IBM expands Opteron lineup

[bbbl67]

IBM falls in love with Opteron all over again | The Register
http://www.theregister.co.uk/2006/08/01/opteron_redux_ibm/

IBM announced 5 server models based on the Socket F/Rev F Opterons.
Despite all of the hoopla surrounding Woodcrest, it does look like the
Opteron momentum keeps growing. A recent article also said that AMD has
now captured 25% of the overall x86 server market already.

Yousuf Khan

 

[Steve]

Last I heard, Socket F, 1207 pins (actually LGA) was to be announced on
July 11 and shipped on August 1.

"With the Opteron Revision F launch, we will see the new processor
cores Santa Ana and Santa Rosa. Santa Ana is actually a dual-core AM2
component targeted specifically for workstations and 1U servers, while
Santa Rosa is an LGA-1207 dual-core Socket F processor. There are no
single-core Opteron Revision F processors. Quad-core Revision G
Opterons (Deerhound) are not anticipated for at least another year.
AMD's new 1207-pin socket, dubbed Socket F... will be AMD's first
land-grid-array (LGA) socket using the same method of contact pads for
the processor found on Intel LGA-775 and LGA-771 processors. The
retention clip is also a radical change for AMD.
http://www.dailytech.com/article.aspx?newsid=2531

And here is all I could find...

"Strangely the new IBM servers are based on existing Socket 940 AMD
Opteron processors instead of the Socket F AMD Opteron revision F
processors that were expected to launch today. AMD evangalists have
confirmed to DailyTech that inventory was supposed to be ready by now,
but problems with motherboard supply and processor logistics has
delayed the launch. http://www.dailytech.com/article.aspx?newsid=3603

I would bet Socket F will do all for the Opteron that Socket AM2 did
for the Athlon... (i.e. not much for performance, but a lot for
confusing buyers and making people wait for no good reason). Speaking
of confusion, Intel also has a chip code-named Santa Rosa. I love a
good fight, I love to see competition, but it looks to me like Intel is
going to open up new cans of whoop-ass with Woodcrest and Conroe alike.

 

[Ed]

I would bet Socket F will do all for the Opteron that Socket AM2 did
for the Athlon... (i.e. not much for performance, but a lot for
confusing buyers and making people wait for no good reason). Speaking
of confusion, Intel also has a chip code-named Santa Rosa. I love a
good fight, I love to see competition, but it looks to me like Intel is
going to open up new cans of whoop-ass with Woodcrest and Conroe alike.

AM2 is nothing like the new Opteron platform.

 

[Yousuf Khan]

And here is all I could find...

"Strangely the new IBM servers are based on existing Socket 940 AMD
Opteron processors instead of the Socket F AMD Opteron revision F
processors that were expected to launch today. AMD evangalists have
confirmed to DailyTech that inventory was supposed to be ready by now,
but problems with motherboard supply and processor logistics has
delayed the launch. http://www.dailytech.com/article.aspx?newsid=3603

I don't think Dailytech is even remotely right on this one (i.e. talking
out its ass). The announcements are all talking about speeds like 533Mhz
& 667Mhz, which are all DDR2 speeds, not DDR1. You need Socket F for DDR2.

I would bet Socket F will do all for the Opteron that Socket AM2 did
for the Athlon... (i.e. not much for performance, but a lot for
confusing buyers and making people wait for no good reason). Speaking
of confusion, Intel also has a chip code-named Santa Rosa. I love a
good fight, I love to see competition, but it looks to me like Intel is
going to open up new cans of whoop-ass with Woodcrest and Conroe alike.

Who's going to be confused by Socket AM2? The old one is for DDR1 and
new one is for DDR2. How much simpler can it get?

Yousuf Khan

 

[Steve]

Who's going to be confused by Socket AM2? The old one is for DDR1 and
new one is for DDR2. How much simpler can it get?

I meant that confusion would result from AMD promising something and
not delivering it. No confusion about what the product is, but
confusion about which to buy, and when. IBM's "new" Opteron servers are
a perfect example. They had to backtrack and ship with socket 940, but
in a month or two they're supposed to have the socket F versions out.

