Doug Siebert (foo@bar.bar) on 2/25/09 wrote:
---------------------------
>anon (anon@anon.com) on 2/24/09 wrote:
>---------------------------
>>Downhill as in shrinking market share? Maybe. I don't know I'm not trying to predict
>>the future. But unlike naysayers have been forecasting for years, IPF right now
>>has been growing, now has a very significant market share, and appears to have a
>>pretty good medium-term commitment from Intel.
>>
>>And I don't think it is at all guaranteed that market share will be relegated to
>>PA-RISC replacement. What about SPARC? How long will POWER compete in the high end?
>>Sure many of them will want to try migrating to x86, but as long as Intel segregates
>>IPF into the high end, some will look there.
>
>
>Well "someone" would certainly consider me a nay-sayer, but I never suggested anything

I don't keep track of who is what. But in general, yes I notice there are fewer and fewer loud obnoxious hardcore IPF bashers here.

>other than that IPF would take over pretty much the entire PA-RISC market share,
>most of the OpenVMS/Alpha market share and maybe half the Unix/Alpha market share
>(since they were changing both architecture and OS, the OSF/1 guys had no real incentive
>to choose HP-UX on IPF any more than Sparc, POWER, or Linux)
>
>That's all its done, its "growth" as you define it, has been replacement.

That's "growth" as everybody except you defines it. Look it up.

That's
>like claiming x86-64 has grown by 10000% in the last 4 years, when everyone knew
>that would happen as 32 bit x86 CPUs were replaced by them. Many who bought 64
>bit CPUs would have been fine with 32 bit CPUs, but if they wanted the latest fastest
>CPUs to run x86 code, they ended up with 64 bit CPUs.

Um, I don't really see the logic in your weird redefinitions. Many metrics put x86 in one category so you don't see the difference between 32 and 64. If you put them in different categories, then of course you would have seen large x86-64 growth.

> Its no different for the
>large bulk of IPF sales. People who wanted to run their HP-UX or OpenVMS software
>with the latest and fastest CPUs they could do so bought IPF.

And now that has mostly happened quite well to plan and has been the major contributor to giving IPF a very significant market share, Intel seems like they are willing to put more resources and take some more risks with IPF and see where it takes them.

You think it is a foregone conclusion that IPF market will now slowly shrink back to nothing. I think if Intel saw it the same way, they would not be developing new microarchitectures or bothering with putting it on a cutting edge process. I'd think most of the legacy market could be held with Tukwila and just bumping core count and adding some tweaks now and again.

>
>>I'm not very knowledgeable about the market, but aren't a significant amount of IPF sales windows and linux based?
>>
>
>
>I've worked with a number of Superdome sites before and so far the only time I've
>ever seen Windows run on one was in a little partition just to test. There's some
>Linux, but the huge bulk is HP-UX. Yes, that's anecdotal, but I've never seen any
>hard numbers on how IPF is sold - I remember seeing some numbers a few years back
>claiming that 1/3 of Superdome buyers were running Windows but I suspect that was
>like the site I saw with one partition. I'll bet as a percentage of IPF cores sold
>Windows doesn't amount to more than 2-3%, but we'll probably never see those numbers
>unless someone leaks some internal HP or Microsoft information.

I don't know. I've seen IPF sales breakdowns from a couple of years ago with UNIX at about 50%. You think that if HP sells a superdome with primarily HP-UX and one Windows partition, that gets chalked up as a Windows sale? That would be interesting.

>Where's the incentive to run Linux on IPF? That disappeared when x86 went 64 bit.

At one point, IPF was pretty competitive in HPC and even smaller enterprise installations. Today there is little or no incentive to migrate to Linux on IPF. Possibly Tukwila will create a bit of demand, depending on price, and how it clocks, and how well the core responds to the redesigned memory hierarchy and interconnect. Poulson is in the position to make a big impact (this is not "wait for the next one", it is also possible that will underachieve).

As I said, at a wild guess I think Poulson might be make or break. Either it will not go well, and Intel will move toward putting IPF on life support and scaling x86 up, or it will go well and Intel will continue to proactively invest in IPF and look at taking more high end market from IBM and SPARC.

>Sure, you can buy high RAS hardware for IPF, but anything where x86 doesn't support
>the feature set is either going to be missing or not well tested in Linux, so the
>increased RAS you get from hardware would be lost via software.

What kinds of missing software-activated RAS features do you have in mind? But your generalization is not true of Linux any longer. Linux for example has support for physical CPU hotplug; I think this was pushed mostly by IBM and I'm sure there were POWER and IPF systems that could take advantage of it before any x86. Memory hotplug seems to be driven largely by Japanese IPF vendors. I think it has pretty good support for IPF machine check exceptions, which I think Intel has worked on. Marking bad physical memory via MCE I think comes from SGI (not surprisingly as they have single systems with tens or hundreds of TB of memory).

So no, the myth that Linux is an x86 operating system that is jammed on top of other architectures is... a myth. It has not been true for a long long time.

>>I don't know. But the thing is, the "just wait for the next one" guys are not worse
>>than the "there will not be a next one" guys, as far as I can tell.
>
>
>What about us "the 'just wait for the next one' guys are always wrong" who aren't
>saying there won't be a next one, just that it will eventually be killed as its
>target market continues to shrink.

It's an interesting assertion. I'm not convinced that it is a foregone conclusion that x86 is destined to dominate from netbook to high end server. I suspect that a lot of people who believed that 10 years ago were expecting it to have already happened 5 years ago.

> At first it was sold a bit in workstations and
>was a real player in the HPC market. But the workstation market has long abandoned
>it, and more recently it has become a non-factor in the HPC market as it is just
>too far behind x86 in performance and the 32 bit limitation of x86 was removed.
>All that's left is the HP-UX market, which is not small - Sparc and IBM are still
>in their own similarly sized Unix markets after all - but its nowhere near where
>Intel wanted/expected it to be. I'm just not sure Intel will consider it worth
>its while in the long run, and would probably save money and free up engineering
>effort to increase their x86 lead over AMD by migrating that market to x86 (and
>it'd still be worth it even if AMD succeeded in taking its usual 20%)

Yes it has definitely been going downhill in non-enterprise markets for a long time due to who knows what going on inside Intel. Marketing dept, poor execution, secret agreements with other companies, etc.

As I keep repeating, it seems like Intel is going to have another go at it. If they can execute; a significantly improved microarchitecture, 10-12 cores, on a cutting edge process. It could end up clear performance leader in the high end, and it could be competitive with x86 in some areas like HPC and high end x86 servers.

This is why I don't completely write off IPF just yet. Anybody who does, as I said, is sticking their heads in the sand (and repeating that x86 is faster at specint while ignoring other important factors like Intel's apparent wish to segregate the market and not scale up x86 as far as it could)