IBM Power 7 will be using Opteron motherboards

[YKhan]

Sun is also planning something similar for its processors. And maybe
even Fujitsu with its Sparc64 chips. Wouldn't that be something, all of
the remaining RISC architectures are converging onto the same
motherboards?

IBM's Power7 chip going into Opteron motherboards | The Register
"AMD has secured yet another major partner win thanks to Opteron.
Starting with the Power7 processor, IBM will give up on making its own
Unix/RISC box motherboards. Instead, it will plug the Power chips
directly into slightly modified Opteron boards in an effort to save
money."
http://www.theregister.co.uk/2006/09/22/ibm_power7_opteron/

 

[nobody]

Sun is also planning something similar for its processors. And maybe
even Fujitsu with its Sparc64 chips. Wouldn't that be something, all of
the remaining RISC architectures are converging onto the same
motherboards?

IBM's Power7 chip going into Opteron motherboards | The Register
"AMD has secured yet another major partner win thanks to Opteron.
Starting with the Power7 processor, IBM will give up on making its own
Unix/RISC box motherboards. Instead, it will plug the Power chips
directly into slightly modified Opteron boards in an effort to save
money."
http://www.theregister.co.uk/2006/09/22/ibm_power7_opteron/

and also http://www.theregister.co.uk/2006/09/21/amd_open_opteron/

So the only one major player with proprietary socket/mobo/chipset/etc.
will be Intel. The rest - IBM, Sun, Fujitsu et al. will be using
industry standard (read - AMD spec'ed) infrastructure. You couldn't
make this stuff up ;-)))))))))))))

NNN

 

[YKhan]

and also http://www.theregister.co.uk/2006/09/21/amd_open_opteron/

Yes, more than likely yesterday's PR announcement is probably a prelude
to what IBM and others were really planning. Opteron co-processors? I
don't think so.

So the only one major player with proprietary socket/mobo/chipset/etc.
will be Intel. The rest - IBM, Sun, Fujitsu et al. will be using
industry standard (read - AMD spec'ed) infrastructure. You couldn't
make this stuff up ;-)))))))))))))

Well, yeah, you can, this is a Register story afterall, so we can't
count our chickens quite yet. At best this is an early preview of
something that may happen.

Anyways, all you need now is Itanium to come over to the Opteron
platform, and it'll all be all complete, the Borg will have been
assimililated too.

Yousuf Khan

 

[Del Cecchi]

Sun is also planning something similar for its processors. And maybe
even Fujitsu with its Sparc64 chips. Wouldn't that be something, all of
the remaining RISC architectures are converging onto the same
motherboards?

IBM's Power7 chip going into Opteron motherboards | The Register
"AMD has secured yet another major partner win thanks to Opteron.
Starting with the Power7 processor, IBM will give up on making its own
Unix/RISC box motherboards. Instead, it will plug the Power chips
directly into slightly modified Opteron boards in an effort to save
money."
http://www.theregister.co.uk/2006/09/22/ibm_power7_opteron/

You think laying out and fabbing a board is a major expense? And what
the blazes is a "slightly modified" board? doesn't modifying it throw
away all the cost advantages allegedly acruing to the high volume?

Who writes this stuff?

 

[David Kanter]

You think laying out and fabbing a board is a major expense? And what
the blazes is a "slightly modified" board?

Why it's a motherboard that supports those 7-10K pins required for
high-end MPUs from IBM...

Surely you don't expect anyone to bung in a POWER5/6 class MPU into a
socket with under 5K pins?

doesn't modifying it throw
away all the cost advantages allegedly acruing to the high volume?

Pretty much. Notice how Alpha accrued no benefits from the cheap and
crappy K7 boards. Even the MP boards weren't up to the standards of
the Tsunami chipset...

DK

 

[Peter Matthias]

You think laying out and fabbing a board is a major expense?  And what
the blazes is a "slightly modified" board?  doesn't modifying it throw
away all the cost advantages allegedly acruing to the high volume?

I thinkk qualifying a board is a major expense. With a slightly modified
board I expect some changes in the power supply of the CPU.