Maybe I am going out on a limb, but why will socket F do any more for
Opteron performance than AM2 did for Athlon performance (i.e. not
much)? It would be interesting to see Opteron Santa Rosa performance
compared to Woodcrest performance. Anyone seen any comparisons... or
want to place a bet?

 

[Tony Hill]

I meant that confusion would result from AMD promising something and
not delivering it. No confusion about what the product is, but
confusion about which to buy, and when. IBM's "new" Opteron servers are
a perfect example. They had to backtrack and ship with socket 940, but
in a month or two they're supposed to have the socket F versions out.

Maybe I am going out on a limb, but why will socket F do any more for
Opteron performance than AM2 did for Athlon performance (i.e. not
much)?

I don't anticipate too much of a performance boost for most
applications, though as a general rule servers are more
bandwidth-hungry than desktops. One notable exception to this is SPEC
CFP2000, which I suspect will see about a 10% boost in performance.
Applications that closely mirror CFP2000 will similarly benefit,
though those tend to be found more in the HPC market.

Another thing of note, Socket F does add a 4th Hypertransport channel
vs. the 3 in Socket 940. This won't make a lick of difference for 1
or 2P servers, but it will help a bit for 4P servers and it'll help a
lot for 8P servers. Combine that with the virtualization technology
and it's quite a worthwhile step forward for the 4 and 8P market.

It would be interesting to see Opteron Santa Rosa performance
compared to Woodcrest performance. Anyone seen any comparisons... or
want to place a bet?

For most applications I suspect that Woodcrest will hold the edge.
The ace up AMD's sleeve is that Woodcrest is limited to 1 or 2P
systems, while Opteron can scale to at least 8 sockets and
theoretically beyond that. Now, admittedly, with quad core (or
dual-dual core) chips being released towards the end of this year,
that might not be such a big deal.

 

[bbbl67]

I meant that confusion would result from AMD promising something and
not delivering it. No confusion about what the product is, but
confusion about which to buy, and when. IBM's "new" Opteron servers are
a perfect example. They had to backtrack and ship with socket 940, but
in a month or two they're supposed to have the socket F versions out.

As I said, forget that Dailytech article, they're not even saying it
anymore. Even the article you quoted has been changed; they've taken
out references to it being Socket 940. I don't know how the morons at
Dailytech came up with the idea that the new IBM servers are not going
to be Socket F, when quite clearly they were always talking about DDR2
memories.

Maybe I am going out on a limb, but why will socket F do any more for
Opteron performance than AM2 did for Athlon performance (i.e. not
much)? It would be interesting to see Opteron Santa Rosa performance
compared to Woodcrest performance. Anyone seen any comparisons... or
want to place a bet?

Why does it have to be about performance only? There are other reasons
for switching to DDR2, such as lower power consumption. Also DRAM
makers are going to only do upgrades on DDR2 from this point forward,
while DDR1 will be locked at its current density and speed levels
forever more. So really, performance advantages or not, there are
clearly good reasons for switching to DDR2 now. There comes a time when
one DRAM technology approaches a tipping point over its predecessor,
when it becomes more common than its predecessor, and that's what's now
happened with DDR2 over DDR1.

If AMD hadn't done Socket AM2/F/S1, then people would be complaining
about a lack of DDR2 support. So this is just standard housekeeping
duties for them.

 

[bbbl67]

Another thing of note, Socket F does add a 4th Hypertransport channel
vs. the 3 in Socket 940. This won't make a lick of difference for 1
or 2P servers, but it will help a bit for 4P servers and it'll help a
lot for 8P servers. Combine that with the virtualization technology
and it's quite a worthwhile step forward for the 4 and 8P market.


For most applications I suspect that Woodcrest will hold the edge.
The ace up AMD's sleeve is that Woodcrest is limited to 1 or 2P
systems, while Opteron can scale to at least 8 sockets and
theoretically beyond that. Now, admittedly, with quad core (or
dual-dual core) chips being released towards the end of this year,
that might not be such a big deal.

Virtualization is really meant for the server world. Though the
marketeers are making a brave effort at showing its usefullness in
desktops, it's really all just bullshit in that market. With the
combination of multi-cores/multi-processors and virtualization, you can
do some massive amounts of server consolidation. Not only can multiple
major multi-processor compute servers be replaced with one, you can
even replace tons of minor 1P and 2P servers. You can imagine one 8P
box with quad-cores has access to 64 cores altogether, all joined and
split up again into virtual servers.

