Sun has to switch to Itanium?

[Felger Carbon]

Increasing cache size
helps, but it's a case of fairly rapidly decreasing returns (how much
does your hit rate increase when going from 9MB of cache to 12MB? I
would guess it's MUCH lower than 1%).

Tony, the "hit rate" doesn't matter. What matters is the miss rate.
Honest.

If the miss rate drops from 2% to 1%, then the average memory latency
nearly drops in half ("nearly" because the cache has _some_ latency,
after all). That corresponds to a hit rate of 98% vs 99%, "only" a 1%
increase.

The question about going from 9MB to 13MB cache is (as you have
frequently pointed out whether the increase is cost-effective at
the system level. It's all about bang per buck.

 

[Yousuf Khan]

It is perfectly legal for a contract to require an exclusive product,
and I have reason to believe that some contracts with Intel do just
that. The point is that Intel could get away with it when they had
the whip hand, which they never did with HP and IBM to the extent
that they did with SGI and Dell.

Sure, Intel is perfectly entitled to say that we don't want you to advertise
our competitors products with the ad money that we give you. However, Intel
is anything but entitled to say we don't want you to sell our competitor's
products if you want our ad money. Just one of those fine lines that are
differentiated in anti-trust laws. Intel would never be stupid enough to say
the latter, it would be hauled into court right away by the government.
However, that's not to say that Intel can't put in some suitably vague
clauses in their contracts which would allow it to pull supplies from
customers who are "non-compliant" with their advertising dollars policies.

Yousuf Khan

 

[Nick Maclaren]

|> |> > It is perfectly legal for a contract to require an exclusive product,
|> > and I have reason to believe that some contracts with Intel do just
|> > that. The point is that Intel could get away with it when they had
|> > the whip hand, which they never did with HP and IBM to the extent
|> > that they did with SGI and Dell.
|>
|> Sure, Intel is perfectly entitled to say that we don't want you to advertise
|> our competitors products with the ad money that we give you. However, Intel
|> is anything but entitled to say we don't want you to sell our competitor's
|> products if you want our ad money. Just one of those fine lines that are
|> differentiated in anti-trust laws. Intel would never be stupid enough to say
|> the latter, it would be hauled into court right away by the government.

Ah. Fascinating. I suspect that you have been reading more into
those laws than is there. There are constraints, but they are much
less absolute than you seem to think.


Regards,
Nick Maclaren.

 

[David Svensson]

The "cheaper" system of course, with cheaper being measured as some
sort of TCO (ie a rather fuzzy number).


Enough has been made public about Niagara to figure this one out.
It's a chip with 8 simple SPARC cores on it, each capable of running 4
threads. Even if the individual Itanium chips are MUCH faster,
they'll only be running one or two threads until "Tukwila" (as you
hint at bellow, the code-name "Tanglewood" has been changed to
"Tukwila"). That means that the Itanium will have to be MUCH faster
than the individual Niagara core to perform the same in a heavily
multithreaded environment, at least 4 times faster (assuming dual-core
Itaniums), and quite possibly much more than that if the SMT
performance of Niagara works out like Sun says it will. Considering
that Intel's plans for Itanium over the next few years consist of
little other than increasing the cache size and going to dual-core
designs, I don't think they'll come close. Increasing cache size
helps, but it's a case of fairly rapidly decreasing returns (how much
does your hit rate increase when going from 9MB of cache to 12MB? I
would guess it's MUCH lower than 1%). Dual-core is a good start, but
they'll be rather late to the game (at least a year away still).

Niagara will have very slow 1-issue wide cores most likely a bit
slower than UltraSparc II. Niagara will not have SMT, it is vertical
multithreading. Niagara will not be SMP cabable (, well it is SMP in
itself but it will be alone.)

Sun will only place them in their blade servers, and they will compete
with 1-2 and perhaps 4-way x86 servers. I really don't see that
Niagara is in the same arena as Itanium. You are a bit wrong too about
next years Itanium (Montecito), it will have 2-way multithreading on
each core, thus appearing as 4 CPUs to the OS.

It looks like Sun is stuck with UltraSparc IV in their general purpose
servers and high-end servers until Rock might appear in 2007.

 

[Wow]

Not offering any opinions myself, just relaying an article:

http://www.osnews.com/story.php?news_id=6757

Yousuf Khan

Itanium is a ugly architecture,
While I like the predication because you can describe
both paths of the branches and not take a
branch mis-prediction hit, but....

First, The software-pipelining feature is extremely
convoluted piece to understand. It uses the
predication registers and the stack engine
together to work, but what happens when you are
already using the predication registers??
Plus this makes the RSE engine more complicated.
Alot of the stuff is done behind some scene.
Or with alot of hand waving. Which is bad
when you are programming in Assembly language.
Believe me, I've done programming in IA64 Assembly
language for work and alot of the examples
given by Intel I could not make heads or tails.

Second, because of the instructions bundles/groups
design, you have to pad the instruction slots
with NOOPs if you can't find exactly three instructions
per group. Hence, the instruction cache is effectively very small.