Peter

 

[Yousuf Khan]

You think laying out and fabbing a board is a major expense? And what
the blazes is a "slightly modified" board? doesn't modifying it throw
away all the cost advantages allegedly acruing to the high volume?

Who writes this stuff?

Probably not the board itself, but at least it'll get to save money by
using off-the-shelf chipsets. Or IBM would be able to also develop
chipsets for its own Power architecture, which it can then sell to the
wider Opteron-market too.

Not sure what they meant by slightly modified boards. Maybe it means as
simple as slightly modified with a Power firmware rather than a PC BIOS?
Or maybe some voltage and capacitor specification changes?

Yousuf Khan

 

[YKhan]

Pretty much. Notice how Alpha accrued no benefits from the cheap and
crappy K7 boards. Even the MP boards weren't up to the standards of
the Tsunami chipset...

I think that by the time K7 came out, Alpha was already being
consolidated down. The massive push for increased Alpha volume was
already turning into "forget it" refrain.

Yousuf Khan

 

[Tony Hill]

You think laying out and fabbing a board is a major expense?

I would think that the cost savings would be in not having to design
their own chip-to-chip connection. Hypertransport 3.0 offers an awful
lot of performance and features at a very low cost.

And what
the blazes is a "slightly modified" board? doesn't modifying it throw
away all the cost advantages allegedly acruing to the high volume?

Not if it's just the BIOS that you're modifying. That might be a
possibility for the low-end servers and workstations. The high-end
stuff would still most certainly have to be a custom design.

Who writes this stuff?

The Register, so take it with a BIG grain of salt. I'll wait to hear
some rather more concrete evidence before I put too much stock in it.

Still, it's not outside the realm of possibility, I've certainly heard
a lot of crazier rumors that turned out to be true.

 

[nobody]

Why it's a motherboard that supports those 7-10K pins required for
high-end MPUs from IBM...

Surely you don't expect anyone to bung in a POWER5/6 class MPU into a
socket with under 5K pins?


Pretty much. Notice how Alpha accrued no benefits from the cheap and
crappy K7 boards. Even the MP boards weren't up to the standards of
the Tsunami chipset...

DK

I'd rather put the question this way: why AMD didn't try to take
advantage of Tsunami, at least for server class MP boards? Not sure
if it was possible without a major rework, but it could pave the way
for AMD into server space way before K8.

NNN

 

[Yousuf Khan]

Not if it's just the BIOS that you're modifying. That might be a
possibility for the low-end servers and workstations. The high-end
stuff would still most certainly have to be a custom design.

Plus it might open up IBM's custom-designed chipsets for Power to a
wider audience too. They could sell their Power chipsets to HP or Sun
too, where they could be used with Opteron or UltraSparc.

Opteron should be approaching 40% of the server market pretty soon, this
will be an even more significant revenue generator for chipset designers
than it is already. Broadcom and Nvidia are already competing for this
market as we speak.

Still, it's not outside the realm of possibility, I've certainly heard
a lot of crazier rumors that turned out to be true.

And it's only the second such similar rumour. Sun is already rumored to
be designing future generation Niagara processors with Opteron
connectors. This way they can share the same system boards and cases for
both types of processors.

Yousuf Khan

 

[Yousuf Khan]

I'd rather put the question this way: why AMD didn't try to take
advantage of Tsunami, at least for server class MP boards? Not sure
if it was possible without a major rework, but it could pave the way
for AMD into server space way before K8.

NNN

Tsunami was cancelled wasn't it? It never came out.

Anyways back then with K7 Athlon MP, AMD could only do 2-way SMP. That
just happens to be the meat of the Xeon market. There was no way Intel
was going to let AMD play in that field, unless AMD could bully its way
into that field by itself. K7 wasn't enough of a weapon to bully its way
into that field, but Opteron was.

Yousuf Khan

 

[YKhan]

The Register, so take it with a BIG grain of salt. I'll wait to hear
some rather more concrete evidence before I put too much stock in it.

Well, I don't know if this counts as corroboration, because they are
quoting the Register article, but at least it indicates that these guys
somehow believe the Register.