Yousuf Khan

 

[George Macdonald]

As I said, forget that Dailytech article, they're not even saying it
anymore. Even the article you quoted has been changed; they've taken
out references to it being Socket 940. I don't know how the morons at
Dailytech came up with the idea that the new IBM servers are not going
to be Socket F, when quite clearly they were always talking about DDR2
memories.

The pic here http://www.marketwire.com/mw/release_html_b1?release_id=149564
is not very good but I'd guess socket F... and worth noting that we have
another F1 team with Opteron cluster/server. Still dunno what happened to
Sauber's rig.

 

[bbbl67]

The pic here http://www.marketwire.com/mw/release_html_b1?release_id=149564
is not very good but I'd guess socket F... and worth noting that we have
another F1 team with Opteron cluster/server. Still dunno what happened to
Sauber's rig.

Yeah, knew about Red Bull getting themselves some Opterons. The press
release didn't make it clear how many they were buying. But Red Bull
took over the Ferrari engine deal from Sauber after they were bought by
BMW, so it looks like every customer team that Ferrari supplies, end up
going with Opteron clusters. So the Ferrari sponsorship is working out
well for their Formula 1 supercluster marketshare numbers.

I'm sure that the Sauber cluster is also working away quietly in the
background there, it's not that easy to replace that many Opterons in
one sitting. I'm sure Intel would rather they got replaced by Core 2
Xeons rather than Pentium Xeons, just to showcase their best
technology. So considering the problems Intel is having in producing
Core 2's at the moment, Sauber will likely be still running Opterons
well into next year.

Yousuf Khan

 

[Tony Hill]

Yeah, knew about Red Bull getting themselves some Opterons. The press
release didn't make it clear how many they were buying. But Red Bull
took over the Ferrari engine deal from Sauber after they were bought by
BMW, so it looks like every customer team that Ferrari supplies, end up
going with Opteron clusters. So the Ferrari sponsorship is working out
well for their Formula 1 supercluster marketshare numbers.

Yup, though AMD actually managed to get some other teams as well.
Renault and Williams were also using Opteron clusters, and I think
McLaren as well (after their deal with Intel fell through). Toyota
and Sauber-BMW are the only ones I know in the Intel camp on the
F1-supercluster battle. I'm not really sure were Honda F1 lie, nor
the little teams (Toro Rosso, Midland and Super Aguri).

I'm sure that the Sauber cluster is also working away quietly in the
background there, it's not that easy to replace that many Opterons in
one sitting. I'm sure Intel would rather they got replaced by Core 2
Xeons rather than Pentium Xeons, just to showcase their best
technology. So considering the problems Intel is having in producing
Core 2's at the moment, Sauber will likely be still running Opterons
well into next year.

The deal was announced some time ago, so I would say that Intel has
already moved in a whole cluster of Xeons and/or Itaniums. They may
well get upgraded to Core 2 chips over time, but I suspect Intel had
some sort of cluster in there pretty early. After all, the whole
Sauber/AMD deal was a bit of good marketing for AMD chips.

As for poor ol' Albert, it's tough to say what happened to him, a
super-cluster in the prime of his career cut short by marketing
decisions totally unrelated to his performance! Maybe he got sold to
one of the above-mentioned little teams? I sure they could make good
use of a 500-chip Opteron cluster!

ps. speaking of Sauber-BMW, congrats to Robert Kubica on his debut,
impressive stuff for a young-un tossed in to the deep end of F1
racing... even if his success does mean the end of the line for old
Jacques. Maybe it'll give Jacques more time to work on his musical
career?

 

[YKhan]

The pic here http://www.marketwire.com/mw/release_html_b1?release_id=149564
is not very good but I'd guess socket F... and worth noting that we have
another F1 team with Opteron cluster/server. Still dunno what happened to
Sauber's rig.

Oh BTW, AMD is going to be the title sponsor of Nascar's Watkins Glen
race.