Third, There is NO integer multiple nor integer divide!!!
So you have to move the values over to the
Floating point unit. Well, what happens if the
FPU is busy doing loops?? Do you want to
interrupt whatever the FPU is doing just for
your integer multiplication that integer ALUs can't do??

If you look carefully at the IA64, It's really a
descendant of HP's PA RISC architecture.
IA64's no integer multiply comes from HP-PA,
so is the Shift-Left-and-Add instructions.
Even the instruction encoding within the groups are
the same as HP-PA RISC.

IA64 = HP PA RISC instructions/encoding +
VLIW/EPIC concepts/ideas +
x86 emulation

 

[John D Groenveld]

What's Dell's Opteron strategy?

John
(e-mail address removed)

 

[John D Groenveld]

Why would Sun settle with Microsoft?

To concentrate on IBM and Intel.

John
(e-mail address removed)

 

[Eugene Nalimov]

"Yousuf Khan" <[email protected]> wrote in message

Itanium is a ugly architecture,
While I like the predication because you can describe
both paths of the branches and not take a
branch mis-prediction hit, but....

First, The software-pipelining feature is extremely
convoluted piece to understand. It uses the
predication registers and the stack engine
together to work, but what happens when you are
already using the predication registers??
Plus this makes the RSE engine more complicated.
Alot of the stuff is done behind some scene.
Or with alot of hand waving. Which is bad
when you are programming in Assembly language.
Believe me, I've done programming in IA64 Assembly
language for work and alot of the examples
given by Intel I could not make heads or tails.

Then let it to the compilers.

Second, because of the instructions bundles/groups
design, you have to pad the instruction slots
with NOOPs if you can't find exactly three instructions
per group. Hence, the instruction cache is effectively very small.

Had you seen optimized Alpha binaries? They contain lot
of NOPs, too.

Third, There is NO integer multiple nor integer divide!!!
So you have to move the values over to the
Floating point unit. Well, what happens if the
FPU is busy doing loops?? Do you want to
interrupt whatever the FPU is doing just for
your integer multiplication that integer ALUs can't do??

You can do 32x32 integer multiplication on Itanium without
FP unit. Visual C does this.

Thanks,
Eugene

 

[David Magda]

Enough has been made public about Niagara to figure this one out.
It's a chip with 8 simple SPARC cores on it, each capable of
running 4 threads. Even if the individual Itanium chips are MUCH

Of interest may be Ace's Hardware recently posted article describing
"Niagra":

http://www.aceshardware.com/read.jsp?id=65000292

 

[Douglas Siebert]

I have difficulty in believing that you are naive enough to ask that
question seriously. But, just in case you are:

1) Money. SGI is very short of it. Investing in a new chip
takes money to develop and test the hardware and software.

2) Contracts. Some of those with Intel may be exclusive, at
least for now. SGI signed the deal with Intel when Intel had the
whip hand.

3) Commercial politics. Intel bears serious grudge, and SGI
can't afford to be on the receiving end from Intel.

I think that SGI signing on with IA64 early means they probably don't
have any contractual reason they couldn't go with Opteron. Back in
the late 90s when they signed on, Intel didn't have any reason to think
that AMD could ever be a threat to the 64 bit market Itanium was
targeting, so it would have never been mentioned in the contract.

1) is valid, however since SGI used much the same hardware to work
with MIPS and IA64, making it work with Opteron shouldn't be all that
difficult, especially since its architecture would lend itself
perfectly to SGI's cellular architecture.

3) is the big one, SGI would have to be ready to give up any sort of
preferential treatment regarding pricing or allocation they may have
been able to get in exchange for agreeing to phase out MIPS. But unless
IA64 suddenly takes off, I don't think SGI has to worry about shortages
meaning they can't get any parts as sometimes happens with Pentiums and
uncooperative OEMs. HP probably gets first dibs on everything anyway.

 

[=?ISO-8859-1?Q?Jan_Vorbr=FCggen?=]

3) is the big one, SGI would have to be ready to give up any sort of

preferential treatment regarding pricing or allocation they may have
been able to get in exchange for agreeing to phase out MIPS.

Is a slight (<50%) change in processor price going to have any impact on
product pricing for the stuff SGI is selling? I strongly doubt that. And
allocation for IA64 processors - is that really happening?

Jan

 

[Nick Maclaren]

|>
|> > 3) is the big one, SGI would have to be ready to give up any sort of
|> > preferential treatment regarding pricing or allocation they may have
|> > been able to get in exchange for agreeing to phase out MIPS.
|>
|> Is a slight (<50%) change in processor price going to have any impact on
|> product pricing for the stuff SGI is selling? I strongly doubt that. And
|> allocation for IA64 processors - is that really happening?

The issues have very little to do with pricing, and a great deal
to do with influence on the details and access to very detailed
(internal) information. This statement is not specific to SGI and
Intel, but applies to almost all such deals - such as IBM and Apple.
It also applies to committments to delivery and dates, like the
need to avoid fiascos like the "it's off - no, it's on" Merced.

I have no idea of what is in any of the SGI/Intel contracts, or
I couldn't post. Nor do I have any first- or second-hand data on
on how Intel behaves behind closed doors, but there is a fair amount
of indirect information around.


