Bob Gezelter writes: On December 16th, an article was posted on OSNews that stated, in effect, that HP was “Exiting Itanium”. A careful review of the facts suggests that this press report was based upon an incomplete understanding of the HP-Intel arrangememnts. I have just published an article on OpenVMS.org, based directly on public published information, containing a more complete reprise of this week’s announcements. Update: HP will be investing $3 billion on its Itanium-based server line.
Thank you to post it. I’m hoping this announce will end the buzz that started with the previous announce.
I am mad to see the press is using more the Itanium subject at the same level of the relation between Brad and Jennifer instead on focusing more about the platform itself.
It is also interesting to see that press often point out a bad communication to explain the current situation of Itanium while there is more and more people asking for more specific details and there is no real journalistic effort on trying to answer these concerns.
is openvms still being developed? and future versions to be released?
Of course. It was one of the few good things HP actually got out of Compaq. The rest of Compaq was just duplication and headaches.
But to continue the thread in the previous post about writing compilers for the EPIC architecture someone asked *why* it was hard.
I’m sort of curious myself – I’m no wizzard but it strikes me as a bit odd (having studied the Itanium in depth and knowing a few bits and pieces of compilers). I’m not saying that this is not the case – I’m mostly curious and would really appreciate a pointer or two. Would make up some nice christmas reading/pondering
It is a long subject and it would be hard to answer in this place. I will try to give a hint but you can contact me by mail. I can put you in contact with people involved in this area (compiler for Itanium). They will be happy to answer you.
It is hard because as you probably know, the target code is not similar as it is for other architecture. Look at the current litterature about this subject (designing a compiler). All of the internal representation are more or less the same and are perfectly suitable for target as ia32, powerpc and others.
Now the notion of bundle make the stuff far more difficult. Even gcc is not at the top with this. That’s one of the difficulties.
HP also got some stronger brand names e.g.
Ipaq was a far stronger and better known brand than hp’s jornada brand.
“…But to continue the thread in the previous post about writing compilers for the EPIC architecture someone asked *why* it was hard.”
It’s not hard to unusually hard to write a compiler for IA64. It is hard to write a good one, however. Why?
The short answer: IA64 was designed around several premises; two important ones: 1) that instruction level parallelism is plentiful in regular source code, and 2) that it is easy to detect it in a compiler. Neither of those premises are universally true.
If everyone’s code was still written in Fortran, then yes, it would be a lot easier to find ILP with an IA64 compiler. But you won’t find as much ILP as IA64 needs to be considered “efficient” on ordinary C/C++ code that is filled with lots and lots of branches, indirect and dependent function calls, and dependent traversal of linked data structures.
IA64 would make a great $20 DSP. ;^)
1) that instruction level parallelism is plentiful in regular source code
Wrong: That if you want more instructions executed in 1ms, you need to execute them in parallel
that it is easy to detect it in a compiler.
Wrong: That the compiler has a better view ahead to move from the back to the front a non-conflicting instruction that can be executed in parallel.
Neither of those premises are universally true.
None of your statement are either.
However, the paragraph following is correct.
For the price/quality you should discuss with some of the institutions who acquired so many Itanium processors.
Wrong: That if you want more instructions executed in 1ms, you need to execute them in parallel
Just to be clear: what I said was that I believe IA64 was designed around at least 2 faulty premises. I believe that ILP is *not* plentiful in regular source code, and I believe that it is *not* easier for a compiler for a language that supports aliasing (C/C++) to detect and schedule conflict- and dependency-free instructions.
regular source code != (SPECfp, Linpack)
For the price/quality you should discuss with some of the institutions who acquired so many Itanium processors.
I have. I have worked in that industry (both in academia and as a vendor) for nearly 15 years, and at least some, perhaps all, of the larger IA64-using shops got Very, Very Good Deals.
Do you really thing that a large company (make that two large companies) don’t have a proper roadmap for product adoption?
How about this for an idea. HP and Intel developed Itaniums. Other companies who want to use Itaniums will be competing with HP, because either they make large iron servers (like SGI) or want to develop medium scale servers.
So, I’m SGI. I compete with HP in the marketplace for servers. I have to buy my processors from my competitor? How logical is that? Not much. Now, no one has a problem buying processors from Intel, right? (works for dell!). Now all of the itanium staff will work for Intel. Seems logical for me, and in fact this can even HELP Itanium adoption.
Those who claim that Itaniums are a failure don’t have any market data to back it up. Yes, they didn’t meet projections, but selling only 100,000 servers instead of 400,000, is that truly a failure? Nope
[i]Yes, they didn’t meet projections, but selling only 100,000 servers instead of 400,000, is that truly a failure? Nope[i]
It’s a failure if they never make back the money that was spent on R&D, which at this point is in the billions. That’s why analysts are so bearish on Itanium, because it was such a massive risk for both Intel and HP and it doesn’t seem to be panning out. Business isn’t just about making sales. You also need to balance between capital expenditure (R&D in this case) and revenue.
for a language that supports aliasing (C/C++) to detect and schedule conflict- and dependency-free instructions.