NASCAR.COM - Watkins Glen Cup race to be sponsored by AMD - Aug 4, 2006
http://www.nascar.com/2006/news/headlines/official/08/04/glen_name/index.html

 

[YKhan]

Yup, though AMD actually managed to get some other teams as well.
Renault and Williams were also using Opteron clusters, and I think
McLaren as well (after their deal with Intel fell through). Toyota
and Sauber-BMW are the only ones I know in the Intel camp on the
F1-supercluster battle. I'm not really sure were Honda F1 lie, nor
the little teams (Toro Rosso, Midland and Super Aguri).

I think I heard about Renault using Opterons, didn't know about
Williams. However, Mclaren had just purchased an Itanium supercluster
from SGI just a couple of years back, which resulted in them losing
their Sun Micro sponsorship.

As for Toro Rosso, they're probably piggybacking on Red Bull, while
Super Aguri is piggybacking on BAR. Don't know who Midland can
piggyback on, they have Toyota engines so maybe they also get some
computing services from Toyota?

The deal was announced some time ago, so I would say that Intel has
already moved in a whole cluster of Xeons and/or Itaniums. They may
well get upgraded to Core 2 chips over time, but I suspect Intel had
some sort of cluster in there pretty early. After all, the whole
Sauber/AMD deal was a bit of good marketing for AMD chips.

I can see them having Xeon chips in there, as that would be compatible
with Opteron software. But I can't see them forcing Itaniums in there.
Unless they started with Itanium software right from the beginning, the
real choice was only Xeon.

As for poor ol' Albert, it's tough to say what happened to him, a
super-cluster in the prime of his career cut short by marketing
decisions totally unrelated to his performance! Maybe he got sold to
one of the above-mentioned little teams? I sure they could make good
use of a 500-chip Opteron cluster!

My assumption was that his brain was replaced little by little. For
example, the original Opteron cluster was done by LinuxNetworx I
believe. So I would think they would just remove the Opteron
rackservers and slowly replace them with Xeon rackservers made by the
same company? Of course with a little upgrade in air conditioning put
in place too.

ps. speaking of Sauber-BMW, congrats to Robert Kubica on his debut,
impressive stuff for a young-un tossed in to the deep end of F1
racing... even if his success does mean the end of the line for old
Jacques. Maybe it'll give Jacques more time to work on his musical
career?

Jacques had his second chance.

Yousuf Khan

 

[nobody]

Virtualization is really meant for the server world. Though the
marketeers are making a brave effort at showing its usefullness in
desktops, it's really all just bullshit in that market. With the
combination of multi-cores/multi-processors and virtualization, you can
do some massive amounts of server consolidation. Not only can multiple
major multi-processor compute servers be replaced with one, you can
even replace tons of minor 1P and 2P servers. You can imagine one 8P
box with quad-cores has access to 64 cores altogether, all joined and
split up again into virtual servers.

Yousuf Khan

I for one don't see this as a benefit. Clusters, Web farms, and such
were invented not only for performance. They mostly provide for
redundancy, to exclude a single point of failure. It might be
possible to put all of enterprise server needs inside one box, but
then a single blown fuse might bring down everything (OK, exagerating,
but the idea is clear). I'd rather have a mission critical
application spread over a bunch of physically separate pizza boxes, so
if one goes down users would not even notice the difference.

NNN

 

[George Macdonald]

Yeah, knew about Red Bull getting themselves some Opterons. The press
release didn't make it clear how many they were buying. But Red Bull
took over the Ferrari engine deal from Sauber after they were bought by
BMW, so it looks like every customer team that Ferrari supplies, end up
going with Opteron clusters. So the Ferrari sponsorship is working out
well for their Formula 1 supercluster marketshare numbers.

Hmm, the Opteron connection may say more about Sauber's "copy" of the
Ferrari chassis than anything to do with engines.;-) I believe the Ferrari
engines get delivered "sealed" anyway so not much rqt for Sauber or Red
Bull to "work" with them. As for Red Bull, looks like they are trying to
get Ferrari to allow the engines to go to STR next year and Flavio will get
them Renault lumps to go with the Webber deal. Ah, the rebirth of umm,
Mecachrome... JIT to foil Cosworth.

