someone (someone@somewhere.com) on 2/25/09 wrote:
>
>About 18 months ago IBM mainframe sales cratered, falling
>something like 50% YoY, in the 2 or 3 quarters before the
>last major platform upgrade (z9 to z10). That was during a
>period of sales growth across all server price segments.

Hey, I'm not arguing about IBM's mainframe business. And
it was obviously a local phenomena, as you yourself point
out, so such a cratering was likely due to local issues
(ie the upgrade cycle timing etc).

However, I'd argue that the drop in the last quarter is
unlikely to be about such local issues. All the markets
dropped, they just dropped at different speeds. And I think
the speed differences can be explained (at least to a
first order - details will obviously always differ) by
simply a big economic contraction.

So I don't see what your IBM example tries to do. I think
you're arguing that the IPF drop is about people waiting for
Tukwila, but as mentioned, the drop was clearly across
architectures.

If your argument is that Tukwila will cause a bump, then
I'm not going to disagree - you'd really hope that a
new CPU SKU would bump the sales. We'll see. But that's
really not germane to anything else, so I'm not seeing the
relevance.

>Here we are looking at the steepest drop off in server
>sales since the dot com bust AND just a quarter or two
>ahead of a major platform upgrade for IPF and its sales
>fall by 5% and you think that is a *bad* sign? Clearly you
>see only what you want to see.

Your argument seems to be that "it could be worse". Sure,
it could be worse. That's not my point.

My point was that you seem to argue that the size of the
relative drop is somehow a sign of IPF taking market share,
but you don't seem to account for any long-term or global
issues in your argument.

Do you really think the drop isn't related to the bigger
slowdown? Do you really not think that big hardware has
longer lead times?

Your IBM "suddenness" argument is not relevant, since my
point is not that changes don't happen, but that the timing
of changes is different across different segments. Within
one segment, sudden changes can happen for local reasons,
but when you have a global reason, the timing will differ.

You've been crowing about the "only 5%" number without
ever admitting to those rather obvious issues. Now when
I call you out on it, you try to do your sing-and-dance-
and-let's-avoid-the-issue number.

Also, I suspect that a bad economy is going to be good
for cheaper platforms in the long run. Yes, sales drop
off, but on the other hand you'll also see just continued
pressure for people to simply convert to cheaper boxes.

The same way netbooks are likely to cannibalize laptops
more in a down economy where a cheap netbook is
picked over a more high-end laptop, that same way I suspect
that you're going to find places that just decide to go
with a cheap x86 server (possibly one they already had:
I'm not saying that this will be a sales uptick for x86),
instead of moving up to something fancier.

End result: that high-end market gets just more and
more stratospheric, and the x86 market continues to grow
and become more common in what would have been seen as
"high end" a few years ago.

And sure, if you keep comparing IPF to just POWER, you'll
never see this, because it very definitely happens to both.
But you shouldn't close your eyes to "high end" (as defined
by "not x86") basically shrinking (not in absolute terms,
but in relative ones).

>SPARC was well into new product cycles both for general
>purpose and CMT product lines in 2H08 and but its sales
>fell ~19% in Q4, even worse than x86. Your long procure-
>ment procedure theory sure didn't do any favors for Sun!

Oh, I agree. When it comes to Sun, I think we look eye to
eye. Look at the Sparc press even before this all - I don't
think it was all that rosy.

>Falling only ~5% under such conditions is IMO a sign
>of a great deal of market strength for IPF, certainly far
>more than x86, SPARC, (and probably Power too but IBM
>goes to great lengths to conceal changes in its overall
>Power sales in its financial reports so it is hard to tell).

I do expect that you're right in that IPF is likely less
"volatile", but that's for another fairly trivial reason:
IPF, of all the ones you mention, is the most "locked in"
of all architectures.

People who buy IPF are often essentially forced to do so
because of HP-PA and Alpha EOL issues, and OpenVMS. That's
a fairly large chunk of the market.

