than the congenitally optimistic initial projections suggested.

Fingolfin_Noldor wrote:I wonder if Intel will finally drive prices down. I was hoping to do an upgrade on my computer come fall.

I'm not familiar with the problems of the x86 design, but what are they?

Chris OFarrell wrote:Think of a CPU like a basic lego block, nice and clean and well designed.

Then spend a decade or two randomly adding on other lego bits and random lego bits to those lego bits.

Starglider wrote: There are a lot of things wrong with Itanium in principle. The design is based on a series of broken assumptions. It's like a PhD student's 'novel processor design' thesis that somehow got made into a decade-long industry trainwreck. It sounds clever until you actually look at trying to run real code on it. The very few things that it does well are done far better by true vector processors and Fusion-type architectures will render it utterly irrelevant.

No, it is not. It only eventually managed to achieve semi-acceptable performance by having stupid amounts of cache tossed at it on huge bloated dies. The logic never worked properly and utterly threw away the basic RISC advantage that the design was supposed to be building on, resulting in pathetic clock speeds. It never achieved good price/performance even if you can write custom software for it. It is an absolute bitch to write compilers for and pretty nasty to program in assembler if you're actually shooting for maximum performance (x86 is crufty, but at least it's tolerant). By comparison SPARC is a great architecture from a programming point of view; POWER is nearly as good and has had a string of excellent microarchitectural implementations recently.

Itanium failed for many more reasons than that.

Uraniun235 wrote:I think we could see a transition to x86-64 sometime in the next several years.

phongn wrote:

Uraniun235 wrote:I think we could see a transition to x86-64 sometime in the next several years.

It's slowly happening, mostly because both Intel and AMD's current and future plans all have x64 processors. It doesn't cost all that much more to produce and Microsoft is planning on dropping x86 in the next major version or two of Windows.

phongn wrote:

Uraniun235 wrote:I think we could see a transition to x86-64 sometime in the next several years.

It's slowly happening, mostly because both Intel and AMD's current and future plans all have x64 processors. It doesn't cost all that much more to produce and Microsoft is planning on dropping x86 in the next major version or two of Windows.

Fingolfin_Noldor wrote:

phongn wrote:

Uraniun235 wrote:I think we could see a transition to x86-64 sometime in the next several years.

It's slowly happening, mostly because both Intel and AMD's current and future plans all have x64 processors. It doesn't cost all that much more to produce and Microsoft is planning on dropping x86 in the next major version or two of Windows.

Oh? What are they switching to?

The Kernel wrote:I think that the only reason you'll see the transition is because of memory limitations...there seem to be no other reason for anyone to actually make the move.

phongn wrote:My graduate computer architecture class went over VLIW in some detail - and the issues with compilers is just too overwhelming.

I was hoping back then for a new "clean slate" to replace the mess that is x86 but it was not to be SPARC, MIPS, POWER - all clean ISAs (hell, 68K was pretty clean, too) but almost all crushed by the might of the X86 juggernaut

I'm not so sure about heterogeneous multiprocessing at the CPU-level, to be blunt. I don't really see the advantages there compared to conventional homogeneous designs, especially with such huge transistor budgets as we have now.

Well, conversely, POWER6 is an in-order(!) design for integer and vector processing - and its SPEC scores are trouncing everything else out there. Sometimes there is something to be said about getting rid of that expensive OOOE logic

The Kernel wrote:At this point it would probably be difficult to argue with that, however we have yet to see vector or GPU processing go much of anywhere with the HPC crowd. Time will tell I suppose.

Itanium has achieved a modest success (not given the initial investment, but that's a sunk cost anyway at this point) with the HPC crowd and these are a very fickle bunch, most of them buy the fastest performing machine and don't give much of a crap about compatibility. The fact that Itanium can compete in this market at all proves that it is price/performance competitive.

Also I should point out that Itanium doesn't have the R&D resources thrown at it that x86 does.

You can't even get x86 users to effectively switch to x86-64

I think that the only reason you'll see the transition is because of memory limitations...there seem to be no other reason for anyone to actually make the move.

Starglider wrote: You are kidding right? You do realise vector computing was supercomputing from the Cray-1 onwards right? Massively parallel clusters didn't begin to come in until the late 80s (Connection Machines actually kicked that off but then utterly failed to capitalise on their lead for internal politics reasons). Even now there's a considerable market for custom, dedicated wide-vector processors. However ultimately hoards of small FP cores with SSE equivalents are going to ouperform them. GP-GPU has a lot of people very excited and is getting a lot of press but it's taking a while to ramp. Ultimately though die level integration of massively parallel FP crunching is more flexible, rather easier to use and for many applications faster than FPGA-based co-processors, the previous 'hot trend' that was getting a lot of press and attention.

Itanium barely competes in this market. If it wasn't for the fact that competing architectures (e.g. MIPS, Alpha) were withdrawn just out of fear of Itanium and several supercomputers would not sell you anything else for several years, it wouldn't have gotten any traction at all. As it is Itanium is rapidly on the way out, partners have been steadily jumping ship and Intel has been cutting resources dedicated to it.

ZDNet wrote:Even before Intel launched "Montecito" processors to much fanfare last month, the Itanium Solutions Alliance has been cranking up its engines in anticipation of strong demand for the platform. And, Hewlett-Packard is proving to be its biggest supporter so far.

By the end of the year, the Itanium Solutions Alliance (ISA) expects to have 10,000 applications available on Itanium systems, up from 8,200 currently. The ISA is a partnership formed by major tech companies with the aim of speeding up the development and adoption of software for Itanium-based platforms.
Advertisement

William Wu, who chairs the Asian arm of the ISA, told ZDNet Asia: "The rate of ported applications is phenomenal and a real sign that the industry's appetite for Itanium is getting larger than ever."

