<strong>How much of a performance jump would one see using VSD? If I understand this right it is a way to run the same process through both CPUs making two CPUs a lot more like one really fast one. If VSD is being used on the 970 I would expect to see more than one dual powermac. I imagine it will be like today. Bottom line single, middle and top dual. </strong><hr></blockquote>
More detail is needed about what this VSD is. If it's a layer that makes two completely separate CPUs look and act like two cores on one die, or even two CPUS on a shared bus, then it's really nice but not earthshaking. It would basically allow Apple to radically change the underlying architecture without the applications noticing, but it would not improve performance.
In fact, I wouldn't be surprised if it introduced a little overhead.
[quote]<strong>Question: Is there a way to run a dual system through two separate busses. Where each CPU has it's own bus and they are connect at the system controller? If apple ever mad quads for there Xserves would this not help performance? Just wondering. </strong><hr></blockquote>
Yes. This is how a NUMA architecture is built. It's been popular in e.g. SGI workstations, but until this year - really - it's been prohibitively expensive to design. HyperTransport, RapidIO, and IBM's Giga bus (the one on the 970) change that.
The 970 can only be set up this way, although if I recall there's another bus that can go directly between processors, setting up a fabric of sorts (rather than everything going through a central controller). This would correspond to something like the Reserved SMP support in the 7400, allowing CPUs to keep their respective memory banks coherent without having to go through the controller.
Note that in a NUMA architecture, each CPU has its own memory controller and its own RAM! That's where VSD becomes interesting, if it is what I think it is: It would provide the illusion that the architecture was x number of CPUs on a shared bus, communicating with a monolithic bank of RAM - the current, traditional architecture.
[quote]<strong>Also can anyone come up with a good reason as to why the 7455 is running at 1.42GHz but we had rumors of Moto having trouble getting the 7457 past 1.42Ghz? Kind of weird in my opinion. I was wondering if apple perhaps turned down the 7457 in favor of the 970. Moto thus deciding not to bother with it.</strong><hr></blockquote>
People might have confused the 7455 and the 7457. Also, the 7455 is a mature design running on a mature process (and really, who expected it to make it to 1.42GHz?) while the '57 will be a new design (not radically new, but tweaked for the smaller process) on a new process, so speeds and yields won't be as good. As the process matures, the 7457 should pick up MHz at a decent clip.
The 7457 does not in any way compete with the 970. They're two different chips aimed at two different markets. Mot's embedded customers love the 7455, and they'll love the 7457, regardless of its fitness for professional Apple workstations. There's no question that Mot will produce it. The question is if and how Apple will use it. It's a safe bet that it will continue to power portables and consumer machines after the 970 appears.
Mot's rumored competitor for the 970 was apparently codenamed Eleven, and it apparently got shelved or canned.
<strong>Note that in a NUMA architecture, each CPU has its own memory controller and its own RAM! That's where VSD becomes interesting, if it is what I think it is: It would provide the illusion that the architecture was x number of CPUs on a shared bus, communicating with a monolithic bank of RAM - the current, traditional architecture.</strong><hr></blockquote>
Yes, I believe that's what it is. Most NUMA systems allow direct addressability of all the memory, but non-local memory is slower to retrieve than local memory. The VSD system is probably a mechanism to move around logical memory pages to keep them local to the processor that needs them. Beyond that, it probably transparently supports non-local memory which is not directly addressable and has to be transfered by some other service (DMA, Ethernet, FireWire, etc).
So, if I understand this right, a dual 970 will be a lot more like two computers in one pretending to be one computer. Or in other words, they will both have their own busses and memory etc but will work together like one, like the Borg. I can see that this setup could end up being extremely fast. How many 970s can this be done to? Could we end up with a quad for the xserves? Maybe a really expensive custom quad tower for around $4000...
Can we just ban these stupid "970 when?" threads, because we seem to having one about every two weeks at the moment - and we don't get any more information for all the fuss it causes.
When: as posted loads of times, second half of 2003 at best - MWSF'04 would hopefully be a worst case scenario, but be prepared for it to appear yet later, if at all.
<strong>[quote]Question: Is there a way to run a dual system through two separate busses. Where each CPU has it's own bus and they are connect at the system controller? If apple ever mad quads for there Xserves would this not help performance? Just wondering.<hr></blockquote></strong>
[quote]<strong>Yes. This is how a NUMA architecture is built. It's been popular in e.g. SGI workstations, but until this year - really - it's been prohibitively expensive to design. HyperTransport, RapidIO, and IBM's Giga bus (the one on the 970) change that.
