Xeon Woodcrest Preys On Opteron

  • Thread starter Thread starter Yousuf Khan
  • Start date Start date
Yousuf said:
Today is the day that Intel officially introduces Woodcrest to the public.

http://tomshardware.co.uk/2006/06/26/xeon_woodcrest_preys_on_opteron_uk/index.html


strikes me as an ok 11 page article.
looks like Intel is on the blitz

i'll be interested to see how many
others agree (on the article) given
that many here (me somewhat included,
but i don't go out to Tom's much)
have come to dislike/mistrust Tom's

thank you for the ref. :)

to my view, Tom's use of the word
"preys" in their subject title sez
something (marketing/sales) about Tom's

also page 6 (of 11) also has this:
"Woodcrest and the Xeon 5100 family is the
core of the new generation that is aimed
at surpassing the AMD Opteron family by
better performance per Watt. Although the
platform offers an number of interesting
features, we are still sceptical concerning
the efficiency claims. These only mention
the processor, all components should be
takeninto account. The powerful northbridge
and numerous FB-DIM modules
could possibly change the outcome."

^^^^^^^^^^^^^^^^^^^^^^^

page 11 (the conclusion) also had
pertinent comments

bill
 
Yousuf,
Since when you treat Tom as a credible source?

I don't, I just post dem there interesting links. You can read it if you
want.
Besides such a bold
headline as "Xeon Woodcrest Preys On Opteron" requires a side by side
comparison of 2 similarly configured systems. As hard as I looked, I
could not find any AMD results in the whole article. So IMHO the jury
is still out.

Well, they were trying very hard to fair to AMD in this article, which
impressed me -- it's not often that Tom's even bothers with those kind
of ethics. They did fail towards the end, spewing out some FUD about how
without FB-DIMMs Opterons wouldn't be able to be expanded upto as much
memory as Woodcrests. FB-DIMMs are only a necessity on Woodcrest because
of their single shared memory controller. You can put upto 8 sticks of
FB-DIMMs on one controller. But with just regular old registered DIMMs,
you're probably limited to 4 sticks, but since each Opteron has its own
memory controller, each controller can take 4 on its own, giving them
the same number of sticks (8) in the end. Plus the registered DIMMs are
cheaper, more power-efficient, and faster than the FB-DIMMs.

Yousuf Khan
 
Ryan said:
I read the article at Tom's but actually couldn't find any comparisons
in terms of performance compared to the Opteron.

They did mention that the article was just an introduction to Woodcrest,
and though they did do testing with the chips, they were going to wait
till another article to show those results.
I would have found the whole thing more satisfying if there were some
hard facts and rigorous tests - but I guess this is Tom's - so not
really their house style.

Well, I'm sure there's going to be lots of fun picking at Tom's testing
methodology when that article is released. Fortunately, it looks like
Tom is no longer part of Tom's Guide (haven't seen the guy in ages), so
the level of professionalism has seemingly gone up. Or at least the
weaseliness level has become much more subtle. :-)

Yousuf Khan
 
willbill said:
strikes me as an ok 11 page article.
looks like Intel is on the blitz

Well, you gotta figure that Intel had 3 years to ponder things in the
wilderness. After all of this time, if they had come up with just
another processor, then the only thing that would've prevented AMD from
taking ever more market share was their own manufacturing capacity. It's
to Intel's credit that it doesn't look like they have short-circuited
the development process to rush this thing into production, either.
Three years is definitely not a rush to market. We saw the problems they
got into when trying to respond to competitive pressures when they
rushed out the Pentium 4; P4 wasn't completed yet, and we saw them
adding features incrementally that they skipped early on like
Hyperthreading; also it looks like P4 was always more comfortable at
130nm but they had to rush it in at 180nm to meet the challenge of the
K7 Athlons. And I don't think AMD has been sitting idle here either,
since it too has had 3 years to develop its processors. It just looks
like Conroe has beaten K8L to market by a few months; I don't think in
the long run that this will matter, it takes a long time to reverse
market share trends in the server market. Opteron only started gaining
share about a year after introduction, despite the fact that it was
plain to see how superior it was to what Intel offered at the time.