Regards,
Nick Maclaren.

 

[Andrew Gabriel]

|>
|> > 3) is the big one, SGI would have to be ready to give up any sort of
|> > preferential treatment regarding pricing or allocation they may have
|> > been able to get in exchange for agreeing to phase out MIPS.
|>
|> Is a slight (<50%) change in processor price going to have any impact on
|> product pricing for the stuff SGI is selling? I strongly doubt that. And
|> allocation for IA64 processors - is that really happening?

The issues have very little to do with pricing, and a great deal
to do with influence on the details and access to very detailed
(internal) information. This statement is not specific to SGI and
Intel, but applies to almost all such deals - such as IBM and Apple.
It also applies to committments to delivery and dates, like the
need to avoid fiascos like the "it's off - no, it's on" Merced.

Whilst this might be the case in the x86 space, I doubt it's the
case in the IA64 space. There the boot is on the other foot, with
most of the IHVs for the product having abandoned it, leaving just
HP with any volume at all, and nowhere near enough for the chip to
remain viable. I suspect Intel would welcome any new IHV with
open arms, but with x86-IA32e announced, it's most unlikely any IHV
will commit further development funding to IA64 systems from now
on (except possibly HP who bet the company on it), as it's clear
that isn't going top be Intel's volume chip.

 

[Bernd Paysan]

What's Dell's Opteron strategy?

Tell Intel that they'll build an Opteron based system until Intel's special
Dell discount is big enough.

 

[Yousuf Khan]

Whilst this might be the case in the x86 space, I doubt it's the
case in the IA64 space. There the boot is on the other foot, with
most of the IHVs for the product having abandoned it, leaving just
HP with any volume at all, and nowhere near enough for the chip to
remain viable. I suspect Intel would welcome any new IHV with
open arms, but with x86-IA32e announced, it's most unlikely any IHV
will commit further development funding to IA64 systems from now
on (except possibly HP who bet the company on it), as it's clear
that isn't going top be Intel's volume chip.

Except HP is also the biggest hedger, being now the biggest committer to
Opteron (the only major OEM with a 4-way system so far), as well as Itanium.
But I guess HP has got enough money to go both ways, unlike SGI.

Yousuf Khan

 

[George Macdonald]

Except HP is also the biggest hedger, being now the biggest committer to
Opteron (the only major OEM with a 4-way system so far), as well as Itanium.
But I guess HP has got enough money to go both ways, unlike SGI.

I don't know if it's still like that but I recall years ago that HP wasn't
so much a company with corporate synergy as a group of independent
divisions which did pretty much what they liked... without talking to each
other. I distinctly recall being at a conference/exhibit once where HP had
three separate booths all pushing some variant of "the PC" and who seemed
reluctant to acknowledge each others' existence.

I find it very significant that they now have AMD systems in their
"Corporate PC" repertoire... something that Intel seems to have prevented
happening at various vendors until recently. I get the feeling the tide is
turning - add in Intel's Xeon only 64-bit and the door is open more than a
crack IMO.[I love mixing metaphors]

Rgds, George Macdonald

"Just because they're paranoid doesn't mean you're not psychotic" - Who, me??

 

[Wolfgang S. Rupprecht]

I don't know if it's still like that but I recall years ago that HP wasn't
so much a company with corporate synergy as a group of independent
divisions which did pretty much what they liked... without talking to each
other. I distinctly recall being at a conference/exhibit once where HP had
three separate booths all pushing some variant of "the PC" and who seemed
reluctant to acknowledge each others' existence.

Any large corporation doesn't really speak with one voice. Internally
there will always be divisions competing with each other for corporate
resources and the permission to sell their wares externally. It must
be interesting to be a fly on the conference room wall of Sun or HP.
The lobbying by the old guard to keep out the faster and cheaper
upstart must be intense.

-wolfgang

 

[keith bierman]

Not offering any opinions myself, just relaying an article:

http://www.osnews.com/story.php?news_id=6757

Yousuf Khan

I seem to have stumbled upon this late. Since it seems that everyone has
neglected to google for old stories about the Solaris port to IA64 here
are a few. I am not in a position to make any comment (other than noting
that they are all old stories).

http://techupdate.zdnet.com/techupdate/stories/main/0,14179,2871375,00.html
http://www.sun.com/smi/Press/sunflash/9910/sunflash.991025.4.html
http://news.com.com/2100-1001_3-243489.html

The point, such as it is, that the article's claim "But, the company is
still averse to porting Solaris over to Itanium" is and was clearly
false. Porting occured. What happened after that is either available in
the press or is company confidential :>

 

[George Macdonald]

Any large corporation doesn't really speak with one voice. Internally
there will always be divisions competing with each other for corporate
resources and the permission to sell their wares externally. It must
be interesting to be a fly on the conference room wall of Sun or HP.
The lobbying by the old guard to keep out the faster and cheaper
upstart must be intense.

Some truth in that but when three different divisions were producing near
identical hardware and pushing it separately on the market... I find it
umm, unusual.

Rgds, George Macdonald

"Just because they're paranoid doesn't mean you're not psychotic" - Who, me??