Techworld.com - IBM plans death of Power mobos
http://www.techworld.com/opsys/news/index.cfm?newsid=6957&pagtype=all

Yousuf Khan

 

[Tony Hill]

Plus it might open up IBM's custom-designed chipsets for Power to a
wider audience too. They could sell their Power chipsets to HP or Sun
too, where they could be used with Opteron or UltraSparc.

It's possible, but I don't see it as being likely. You don't see IBM
pushing their X3 chipsets for sale to others and, if anything, the
Opteron/Power chipsets are likely to be more high-end designs than the
X3. In this market the real money to be made is in service and
support, not in the hardware. Selling the chipsets to Sun or HP would
be giving away any advantage that IBM might have in this market for
rather small revenue.

Opteron should be approaching 40% of the server market pretty soon, this
will be an even more significant revenue generator for chipset designers
than it is already. Broadcom and Nvidia are already competing for this
market as we speak.

Neither Broadcom or nVidia produce their own servers, just chipsets.
More to the point, neither of them generate revenue from the service
and support of servers the way IBM does.

Besides, why would IBM want to get into an already competitive
low-margin business?

And it's only the second such similar rumour. Sun is already rumored to
be designing future generation Niagara processors with Opteron
connectors. This way they can share the same system boards and cases for
both types of processors.

That would make very good sense. The market for Niagara servers is
very similar cost-wise to that of Opteron servers, while the low
latency/high bandwidth of hypertransport should mesh well with the
chips main tasks. Niagara already has an integrated memory
controller, so that part shouldn't be a problem. Add in network chip
that hangs directly off a Hypertransport link and you've got a pretty
impressive solution. To top it off, Sun has their own Opteron servers
so they can share a lot of R&D and maybe even some hardware. This one
might be just a rumor, but it is not only quite believable, but it
would seem rather dumb for Sun to take any other choice.

The IBM rumor, on the other hand, is much less clear-cut.

 

[YKhan]

It's possible, but I don't see it as being likely. You don't see IBM
pushing their X3 chipsets for sale to others and, if anything, the
Opteron/Power chipsets are likely to be more high-end designs than the
X3. In this market the real money to be made is in service and
support, not in the hardware. Selling the chipsets to Sun or HP would
be giving away any advantage that IBM might have in this market for
rather small revenue.

True, so even if IBM keeps it all to itself, it can still spread out
the development costs between both the Power & Opteron servers it
sells.

Secondly, the reason IBM has any kind of a "high-end" X3 chipset for
Xeons is due to the Xeons' inherent limitations in the multi-processor
realm. That's not the case with Opteron, so much. So there won't be as
much need for a high-end chipset with these systems anyways. Even the
Newisys Horus chipset seems to have fallen by the wayside, because it's
looking like with AMD increasing the number of HT links in Opteron upto
4 from 3 might be enough to make Horus redundant. So in a similar vein,
IBM chipset development might become redundant.

Neither Broadcom or nVidia produce their own servers, just chipsets.
More to the point, neither of them generate revenue from the service
and support of servers the way IBM does.

Besides, why would IBM want to get into an already competitive
low-margin business?

Never said they do sell services, but they make quite enough money from
chipsets to make a decent enough profit.

That would make very good sense. The market for Niagara servers is
very similar cost-wise to that of Opteron servers, while the low
latency/high bandwidth of hypertransport should mesh well with the
chips main tasks. Niagara already has an integrated memory
controller, so that part shouldn't be a problem. Add in network chip
that hangs directly off a Hypertransport link and you've got a pretty
impressive solution. To top it off, Sun has their own Opteron servers
so they can share a lot of R&D and maybe even some hardware. This one
might be just a rumor, but it is not only quite believable, but it
would seem rather dumb for Sun to take any other choice.

Well, yeah, that was my point when I said, "This way they can share the
same system boards and cases for both types of processors."

The IBM rumor, on the other hand, is much less clear-cut.

Why would that be? If it makes sense for Sun, why not IBM? IBM has its
own set of Opteron servers and its own RISC servers, just like Sun. The
same level of sharing between RISC and Opteron infrastructures could be
driving IBM too.

Yousuf Khan

 

[George Macdonald]

True, so even if IBM keeps it all to itself, it can still spread out
the development costs between both the Power & Opteron servers it
sells.