But it certainly looks like AMD is setting to work big time on
addressing its manufacturing capacity problems. Not only is it upgrading
its two Dresden fabs, it's also going to build a new one in New York state.
i'll be interested to see how many
others agree (on the article) given
that many here (me somewhat included,
but i don't go out to Tom's much)
have come to dislike/mistrust Tom's

thank you for the ref. :)

to my view, Tom's use of the word
"preys" in their subject title sez
something (marketing/sales) about Tom's

Well, I'll say that the headline certainly got my attention. Even if the
article itself wasn't quite what I expected.

Yousuf Khan
 
Besides such a bold
Well, they were trying very hard to fair to AMD in this article, which
impressed me -- it's not often that Tom's even bothers with those kind
of ethics. They did fail towards the end, spewing out some FUD about how
without FB-DIMMs Opterons wouldn't be able to be expanded upto as much
memory as Woodcrests. FB-DIMMs are only a necessity on Woodcrest because
of their single shared memory controller. You can put upto 8 sticks of
FB-DIMMs on one controller.

No. Each channel of FBD can have up to 8 DIMMs. That would be a total
of 32 DIMMs for a Woodcrest system. Currently I have seen
implementations with 16 DIMMs:
http://supermicro.com/products/motherboard/Xeon1333/5000P/X7DB8+.cfm
But with just regular old registered DIMMs,
you're probably limited to 4 sticks, but since each Opteron has its own
memory controller, each controller can take 4 on its own, giving them
the same number of sticks (8) in the end.

See my note above...
Plus the registered DIMMs are
cheaper, more power-efficient, and faster than the FB-DIMMs.

Less RAS and harder to route. Personally, I think Intel made a mistake
and should have a special ULV chipset for Woodcrest that uses RDIMMs.
I don't think FBD is the right technology for blade servers or other
ultra-dense/low power systems at that moment.

Perhaps in a year though, but for a 1U system, it's fine.

DK
 
David said:
No. Each channel of FBD can have up to 8 DIMMs. That would be a total
of 32 DIMMs for a Woodcrest system. Currently I have seen
implementations with 16 DIMMs:
http://supermicro.com/products/motherboard/Xeon1333/5000P/X7DB8+.cfm

I guess that comes down to how conservative the OEM and/or board
manufacturers are with these things, both on the Opteron and Xeon side:

thunderk8hm_spec
http://www.tyan.com/PRODUCTS/html/thunderk8hm_spec.html

The above holds 16 sticks of registered DDR1 memory. It limits the
speed to 333Mhz when using the highest configuration though.

Yousuf Khan
 
But it certainly looks like AMD is setting to work big time on
addressing its manufacturing capacity problems. Not only is it upgrading
its two Dresden fabs, it's also going to build a new one in New York state.

It'll take a while for most of that to kick in, according to AMD's
website.

Fab 36 has started production of 300mm 90nm, and will reportedly be
shipping 65nm in the second half of this year, converting fully by mid
2007, but that will be their main manufacturing site as Fab 30 is
ramped down at the same time.

Fab 30's transformation into Fab 38 won't ramp 300mm 65nm until the
end of 2007 if all goes well, reaching capacity at the end of 2008,
but converting a fab is trickier than building a greenfield plant, and
I'd bet their 300mm output is not going to be very high during the
conversion. In the meantime, the Fab 36 output will partly go to
offset the loss of Fab 30's 200mm capacity.

As for the NY fab:
"The agreement enables construction on the 1.2 million square foot
plant to begin between July 2007 and July 31, 2009 and be fully
operational sometime between December 31, 2012 and December 31, 2014."

They'll certainly have plenty of capacity once all that's online, no
doubt. They're some expensive projects, though, and if the price wars
hit their cash flow, it could get tight for them in the meantime.

If Intel takes back market share with Core 2, particularly in the high
end, I could foresee some of these projects getting delayed.

Interesting times, to be sure.

max
 
No. Each channel of FBD can have up to 8 DIMMs. That would be a total
I guess that comes down to how conservative the OEM and/or board
manufacturers are with these things, both on the Opteron and Xeon side:

That is part of it. Thought, you have to realize that Super Micro
still hasn't hit the peak. You can get 8 DIMMs/channel, so that's 32
max.
thunderk8hm_spec
http://www.tyan.com/PRODUCTS/html/thunderk8hm_spec.html

The above holds 16 sticks of registered DDR1 memory. It limits the
speed to 333Mhz when using the highest configuration though.