Thus I should redirect you to OpenImpact guys whose pointer analysis is strong (and fast) enough to put the headhach away.
got Very, Very Good Deals.
I hope so. When I can get 30% off for 60 workstations (as an academia), I’m hoping they are getting better than me for alf a dozen of thousands of processors. And I am pretty sure other vendors offered interesting price too.
To work close (bost physically and spiritually) with one of them, they appreciate the platform too. Not only the price. Not far from quoting them, they started to like it once they understood gcc accuracy was perhaps nice, but efficiency was just junk.
This announcement is yet another proof that Itanium is one step closer to become dead meat. It looks like there are absolutely no hopes that Itanium prices are going to go down, prospects of volume growth for Itanic are practically non-existant. Even provided that the volumes go up (which they never will) Intel will never be successfull at making the processor cheap enough to compete with Opteron or even Power and UltraSparc. Itanium die size is so huge compared to the competition, that the yeilds for Itanium will always be lower than the competition making the processor intrinsically more expensive than anything else out there. IBM and Sun will always have advantage in this area as they already have volumes and the economies of scale on their processors. Plus both IBM and Sun will always be more successfull at recouping the processor development costs as they can tie the processor costs to the systems costs (they build systems based on their own processors). Both Intel and HP are at a huge disadvantage right now and these face saving measures ($3bn) are not going to change the situation — Itanic most likely will be dead soon.
Maybe HP wants to disengage from IA64 and Intel doesn’t want to see 300 CPU engineers leaving the IA64 world to finish in PPC Wonderland……
poundsmack,
As I noted in my article, 8.2 for Alpha/Integirty (IA64) is being released on or about the end of this year. As I am not inside HP, I don’t recall having heard the exact date, only an approximation. Having used the Field Test vesions (see my article on OpenVMS.org at http://www.openvms.org/stories.php?story=04/12/07/0088240 , at least that I can admit to), it looks pretty good.
The VAX version of 8.2 will ship later. 8.3 is already in progress. The ongoing OpenVMS Roadmaps (as I write this, the latest set is only one month old) can be found at http://h71000.www7.hp.com/openvms/roadmap/openvms_roadmaps.htm.
In terms of meeting quality and schedule goals, OpenVMS seems to be pretty much on their published timetables.
I hope that the above is helpful.
– Bob Gezelter, http://www.rlgsc.com
This announcement is yet another proof that Itanium is one step closer to become dead meat. It looks like there are absolutely no hopes that Itanium prices are going to go down, prospects of volume growth for Itanic are practically non-existant.
True, the problem with Intel is that they’re unwilling to make the investments to encourage no only software vendors to come on board, but to also encourage third party hardware vendors to jump on board.
How many third party motherboard manufacturers are out there for Itanium? none. Can I go into my local hardware vendors shop and purchase an Itanium in a box? I can do the same with SUN, I just simply have to ring up SUN, and order a CPU directly from them, and if I want to purchase a number of system boards, its just a matter of hammering out a contract.
Ring up Intel and they’ll tell you to piss off if you inquire about purchasing an Itanium processor or board for research purposes.
If Intel don’t want to play ball with the developer community, then so be it, but mark my words, they only have themselves to blame when Itanium ultimately fails.
No ISV and IHV network equals failure. No workstations selling in volume equals failure. There is a laundry list a mile long of the number of companies who said, “we’ll contrate on the server market”, and in a few years, they were going under. SGI and their “super computer” market, which is now an ultra niche where clusters are the in thing, and big frigging SMP configurations are no longer giving a decent bang for the buck.
Workstations will always outsell servers, meaning, workstations play an important role in providing volume, and thus allowing the company to reduce the cost of production.
Even provided that the volumes go up (which they never will) Intel will never be successfull at making the processor cheap enough to compete with Opteron or even Power and UltraSparc. Itanium die size is so huge compared to the competition, that the yeilds for Itanium will always be lower than the competition making the processor intrinsically more expensive than anything else out there. IBM and Sun will always have advantage in this area as they already have volumes and the economies of scale on their processors. Plus both IBM and Sun will always be more successfull at recouping the processor development costs as they can tie the processor costs to the systems costs (they build systems based on their own processors). Both Intel and HP are at a huge disadvantage right now and these face saving measures ($3bn) are not going to change the situation — Itanic most likely will be dead soon.
And the fact that both IBM and SUN are creating alliances to drop the cost of development even further, just “watch this space”(tm) with massive price drops of SPARC and POWER (incl. PowerPC) CPUs, to the point that they’re priced around the Xeon price point.
iAMD64 – oops EMT64 – is killing IA64. Nice to see Intel doing Seppuku. IA64 will be targeted to niche segment : HPC and no more, no volume thereby no future!
“Ring up Intel and they’ll tell you to piss off if you inquire about purchasing an Itanium processor or board for research purposes.”
One reason for this, perhaps, is that Sun has long been a direct-to-customer company for their whole product line, while Intel mainly does business with OEMs. The OEM customer base is a wholly different culture than sysadmins ordering hardware for their businesses.