The 970 can only be set up this way, although if I recall there's another bus that can go directly between processors, setting up a fabric of sorts (rather than everything going through a central controller). This would correspond to something like the Reserved SMP support in the 7400, allowing CPUs to keep their respective memory banks coherent without having to go through the controller.</strong><hr></blockquote>
Would it then be okay to expect a quad 970 on a theoretically 3.6 Ghz bus?
[quote]Originally posted by Clive:
<strong>Can we just ban these stupid "970 when?" threads, because we seem to having one about every two weeks at the moment - and we don't get any more information for all the fuss it causes.</strong><hr></blockquote>
Some of us dont live their lives on AI, and therefore dont know all the gossip.
And if you dont like the smell in the bakery, then stay away!
Some of us dont live their lives on AI, and therefore dont know all the gossip.
And if you dont like the smell in the bakery, then stay away! </strong><hr></blockquote>
You don't have to spend your life here, it's been a recurring theme for about four or five months now, every couple of weeks. It's not adding any new information and just makes the whole board less useful.
It's easy to say "if you don't like...", but I want to read fresh info too, all I get is the same stale old dough!
<strong>So, if I understand this right, a dual 970 will be a lot more like two computers in one pretending to be one computer. Or in other words, they will both have their own busses and memory etc but will work together like one, like the Borg. I can see that this setup could end up being extremely fast. How many 970s can this be done to? Could we end up with a quad for the xserves? Maybe a really expensive custom quad tower for around $4000...</strong><hr></blockquote>
It would depend on how the machine is built. A typical dual 970, however, would most likely have a dual ported companion chip and the only thing that the 970's would not share would be the FSB. The same is probably true of quad processors. More than 4 ports would result in an unruly number of pins on the chipset, however, and the chipset would have to be a very fast memory controller in order to satisfy the appetite of 4 970s. Beyond this level the system would have to start using something like RapidIO, PCI-X, or some other system/board level bus to exchange data and this would cause a performance drop on anything they share. They could all still connect to the same I/O subsystem, however. The next step is from systems to clusters and the interconnects become Ethernet, FibreChannel, FireWire, or something along those lines. At that point only the communications channel is shared and each machine has its own I/O and even power supply.
VSD isn't some magic solution, but it does allow the easier development of clustered applications.
It would depend on how the machine is built. A typical dual 970, however, would most likely have a dual ported companion chip and the only thing that the 970's would not share would be the FSB. The same is probably true of quad processors. More than 4 ports would result in an unruly number of pins on the chipset, however, and the chipset would have to be a very fast memory controller in order to satisfy the appetite of 4 970s. Beyond this level the system would have to start using something like RapidIO, PCI-X, or some other system/board level bus to exchange data and this would cause a performance drop on anything they share. They could all still connect to the same I/O subsystem, however. The next step is from systems to clusters and the interconnects become Ethernet, FibreChannel, FireWire, or something along those lines. At that point only the communications channel is shared and each machine has its own I/O and even power supply.
VSD isn't some magic solution, but it does allow the easier development of clustered applications.</strong><hr></blockquote>
How about the companion chip having 2 ports to each CPU plus a RIO port to another companion chip. This would require seperate memory spaces for each pair of processors, but it would allow for a low pin count on the companion chips.
<strong>How about the companion chip having 2 ports to each CPU plus a RIO port to another companion chip. This would require seperate memory spaces for each pair of processors, but it would allow for a low pin count on the companion chips.</strong><hr></blockquote>
Heh, I wasn't trying to develop and exhaustive list. There are very many possible topologies, unlike MPX which is really only usable as a single shared bus and snooping between buses isn't inherently supported (although could be done).
As far as I understood from previous discussions, it is not the technical side that prevents Apple from going quad or octo but patent-issues. It cost Apple half a fortune to buy back the rights for dual-configurations from one of the former Mac-clone manufacturers.
As far as I understood from previous discussions, it is not the technical side that prevents Apple from going quad or octo but patent-issues. It cost Apple half a fortune to buy back the rights for dual-configurations from one of the former Mac-clone manufacturers.</strong><hr></blockquote>
You have some reference for this?
I can't see that this has anything to do with patent, because it would preclude *any* dual/quad/octo/etc processor machines - and that's clearly not the case.
Possibly Apple granted someone an exclusive licence to develop multiple processor machines - ie they couldn't do it themselves, under contract terms. But again this sounds unlikely to me.
AFAIR the only quad processor Macs were produced by Daystar Digital.
I believe that there have been, and probably are still, multiple vendors for multiple processors PPC boards/machines (running *nix). Easiest example is the upgrade market where you can easily find multiple vendors for dual processor Mac upgrades.
Possibly Apple granted someone an exclusive licence to develop multiple processor machines - ie they couldn't do it themselves, under contract terms.
</strong><hr></blockquote>
You are right, this was it! (Sorry, I did not know how to express that in English).