That's right, you cannot use a maxed out Opteron memory system without
dropping the bandwidth. FBD doesn't suffer from that problem.

Now the interesting question is how will memory capacity change when
going to DDR2...it sure won't increase.

DK
 
Ryan said:
I read the article at Tom's but actually couldn't find any comparisons
in terms of performance compared to the Opteron.

I would have found the whole thing more satisfying if there were some
hard facts and rigorous tests - but I guess this is Tom's - so not
really their house style.

Ryan

There are a ton of tests out (2cpu.com, anandtech.com, etc) and all
show about the same; woodcrests and conroes pretty much blowing the
doors off of opterons.

I think the fellow who made the comment about it taking a long time to
gain share in the server market is wrong; its an entirely different
thing to convince Intel houses to switch to AMD than it is to convince
them to switch back to Intel. Lots of people in the UP or DP market use
opterons simply because they are the fastest available processor. With
that not being the case there's an awfully large segment that has no
reason to use opterons any longer.

Its a lot easier to make a case to use Intel than it is to use AMD.

TM

TM
 
It'll take a while for most of that to kick in, according to AMD's
website.

Fab 36 has started production of 300mm 90nm, and will reportedly be
shipping 65nm in the second half of this year, converting fully by mid
2007, but that will be their main manufacturing site as Fab 30 is
ramped down at the same time.

Fab 30's transformation into Fab 38 won't ramp 300mm 65nm until the
end of 2007 if all goes well, reaching capacity at the end of 2008,
but converting a fab is trickier than building a greenfield plant, and
I'd bet their 300mm output is not going to be very high during the
conversion. In the meantime, the Fab 36 output will partly go to
offset the loss of Fab 30's 200mm capacity.

As for the NY fab:
"The agreement enables construction on the 1.2 million square foot
plant to begin between July 2007 and July 31, 2009 and be fully
operational sometime between December 31, 2012 and December 31, 2014."

They'll certainly have plenty of capacity once all that's online, no
doubt. They're some expensive projects, though, and if the price wars
hit their cash flow, it could get tight for them in the meantime.

If Intel takes back market share with Core 2, particularly in the high
end, I could foresee some of these projects getting delayed.

Interesting times, to be sure.

max
Sanders may be out of the corner office, but his spirit is alive at
AMD - "real men have fabs"
NNN
 
max said:
Fab 30's transformation into Fab 38 won't ramp 300mm 65nm until the
end of 2007 if all goes well, reaching capacity at the end of 2008,
but converting a fab is trickier than building a greenfield plant, and
I'd bet their 300mm output is not going to be very high during the
conversion. In the meantime, the Fab 36 output will partly go to
offset the loss of Fab 30's 200mm capacity.

Yeah, but I'm sure the days of just abandonning a fab are over. How are
you going to explain to the environmentalists and governments why a
perfectly good building is just sitting there empty with no other uses
for it other than demolition? I think for that reason, AMD has decided
to build two buildings in Germany, so that it can probably partially
shut one down, upgrade it, while the other still works at full
capacity. It'll likely do the same thing in New York (although NY
taxpayers probably don't know it yet). :-)

In the meantime, they also have Chartered to take on any extra capacity
requirements. Also there is a fab being built in India by SemIndia,
with AMD technology in it. This is probably a pilot to see if the
Indians can get enough willpower together to build out the proper
infrastructure to run a fab reliably. AMD won't have to spend any money
on this one, but they can invest in it once they see it's running
reliably and they'd have some more capacity there too.

They'll certainly have plenty of capacity once all that's online, no
doubt. They're some expensive projects, though, and if the price wars
hit their cash flow, it could get tight for them in the meantime.

If Intel takes back market share with Core 2, particularly in the high
end, I could foresee some of these projects getting delayed.

Intel might take back some marketshare in the desktop segment, and
possibly in the laptop segment, but it's going to have to face the same
issue to retake marketshare in the server market that AMD had in
initially capturing it -- server market is slow to switch. AMD had a 3
full years of superiority to start to change the momentum around in
that market. Intel will at best only have six months of superiority
before AMD catches back up again (if only because of 65nm, even without
any architectural improvements). Not enough time to make much impact on
AMD. Of course, that's not to say that AMD won't have to lower its
prices somewhat in the server market, but still there's enough money to
be made here.