Secondly, the reason IBM has any kind of a "high-end" X3 chipset for
Xeons is due to the Xeons' inherent limitations in the multi-processor
realm.

Yeah, it was umm, interesting to see X3 was followed shortly by Intel's own
DIB chipset. I haven't seen too much more on how they compare
performance-wise?

That's not the case with Opteron, so much. So there won't be as
much need for a high-end chipset with these systems anyways. Even the
Newisys Horus chipset seems to have fallen by the wayside, because it's
looking like with AMD increasing the number of HT links in Opteron upto
4 from 3 might be enough to make Horus redundant. So in a similar vein,
IBM chipset development might become redundant.

I notice that many of the newer Opteron chipset mbrds include a HTX slot -
not sure how much of a backplane you need to do a large cluster with that
but the infrastructure/enclosure engineering gets important there too.

Never said they do sell services, but they make quite enough money from
chipsets to make a decent enough profit.

I have my doubts that Broadcom/Serverworks is going to survive this one...
as a chipset designer/mfr. I get the impression they're hanging on.

Well, yeah, that was my point when I said, "This way they can share the
same system boards and cases for both types of processors."


Why would that be? If it makes sense for Sun, why not IBM? IBM has its
own set of Opteron servers and its own RISC servers, just like Sun. The
same level of sharing between RISC and Opteron infrastructures could be
driving IBM too.

I assume you're talking about >2S chipset/mbrd systems here? The 2S is
pretty well covered by the merchant chipsets and even an excellent design
would not bring a tangible performance benefit. I see opportunities in the
4S and up though - the return of the glasshouse is taking a grip and I've
pondered the possible viability of an x86 "mainframe" here; a hybrid
Power/x86 mainframe might be just the ticket. At some point, I'd think
there has to be an advantage in having an integrated "big box" vs. a bunch
of 1U or 2U modules in a generic chassis.

 

[David Kanter]

Secondly, the reason IBM has any kind of a "high-end" X3 chipset for

Yeah, it was umm, interesting to see X3 was followed shortly by Intel's own
DIB chipset. I haven't seen too much more on how they compare
performance-wise?

They are hard to compare. X3 is for > 4 Sockets, blackford for 2
sockets.

It has a much larger snoop filter (it spans around 200MB of cache),
while Blackford spans about 20 or 30MB of cache.

The X3 uses a slightly slower bus as well, but Intel's SF
implementation has problems with OLTP-like workloads. It's really hard
to say which is higher performance because its an apples to mangos kind
of thing.

David

 

[Tony Hill]

True, so even if IBM keeps it all to itself, it can still spread out
the development costs between both the Power & Opteron servers it
sells.

Yup, that would be the main advantage for IBM.

Secondly, the reason IBM has any kind of a "high-end" X3 chipset for
Xeons is due to the Xeons' inherent limitations in the multi-processor
realm. That's not the case with Opteron, so much. So there won't be as
much need for a high-end chipset with these systems anyways. Even the
Newisys Horus chipset seems to have fallen by the wayside, because it's
looking like with AMD increasing the number of HT links in Opteron upto
4 from 3 might be enough to make Horus redundant. So in a similar vein,
IBM chipset development might become redundant.

Except at the very high-end, yes. The Opteron still probably won't
scale all that great beyond 4 sockets with a glue-less design,
certainly going beyond 8 sockets is out of the question. IBM could be
well served by designing their Power7 chips to use Hypertransport and
designing high-end chipsets to go along with this for their top-end
Power servers. If needed they could also incorporate Opteron chips
into these high-end servers, though more likely the transfer would
flow the other way, ie Power chips with 3rd party chipsets for the
lower range of Power servers. I'm not sure that there's enough of a
market for 4+ socket x86 servers, especially given that quad-core
processors will be widely available in the timeframe we're talking
about. However there does still seem to be a market for lower end
(1-4 socket) Power servers.

Never said they do sell services, but they make quite enough money from
chipsets to make a decent enough profit.

nVidia does, though I think that's more due to their desktop chipset
sales. Their server chips are really just lightly modified desktop
chipsets. If they had to develop these chips all on their own than
the volume probably wouldn't be enough to make much money, even with
the high profit margins of server parts.