As far as I can remember this license dates back to the times of the PPC 604.
BTW: Did you ever wonder why no third-party supplier since that time offered quad- or octo-configurations? This could have been a beautiful niche for them!
You are right, this was it! (Sorry, I did not know how to express that in English).
As far as I can remember this license dates back to the times of the PPC 604.
BTW: Did you ever wonder why no third-party supplier since that time offered quad- or octo-configurations? This could have been a beautiful niche for them!
I may be off here, but if I remember correctly Daystar developed the Dual cards, and the system extensions to utalize them. Apple licensed this technology from Daystar to build their dual computers. Apple designed a mother board with 2 daughter card slots that Daystar used to build Quads (2 dual processor daughter cards).
The reason that no one has since built a Quad system is that Apple stoped building a motherboard that would support 2 daughter cards. I'm not a hardware engineer, but I would be willing to bet that there are hardware limitations that prohibit putting 2 G4's on one daughter card. Also, from what I have read on these boards, issues like system bus speed and memory speeds would make this "upgrade" a bad value.
One last comment that I only have a laymans grasp of the legal issues. Apple can get around any patent issues with the hardware by engineering their own solution, or using one developed by IBM to get Quad configurations. One the software side, the extensions were for OS 7/8/9... OS X has a new core which natively supports MP systems so they would not have to worry about paying licensing fees to Daystar for MP support once they leave OS 9 complealy behhind.
For those hoping for an NAB release; We just received an update of the full conference schedule. As far as I can tell, Apple in not involved in any of the Key Notes, Special Events, Super Sessions, Leaderships Sessions, or Workshops.
There will be a couple of sessions in the two-day Digital Video Production Workshop on Final Cut Pro and DVD Studio Pro, but no Apple employees are scheduled to be involved.
This, of course, doesn't obviate Apple from doing an intro on the vendor floor, but it seems unlikely for a product this significant.
It doesn't appear to me that Apple has any plans for a Mac with more than 2 CPUs for at least the next couple of years.
I'm not saying anything about how great it would or wouldn't be, or whether they can or can't do it for technical or legal reasons. I'm just saying that they won't.
Comments
<strong>How much of a performance jump would one see using VSD? If I understand this right it is a way to run the same process through both CPUs making two CPUs a lot more like one really fast one. If VSD is being used on the 970 I would expect to see more than one dual powermac. I imagine it will be like today. Bottom line single, middle and top dual. </strong><hr></blockquote>
More detail is needed about what this VSD is. If it's a layer that makes two completely separate CPUs look and act like two cores on one die, or even two CPUS on a shared bus, then it's really nice but not earthshaking. It would basically allow Apple to radically change the underlying architecture without the applications noticing, but it would not improve performance.
In fact, I wouldn't be surprised if it introduced a little overhead.
[quote]<strong>Question: Is there a way to run a dual system through two separate busses. Where each CPU has it's own bus and they are connect at the system controller? If apple ever mad quads for there Xserves would this not help performance? Just wondering. </strong><hr></blockquote>
Yes. This is how a NUMA architecture is built. It's been popular in e.g. SGI workstations, but until this year - really - it's been prohibitively expensive to design. HyperTransport, RapidIO, and IBM's Giga bus (the one on the 970) change that.
The 970 can only be set up this way, although if I recall there's another bus that can go directly between processors, setting up a fabric of sorts (rather than everything going through a central controller). This would correspond to something like the Reserved SMP support in the 7400, allowing CPUs to keep their respective memory banks coherent without having to go through the controller.
Note that in a NUMA architecture, each CPU has its own memory controller and its own RAM! That's where VSD becomes interesting, if it is what I think it is: It would provide the illusion that the architecture was x number of CPUs on a shared bus, communicating with a monolithic bank of RAM - the current, traditional architecture.
[quote]<strong>Also can anyone come up with a good reason as to why the 7455 is running at 1.42GHz but we had rumors of Moto having trouble getting the 7457 past 1.42Ghz? Kind of weird in my opinion. I was wondering if apple perhaps turned down the 7457 in favor of the 970. Moto thus deciding not to bother with it.</strong><hr></blockquote>
People might have confused the 7455 and the 7457. Also, the 7455 is a mature design running on a mature process (and really, who expected it to make it to 1.42GHz?) while the '57 will be a new design (not radically new, but tweaked for the smaller process) on a new process, so speeds and yields won't be as good. As the process matures, the 7457 should pick up MHz at a decent clip.
The 7457 does not in any way compete with the 970. They're two different chips aimed at two different markets. Mot's embedded customers love the 7455, and they'll love the 7457, regardless of its fitness for professional Apple workstations. There's no question that Mot will produce it. The question is if and how Apple will use it. It's a safe bet that it will continue to power portables and consumer machines after the 970 appears.