The laptop market isn't very concerned with performance. It's more
concerned with price and value. At least the consumer laptop market is,
which is the most dynamic part of the laptop market. So Merom won't
have much impact on Intel's laptop marketshare. In fact, it will be
Sonoma, Intel's two generation-old Pentium M processor that will have
more impact on its marketshare here. There was a story today that Intel
is sitting on 4 million units of unsold Sonomas that it has to clear
from inventory; it's going to sell them for the same price as
Celeron-M's!

http://www.digitimes.com/mobos/a20060629PR210.html

So that just leaves just Conroe on the desktop. That's probably the one
that will make the biggest impact to AMD marketshare in the meantime.
The desktop market is the most fickle and dynamic. The prices here are
already so low that people don't worry so much if they pay a little
extra here and there.

Yousuf Khan
 
They'll certainly have plenty of capacity once all that's online, no
Intel might take back some marketshare in the desktop segment, and
possibly in the laptop segment, but it's going to have to face the same
issue to retake marketshare in the server market that AMD had in
initially capturing it -- server market is slow to switch.

Actually, I think the real issue why it was slow for AMD was that they
had to build up their brand reputation and get OEM design wins. I
think for the DP market, Intel can easily get the design wins...
AMD had a 3
full years of superiority to start to change the momentum around in
that market. Intel will at best only have six months of superiority
before AMD catches back up again (if only because of 65nm, even without
any architectural improvements).

Do you really think that AMD's 65nm compaction will match Woodcrest?
I'm a little skeptical here. I'd also point out that AMD's presence in
4S market won't be challenged till Q3 when Tulsa comes out, and it has
yet to be determined whether that will be enough.
Not enough time to make much impact on
AMD. Of course, that's not to say that AMD won't have to lower its
prices somewhat in the server market, but still there's enough money to
be made here.
The laptop market isn't very concerned with performance. It's more
concerned with price and value. At least the consumer laptop market is,
which is the most dynamic part of the laptop market. So Merom won't
have much impact on Intel's laptop marketshare. In fact, it will be
Sonoma, Intel's two generation-old Pentium M processor that will have
more impact on its marketshare here. There was a story today that Intel
is sitting on 4 million units of unsold Sonomas that it has to clear
from inventory; it's going to sell them for the same price as
Celeron-M's!

http://www.digitimes.com/mobos/a20060629PR210.html

So that just leaves just Conroe on the desktop. That's probably the one
that will make the biggest impact to AMD marketshare in the meantime.
The desktop market is the most fickle and dynamic. The prices here are
already so low that people don't worry so much if they pay a little
extra here and there.

The other issue which you are missing is that AMD has no marketshare in
mobile. They have about 15% of the consumer and small business (< 10
employees) market, and it goes rapidly downhill from there. AMD has no
presence in the 1000 employee+ market, and that won't change in the
near future.

DK
 
I think the fellow who made the comment about it taking a long time to
gain share in the server market is wrong; its an entirely different
thing to convince Intel houses to switch to AMD than it is to convince
them to switch back to Intel. Lots of people in the UP or DP market use
opterons simply because they are the fastest available processor. With
that not being the case there's an awfully large segment that has no
reason to use opterons any longer.

Its a lot easier to make a case to use Intel than it is to use AMD.


The server market isn't simply based on performance alone. A lot of it
depends on infrastructure requirements. Once these people made the
switch to Opteron, they had to put in place the tools and knowledge to
operate them. They don't make the switch that easily. It took AMD a
long time to convince these people to make the switch, mainly because
they were hoping to see Intel come back with an answer AMD which didn't
come for years. Now at best, Intel's advantage is there for six months
until AMD catches right back up when it gets to 65nm. AMD doesn't even
have to make architectural improvements, just the switch to 65nm will
result in improved power consumption and good enough performance
increases. Once the architectural improvements come too, it will result
in equal or better performance from AMD once again. There's also the
upgrade market where there will be a pre-existing base of Opteron
servers who will want a faster Opteron, since a Xeon won't fit in
there. And many businesses expand out with projects with homogenous
servers, so they don't want to switch to a different platform if they
don't have to as their project expands.