In contrast, people who buy x86 and Sparc (and to a lesser
degree POWER) tend to have a lot more flexibility. They
aren't, for example, fighting an EOL of PA-RISC machines.

More flexibility = more freedom = more volatility. Both in
timing and in size of the drop.

But again, there's the other side of the coin. That more
volatile drop is also paired with a more volatile uptick.
x86 simply grows bigger too. Especially if you look at
volume rather than just revenue - and volume matters,
because it is volume that drives all those nice
networking effects.

(I forget the numbers and am too lazy to look them up, but
I suspect that the x86 dip in revenue was a lot bigger than
the dip in actual boxes shipped. Companies flush with cash
will buy beefier boxes "just in case". Companies that try
to cut down will tend to buy the boxes they need).

>Let's turn this around and put you on the spot instead
>of Itanium. What is your criterion for success fo IPF?
>20% of server sales? Greater than x86? More than 50%?
>Taking every last dollar spent on servers? Come on, step
>up and be a man. Draw a line in the sand and make your
>agenda clear to all.

Be a man? I'm not the one avoiding the issue.

But seriously, my success requirement for IPF used to be
that they actually get a thriving and lively development
community with good workstations that were competitive
from that standpoint. That never happened. Yes, there were
workstations, and some people even called them "good", but
those people were on some serious drugs and/or had lost
their hearing due to age or exposure to other IPF machines.

So now? What can happen to convince me there's life in
that old dog? Honestly, it needs to actually be technically
better than something else. Not "last in the heap". Faster,
cheaper, more capable.

Until that happens, IPF can't get any new market.

Sure, it can continue to get the HP-PA and Alpha/VMS sales,
and outside of that, it can even continue to grow with the
market (and by that I don't mean "in lock-step, quarter
by quarter" - that's not how anything works).

So there is no real reason to expect that it really takes
any new and relevant markets with x86 just being better in
every technical way, and with a much more lively community
around it.

If you don't want to bring up x86 (you generally don't),
then the only real competition is POWER. And in order for
IPF to be successful there, it needs to be better.
And I'm not seeing it.

Maybe HP can become such a wonderful service provider that
they blow IBM out of the water despite an inferior CPU
architecture - sure, that could happen. There are clearly
tons of totally non-technical reasons for IPF to take the
crown. However, they don't seem to actually ever happen.

You used to talk about "merchant cpu". It's gone. It's
now IBM vs HP. I didn't believe it back then, these days
nobody believes in the whole "merchant CPU" thing.
Sparc is more merchant than IPF is.

Yes, yes, there is Japan. I know about Japan. They like
odd architectures. They've always done them. They did them
for the PC space, for the supercomputer space, for the home
computer space, for every damn possible market space there
is for chissake. And it never ever made much of a
difference in the big picture. Japan is not sufficient.

So what does it take? The big price is long gone. IPF on
workstations is not even remotely realistic. The only part
remaining is the IPF-vs-Power thing, and as a technical
person, I think IPF needs to show that it is simply better
than POWER.

And for the last decade, IPF has failed miserably. It has
never been even temporarily better than POWER, and
most of the time it has been trailing by quite a big
way in performance and in features.

So if you ask for a success from my perspective, here's
one: IPF actually being clearly better than POWER. Even for
just a short while.

Of course, as explained, I think IPF has already lost all
the really interesting luster people tried to give
it, so that success is really more of a "most improved"
kind of accolade, but dang, it would be something glorious
for IPF to see even just once.

And the sad part is, Intel definitely can do it. I
obviously think Itanium is a horribly failed architecture
from a technical standpoint, but Intel should be able to
just use its clearly superior process to just overpower
IBM on that front.

Is Tukwila going to come close to POWER? Yes, I believe
that it will be able to basically get to parity, with a
"win some, lose some" situation.

Of course, that's when comparing to something IBM released
basically two years ago. Does IBM have something up its
sleeve to react if Tukwila looks good? I honestly don't
know. But I'd also be honestly surprised if they don't.

Ok, that was a much longer answer than I really expected.

Linus