Broadcom, on the other hand, probably isn't doing too hot with their
Opteron chipsets. I'd be quite surprised if they were breaking even,
let alone making a profit. They don't break down their profits by
division, but the old Serverworks group didn't even get a mention in
their latest financial report, unlike the guys that made the Bluetooth
chip to go into the controller for the Nintendo Wii. Seems to me like
this is a pretty fringe group who's time may be rather limited.

Why would that be? If it makes sense for Sun, why not IBM? IBM has its
own set of Opteron servers and its own RISC servers, just like Sun. The
same level of sharing between RISC and Opteron infrastructures could be
driving IBM too.

You don't see Sun rushing to move their UltraSparc chips to
hypertransport, do you? It might make some sense, just as it makes
some sense for IBM to move Power there, but it's definitely not the
same thing as with Niagara chips. Different markets with different
needs. Where Niagara is pretty much a win-win situation for Sun,
UltraSparc and IBM's Power are much less so.

 

[bbbl67]

Except at the very high-end, yes. The Opteron still probably won't
scale all that great beyond 4 sockets with a glue-less design,
certainly going beyond 8 sockets is out of the question. IBM could be
well served by designing their Power7 chips to use Hypertransport and
designing high-end chipsets to go along with this for their top-end
Power servers. If needed they could also incorporate Opteron chips
into these high-end servers, though more likely the transfer would
flow the other way, ie Power chips with 3rd party chipsets for the
lower range of Power servers. I'm not sure that there's enough of a
market for 4+ socket x86 servers, especially given that quad-core
processors will be widely available in the timeframe we're talking
about. However there does still seem to be a market for lower end
(1-4 socket) Power servers.

Since Socket F is supposed to be designed for future Opterons with 4 HT
links rather than 3 links, it's supposed to be good for 16 socket
glueless systems. With quad-core that would mean 64 processor cores.
Systems I see with that many processors tend to be never used as a
single monolithic system, but they are almost always split up into
virtualized partitions. Even if you're doing HTPC work, often the work
can be split up over network cabling rather than system backplanes.

nVidia does, though I think that's more due to their desktop chipset
sales. Their server chips are really just lightly modified desktop
chipsets. If they had to develop these chips all on their own than
the volume probably wouldn't be enough to make much money, even with
the high profit margins of server parts.

Well, Nvidia discovered the main advantage of Hypertransport early: let
the CPU do all of the heavy-lifting. So it wouldn't surprise me that
Nvidia is supplying the same chipsets for both the desktop and server
markets. What else would you need extra between these two markets?

Broadcom, on the other hand, probably isn't doing too hot with their
Opteron chipsets. I'd be quite surprised if they were breaking even,
let alone making a profit. They don't break down their profits by
division, but the old Serverworks group didn't even get a mention in
their latest financial report, unlike the guys that made the Bluetooth
chip to go into the controller for the Nintendo Wii. Seems to me like
this is a pretty fringe group who's time may be rather limited.

They were supposedly cooperating with Sun on their Galaxy servers, but
when the Galaxies were finally introduced they came with Nvidia
chipsets.

You don't see Sun rushing to move their UltraSparc chips to
hypertransport, do you? It might make some sense, just as it makes
some sense for IBM to move Power there, but it's definitely not the
same thing as with Niagara chips. Different markets with different
needs. Where Niagara is pretty much a win-win situation for Sun,
UltraSparc and IBM's Power are much less so.

Well, yeah they were talking about both the UltraSparc and UltraSparc
T1 (Niagara) chips. However, I have my doubts how much longer Sun is
going to keep developing the Ultrasparcs, it seems to me that they are
replacing them with Fujitsu's Sparc64's.

Yousuf Khan

 

[Keith]

I see opportunities in the
4S and up though - the return of the glasshouse is taking a grip and I've
pondered the possible viability of an x86 "mainframe" here; a hybrid
Power/x86 mainframe might be just the ticket.

Aak!! 615 Alert!!! Run for your lives!!! ;-)