In short these were all of the issues that prevented Opteron from
getting much traction at all in its first year, but which made it
explode in the second and third years. Intel will not get a 3 year lead
to work with here.

Yousuf Khan
 
David said:
Actually, I think the real issue why it was slow for AMD was that they
had to build up their brand reputation and get OEM design wins. I
think for the DP market, Intel can easily get the design wins...

Well yeah, that really is the real issue isn't it? Why did it take so
long for AMD to rack up OEM design wins? But I don't think, as you do,
it had anything to do with simply brand recognition. The design wins
seemed to start piling up all of a sudden within weeks of AMD launching
its lawsuit against Intel, as the spotlight of scrutiny shone on Intel.
I won't bore you with a rehash of all of the examples, you should know
them by now.
Do you really think that AMD's 65nm compaction will match Woodcrest?

Sure, it will catch up to the point where it's going to be close enough
so it doesn't matter. Even without architectural improvements, just a
simple die shrink will be enough to reduce power consumption and they
could also add larger caches if they wanted to at that point. But it
looks like AMD is set to introduce various architectural improvements
and that it's just waiting for the 65nm process to become available, so
it won't necessarily be just a simple die shrink.

Already there's indications that AMD's memory controller and
Hypertransport are getting ready to reassert their superiority within a
few months over the FSB. There are performance DDR2 modules coming out
at 1200Mhz/CAS4, 50% higher than DDR2-800, which only an AMD processor
can handle. The dual-channel DDR2-800 has already saturated the 1066Mhz
Conroe bus. If Conroe wants to use the DDR2-1200, it would need a
chipset revision that can handle 1600Mhz FSB.
The other issue which you are missing is that AMD has no marketshare in
mobile. They have about 15% of the consumer and small business (< 10
employees) market, and it goes rapidly downhill from there. AMD has no
presence in the 1000 employee+ market, and that won't change in the
near future.

15% is their overall laptop marketshare, not just their consumer and
small business share. But they seem to be holding closer to 30%
marketshare in the retail notebook market. But anyways, that miniscule
15% AMD overall marketshare seems to have been enough to cause Intel to
miss their mark by 4 million Sonoma chips.

As for the large business (1000+ employee) laptop market, sure Intel
probably holds over 90% of the market there, if not 100%. But frankly,
it's the least interesting part of the laptop market, even for Intel.
It's the oldest, most mature part of the laptop market, with the
slowest growth rates. This market probably provides Intel with some
steady upgrade business too, but this business is not growing. Large
businesses tend to layoff people more than they hire them; and when
they do hire somebody, they tend to assign them with a laptop from
somebody they previously had layed-off.

That's why Intel and everybody else are interested in getting more
consumers to buy laptops. But the consumer won't buy the expensive
laptops that businesses tend to buy. Frankly, a business doesn't care
what the price of the laptop is, because when a business buys a $2000
laptop, they're not really looking at the $2000 final price tag,
they're probably only looking at the $100/month lease-rate. But the
consumer looks at the $2000 price tag, and sees $2000. Consumers don't
tend to lease their laptops, unless you consider using a credit card a
kind of a lease? So value for the money has become the top priority in
the consumer laptop space.
 
YKhan said:
Well yeah, that really is the real issue isn't it? Why did it take so
long for AMD to rack up OEM design wins? But I don't think, as you do,
it had anything to do with simply brand recognition.

Well you are entitled to your opinion.
The design wins
seemed to start piling up all of a sudden within weeks of AMD launching
its lawsuit against Intel, as the spotlight of scrutiny shone on Intel.
I won't bore you with a rehash of all of the examples, you should know
them by now.


Sure, it will catch up to the point where it's going to be close enough
so it doesn't matter.

Not a chance, look at the gap in SPECint scores. There is no way that
AMD can catch up with a simple die shrink, and you're deceiving
yourself if you think that is so.

Ditto for TPC-C or SPECjbb2005.
Even without architectural improvements, just a
simple die shrink will be enough to reduce power consumption and they
could also add larger caches if they wanted to at that point.