I'm sure that the Sauber cluster is also working away quietly in the
background there, it's not that easy to replace that many Opterons in
one sitting. I'm sure Intel would rather they got replaced by Core 2
Xeons rather than Pentium Xeons, just to showcase their best
technology. So considering the problems Intel is having in producing
Core 2's at the moment, Sauber will likely be still running Opterons
well into next year.

I guess Intel has to relent on the Itanium thingy there. The info I get
from inside F1 doesn't really tell how much custom software there might be
so I dunno how easy it would be to switch to Itanic; certainly the
mechnical/surfacing CAD stuff is mostly off the shelf -- e.g. CATIA -- but
there could be some aero/CFD which is "in-house" and then there's the stuff
which is so secret nobody knows. Add in things like PTC's
"decertification" of Itanium and I'd think Itanium is dead in that
market... among others. Even if you don't use PTC's stuff, it's always
nice to have the option without a hardware revamp.

 

[bbbl67]

I for one don't see this as a benefit. Clusters, Web farms, and such
were invented not only for performance. They mostly provide for
redundancy, to exclude a single point of failure. It might be
possible to put all of enterprise server needs inside one box, but
then a single blown fuse might bring down everything (OK, exagerating,
but the idea is clear). I'd rather have a mission critical
application spread over a bunch of physically separate pizza boxes, so
if one goes down users would not even notice the difference.

Well, I'm sure the replacement of several servers with just 1 box is
just an oversimplified example. The real scenario might be something
like replacing 100 boxes with 10 boxes. This would save on electricity
and airconditioning costs, beyond just straight hardware consolidation.
I don't think if you have just 10 boxes that you're really pining to
replace them with just one box.

As for failover and clustering, I would think that a large box would be
better able to handle the extra loads of a failover than a small box.
If you had an Oracle super-server with 8 processors, then you'd pretty
much have to fail it over to another 8 processor server sitting as
backup (at best you might be able to get away with a 4 processor server
as its backup, but not a cheap 2 processor box). When you have a
64-core hyper-server, it can be backed up by another 64-core
hyper-server. Programs like Symantec's Veritas Cluster Server and
VMWare's own virtualization program allow these sorts of failovers
between different machines.

 

[bbbl67]

Hmm, the Opteron connection may say more about Sauber's "copy" of the
Ferrari chassis than anything to do with engines.;-) I believe the Ferrari
engines get delivered "sealed" anyway so not much rqt for Sauber or Red
Bull to "work" with them. As for Red Bull, looks like they are trying to
get Ferrari to allow the engines to go to STR next year and Flavio will get
them Renault lumps to go with the Webber deal. Ah, the rebirth of umm,
Mecachrome... JIT to foil Cosworth.

Yeah, Sauber only talked about using it for computational fluid
dynamics work. It seems like they came up with one innovation this
year, using their cluster, those long vertical aerodynamic "towers" put
right in the path of the driver's vision. Those were promptly banned
the next race.

Next year they're going to lock down the engine specs for the whole
year, and implement a rev limit on the engines.

I guess Intel has to relent on the Itanium thingy there. The info I get
from inside F1 doesn't really tell how much custom software there might be
so I dunno how easy it would be to switch to Itanic; certainly the
mechnical/surfacing CAD stuff is mostly off the shelf -- e.g. CATIA -- but
there could be some aero/CFD which is "in-house" and then there's the stuff
which is so secret nobody knows. Add in things like PTC's
"decertification" of Itanium and I'd think Itanium is dead in that
market... among others. Even if you don't use PTC's stuff, it's always
nice to have the option without a hardware revamp.

I don't know, I'm stilly pretty certain that McLaren is using SGI
Itanium clusters, they just bought them a couple of years back, can't
see them having them replaced already. At least not without marketing
and sponsorship pressure.

Yousuf Khan

 

[nobody]

If you had an Oracle super-server with 8 processors, then you'd pretty
much have to fail it over to another 8 processor server sitting as
backup (at best you might be able to get away with a 4 processor server
as its backup, but not a cheap 2 processor box).