In the first six months of 2005 alone, he said, Intel added 1,000 applications onto Itanium, including Oracle's e-Business Suite and Fusion middleware, SunGard's financial services applications, Microsoft's SQL, SAS' v9 and Symantec Storage Foundation.

Currently, ISA has 11 global centers for ISVs to access development tools and Itanium hardware systems. By the end of the year, there will be 20 such centers, said Wu.

However, Rajnish Arora, IDC Asia-Pacific's enterprise server and workstation research director, noted that while Intel may claim to have big numbers in application support for Itanium, the true picture may be clearer if one discounts the applications previously available on the HP-UX operating system running on PA-RISC and Alpha machines.

According to IDC, HP is the primary vendor selling Itaniums systems in the Asia-Pacific, excluding Japan region, accounting for 95.2 percent of total unit shipments and 91.8 percent of revenues sold in the first quarter of 2006. Other Itanium vendors include Silicon Graphics (SGI), Unisys, Samsung and Bull.

In total, 1, 914 Itanium systems were shipped in that quarter, raking in revenues worth US$73.6 million for the region.

Arora said: "I would ask how many, out of all Windows, Linux and Unix ISVs, [has Intel] never had an engagement [with previously]. The number of new ISVs would be a good indicator in gauging how much new success [Intel has] broken with Itanium."

It is one thing to get existing ISVs running software on HP-UX or Alpha Unix and move them to Itanium, and another to encourage "brand new ISVs" which had never previously offered their applications on HP systems to port their applications on to HP-UX running on Itanium systems, he pointed out.

"It's largely a HP play in the market," he said. "Apart from HP, there are not many other vendors which can design, market and sell [enterprise-class Itanium] systems."

Arora added that in the Asia-Pacific region, Itanium-based Unix server sales are primarily driven by existing PA-RISC/HP-UX and Alpha/Tru64 Unix enterprise clients that are either migrating existing workloads or deploying new ones.

"The number of customers migrating off Power and Sparc to Itanium is still a fairly small number," he said.

According to IDC, Itanium-based systems grew 41.8 percent year-over-year worldwide in the first quarter of 2006, generating US$640 million in revenue. This segment now represents 11.2 percent of all non-x86 server revenue.

In addition, there are other key players in the overall server market apart from HP.

Roadblocks ahead
According to IDC, in the first quarter of 2006, HP and IBM tied for number one position with 28.1 percent and 27.9 percent share, respectively, in terms of global server revenues.

Dell Computer maintained third place with 11.1 percent market share in the same quarter. Although the company's year-over-year revenue growth slowed to 3.6 percent for the quarter, Dell managed to gain new market share.

Sun experienced a 5.8 percent year-over-year revenue growth, and increased its overall market share to 10.8 percent, from 10 percent in the same quarter last year.

Intel, which pits Itanium processors directly against Sun Microsystems' Sparc and IBM's Power chips, claims the company has made headway against its competitors.

Citing IDC figures, Intel's Wu noted that in the first quarter of 2006, global revenue from Itanium systems represented 45 percent of revenue from Sparc systems and 42 percent of revenue from Power systems. In fact, he added that Intel's fourth quarter-2005 Itanium system revenue surpassed Sparc in both Japan and Korea.

Arora noted though that uptake for Itanium servers on Windows and Linux has slowed down significantly, primarily due to the performance boost that x86 servers have had with the introduction of new technologies such as 32-/64-bit hybrid systems and dual-core systems from AMD and Intel, in the past two years.

IDC Asia-Pacific figures found that in 2005, 58 percent of Itanium systems were sold with the Unix operating system. In the first quarter of 2006, this number increased to 72.2 percent.

In comparison, about 30.4 percent of all Itanium systems sold last year were on Linux. By the first quarter of this year, this dropped to 18.4 percent. Similarly, 11.6 percent of Itanium systems sold last year ran on the Windows platform, but this number dropped to 9.4 percent in the first quarter of 2006.

Without the support of other major server vendors such as Dell and IBM, it seems that for now, at least, HP will remain Intel's biggest ally in pushing Itanium.

Arora explained: "HP's large share of Itanium server sales both in the Asia-Pacific region and worldwide, has further expanded in the past one year after IBM and Dell decided to exit, at least temporarily, due to lackluster demand for Windows and Linux workloads on Itanium."

Itanium had a ridiculous amount of R&D resources thrown at it, far more than the Opteron got or that Sun or IBM throws at their processors. The only design that got more was the P4, and I'm not even sure about that; sources inside Intel consistently report that all the best talent went to work on Itanium in the late 90s and early 2000s. Stop trying to be an apologist. Itanium did not work and attempts to make it competitive were the processor equivalent of lipstick on a pig. It is dying and will probably be gone by 2010.

Outright lie. Linux supported it at launch. Fully x86-64 servers were possible within a year, and many new servers and HPC installations were in fact to take advantage of the large flat memory addressing. Vista is 64-bit by default, 64-bit binaries are now the norm even for such consumer things as games and Microsoft Office.

Ah, and it finally becomes clear that you are a pathetic Intel fanboy. 64-bit integer processing is extremely useful in many applications, such as cryptography and anything involving large bitfield manipulation (I use 64-bit integers pervasively in our code). The doubled register set is helpful for performance in nearly all applications and easy for compilers to take advantage of; x86 is actually register bottlenecked, whereas IA64's grossly inflated register set is too big to get much use out of in most situations.

phongn wrote:

The Kernel wrote:I think that the only reason you'll see the transition is because of memory limitations...there seem to be no other reason for anyone to actually make the move.

It's a reasonable enough reason, though. The added GPRs and ditching of x87 is just a bonus, too