See you're missing the point. Right now, AMD is not performance
competitive with Intel in servers. They won't be able to reduce power
consumption on 65nm, they will be too busy ramping up clockspeed to try
and get back to performance parity.
But it
looks like AMD is set to introduce various architectural improvements
and that it's just waiting for the 65nm process to become available, so
it won't necessarily be just a simple die shrink.

Yes it will. AMD's first design will be a compaction, and after that
they will do K8L.
Already there's indications that AMD's memory controller and
Hypertransport are getting ready to reassert their superiority within a
few months over the FSB. There are performance DDR2 modules coming out
at 1200Mhz/CAS4, 50% higher than DDR2-800, which only an AMD processor
can handle.

Hello? Nobody in the server world uses out of spec modules. DDR2-800
is the top of the line for servers. Do you really think AMD will be
able to out perform Intel with a simple upgrade to HT and the memory
controller? If so you are ignoring reality. AMD's modifications might
get them 10%...
The dual-channel DDR2-800 has already saturated the 1066Mhz
Conroe bus.

Who cares about Conroe? AMD makes their money in the server market,
and that is where Intel is going to hit the hardest.
If Conroe wants to use the DDR2-1200, it would need a
chipset revision that can handle 1600Mhz FSB.

Who cares about out of spec overclocked modules?
15% is their overall laptop marketshare, not just their consumer and
small business share.

Prove it. I've seen presentations from AMD that claim 15% for consumer
markets WW, and it goes down from there.
But they seem to be holding closer to 30%
marketshare in the retail notebook market. But anyways, that miniscule
15% AMD overall marketshare seems to have been enough to cause Intel to
miss their mark by 4 million Sonoma chips.

You don't have a clue. AMD has no presence in the mobile space and
won't in the near future.
As for the large business (1000+ employee) laptop market, sure Intel
probably holds over 90% of the market there, if not 100%.

Try 99%, according to AMD.
But frankly,
it's the least interesting part of the laptop market, even for Intel.
It's the oldest, most mature part of the laptop market, with the
slowest growth rates. This market probably provides Intel with some
steady upgrade business too, but this business is not growing. Large
businesses tend to layoff people more than they hire them; and when
they do hire somebody, they tend to assign them with a laptop from
somebody they previously had layed-off.

That's why Intel and everybody else are interested in getting more
consumers to buy laptops. But the consumer won't buy the expensive
laptops that businesses tend to buy. Frankly, a business doesn't care
what the price of the laptop is, because when a business buys a $2000
laptop, they're not really looking at the $2000 final price tag,
they're probably only looking at the $100/month lease-rate. But the
consumer looks at the $2000 price tag, and sees $2000. Consumers don't
tend to lease their laptops, unless you consider using a credit card a
kind of a lease? So value for the money has become the top priority in
the consumer laptop space.

No it's not. But you are welcome to believe what ever you want.

DK
 
There are a ton of tests out (2cpu.com, anandtech.com, etc) and all
show about the same; woodcrests and conroes pretty much blowing the
doors off of opterons.

I think the fellow who made the comment about it taking a long time to
gain share in the server market is wrong; its an entirely different
thing to convince Intel houses to switch to AMD than it is to convince
them to switch back to Intel. Lots of people in the UP or DP market use
opterons simply because they are the fastest available processor. With
that not being the case there's an awfully large segment that has no
reason to use opterons any longer.

Its a lot easier to make a case to use Intel than it is to use AMD.

TM

TM

Correct me if i'm wrong, but I thought that Woodcrest and Conroe were
not yet generally available. I'll take these numbers with a pinch of
salt until I see retail processors on retail boards, or at least oem
machines. I agree that the performance is looking promising.

If Intel's engineering matches their marketeering on we'll have some
jolly nice kit coming along. I wouldn't however recommend a
non-available system for general use - too slow:)

Ryan
 
Correct me if i'm wrong, but I thought that Woodcrest and Conroe were
not yet generally available.

You're wrong. Woodcrest is quite available. Conroe won't be released
for a little while.
I'll take these numbers with a pinch of
salt until I see retail processors on retail boards, or at least oem
machines. I agree that the performance is looking promising.

Go to an OEM website.

DK
 
Back
Top