While I am more familliar with MS SQL Server than Oracle, I would
prefer a cluster of 3 or 4 el cheapo 2 socket boxes to one 8 socket
monster. This way, even if one of the boxes fails, the other boxes
will take its workload, and the users will in worst case notice slower
response time, if anything at all. If the big box goes down for
whatever reason, all users go down with it, or you have to keep
another big box as a hot spare. Since the 8 socket server costs waaay
more than 4x2 socket, the cluster setup would be also more
economically sound solution. Admittedly administering a cluster is
more of pain in you know where comparing to a single box, so it all
depends on particular application you are running.

NNN

 

[George Macdonald]

Yeah, Sauber only talked about using it for computational fluid
dynamics work. It seems like they came up with one innovation this
year, using their cluster, those long vertical aerodynamic "towers" put
right in the path of the driver's vision. Those were promptly banned
the next race.

Next year they're going to lock down the engine specs for the whole
year, and implement a rev limit on the engines.

What's apparently been tentatively agreed is an engine lock-down from the
2006 Chinese GP, with a 19k limit for 2008... and then the funny part: fuel
economy and energy reuse. For cars which get ~1.5mpg this is hilarious -
Mad Max is showing signs of taking after his father after all.

I don't know, I'm stilly pretty certain that McLaren is using SGI
Itanium clusters, they just bought them a couple of years back, can't
see them having them replaced already. At least not without marketing
and sponsorship pressure.

Could be but, with the size of McLaren's ops in their swish new umm,
premises, I'm sure there are other things they could find to do with
Itaniums... given available software, which is the real problem. Certainly
it was Mclaren who was the intended Intel partner to be announced at Monza
last year... before AMD+FIA went and spoiled the party... sending Intel off
in an "incandescent" hissy fit.

 

[Tony Hill]

I think I heard about Renault using Opterons, didn't know about
Williams.

I think the Williams deal actually came about from their (now defunct)
deal with HP. Interesting that HP would supply them with an Opteron
cluster rather than Itanium ones.

However, Mclaren had just purchased an Itanium supercluster
from SGI just a couple of years back, which resulted in them losing
their Sun Micro sponsorship.

Ahh, you're quite right, just last summer actually.

As for Toro Rosso, they're probably piggybacking on Red Bull, while
Super Aguri is piggybacking on BAR.

A certain degree of piggybacking may occur, but the teams are both
doing their own development. Particularly Super Aguri is supposed to
be doing their own cars, albeit with some hand-me-downs from the Honda
factory team (formerly BAR Honda until Honda bought out the remaining
stake in BAR). They've got their own wind tunnel and that means that
they will almost certainly need their own supercluster to process the
data.

Don't know who Midland can
piggyback on, they have Toyota engines so maybe they also get some
computing services from Toyota?

Doubt it. Midland has already lost their Toyota engine deal to
Williams for the next 3 years. Besides that deal seemed to be pretty
exclusive to just engine supplies. At this point it would seem that
Midland is going to be the only really independent team on the grid
next year. The smart money seems to be on them getting Cosworth
engines, so they'll be pretty much on their own when it comes to
anything beyond the drive-train (though I suppose the same will be
true for Williams as well).

I can see them having Xeon chips in there, as that would be compatible
with Opteron software. But I can't see them forcing Itaniums in there.
Unless they started with Itanium software right from the beginning, the
real choice was only Xeon.

It all depends on the software that they're using. If they recompile
the important bits than it shouldn't take too much to hand it off to
Itaniums. On the other hand, if they're using lots of precompiled
software or hand-tuned stuff then Xeons would probably be a better
bet. Either way, Intel probably tossed enough money their way to make
it worth their while.

My assumption was that his brain was replaced little by little. For
example, the original Opteron cluster was done by LinuxNetworx I
believe. So I would think they would just remove the Opteron
rackservers and slowly replace them with Xeon rackservers made by the
same company? Of course with a little upgrade in air conditioning put
in place too.

Yeah, I wonder how the people working with the system thought about
the whole deal? It's not like this was a decision based on technical
superiority or anything, it was all about much-needed sponsorship
money.

Jacques had his second chance.

Only his second chance? I think Jacques was on at least his 3rd or
4th chance by this stage of the game! Ohh well, maybe he can head
over to Champ Car and join Forsythe alongside Paul Tracy!