Mot's rumored competitor for the 970 was apparently codenamed Eleven, and it apparently got shelved or canned.
<strong>Note that in a NUMA architecture, each CPU has its own memory controller and its own RAM! That's where VSD becomes interesting, if it is what I think it is: It would provide the illusion that the architecture was x number of CPUs on a shared bus, communicating with a monolithic bank of RAM - the current, traditional architecture.</strong><hr></blockquote>
Yes, I believe that's what it is. Most NUMA systems allow direct addressability of all the memory, but non-local memory is slower to retrieve than local memory. The VSD system is probably a mechanism to move around logical memory pages to keep them local to the processor that needs them. Beyond that, it probably transparently supports non-local memory which is not directly addressable and has to be transfered by some other service (DMA, Ethernet, FireWire, etc).
<strong>IBM are under a legal obligation, as a result of an anti-trust settlement, to announce (reasonably) accurately their forthcoming plans.
[...]
michael</strong><hr></blockquote>
/rant
Ok, not to pick on you Michael, because I see this particular claim over and over and over...but it isn't true anymore:
<a href="http://www.usdoj.gov/opa/pr/1996/July96/324.at.html"; target="_blank">http://www.usdoj.gov/opa/pr/1996/July96/324.at.html</a>;
That is one of the problems with sites like this, people make a claim, and then it is parroted over and over without seeing if it is true.
/end rant
When: as posted loads of times, second half of 2003 at best - MWSF'04 would hopefully be a worst case scenario, but be prepared for it to appear yet later, if at all.
<strong>[quote]Question: Is there a way to run a dual system through two separate busses. Where each CPU has it's own bus and they are connect at the system controller? If apple ever mad quads for there Xserves would this not help performance? Just wondering.<hr></blockquote></strong>
[quote]<strong>Yes. This is how a NUMA architecture is built. It's been popular in e.g. SGI workstations, but until this year - really - it's been prohibitively expensive to design. HyperTransport, RapidIO, and IBM's Giga bus (the one on the 970) change that.
The 970 can only be set up this way, although if I recall there's another bus that can go directly between processors, setting up a fabric of sorts (rather than everything going through a central controller). This would correspond to something like the Reserved SMP support in the 7400, allowing CPUs to keep their respective memory banks coherent without having to go through the controller.</strong><hr></blockquote>
Would it then be okay to expect a quad 970 on a theoretically 3.6 Ghz bus?
[quote]Originally posted by Clive:
<strong>Can we just ban these stupid "970 when?" threads, because we seem to having one about every two weeks at the moment - and we don't get any more information for all the fuss it causes.</strong><hr></blockquote>
Some of us dont live their lives on AI, and therefore dont know all the gossip.
And if you dont like the smell in the bakery, then stay away!
[ 02-07-2003: Message edited by: T'hain Esh Kelch ]</p>
<strong>
And if you dont like the smell in the bakery, then stay away!
Now I feel like freshly baked warm bread damn it <img src="embarrassed.gif" border="0">
[ 02-07-2003: Message edited by: Telomar ]</p>
<strong>
Some of us dont live their lives on AI, and therefore dont know all the gossip.
And if you dont like the smell in the bakery, then stay away!
You don't have to spend your life here, it's been a recurring theme for about four or five months now, every couple of weeks. It's not adding any new information and just makes the whole board less useful.
It's easy to say "if you don't like...", but I want to read fresh info too, all I get is the same stale old dough!
Admin, please lock this thread!
<strong>So, if I understand this right, a dual 970 will be a lot more like two computers in one pretending to be one computer. Or in other words, they will both have their own busses and memory etc but will work together like one, like the Borg. I can see that this setup could end up being extremely fast. How many 970s can this be done to? Could we end up with a quad for the xserves? Maybe a really expensive custom quad tower for around $4000...</strong><hr></blockquote>
It would depend on how the machine is built. A typical dual 970, however, would most likely have a dual ported companion chip and the only thing that the 970's would not share would be the FSB. The same is probably true of quad processors. More than 4 ports would result in an unruly number of pins on the chipset, however, and the chipset would have to be a very fast memory controller in order to satisfy the appetite of 4 970s. Beyond this level the system would have to start using something like RapidIO, PCI-X, or some other system/board level bus to exchange data and this would cause a performance drop on anything they share. They could all still connect to the same I/O subsystem, however. The next step is from systems to clusters and the interconnects become Ethernet, FibreChannel, FireWire, or something along those lines. At that point only the communications channel is shared and each machine has its own I/O and even power supply.
VSD isn't some magic solution, but it does allow the easier development of clustered applications.
<strong>
It would depend on how the machine is built. A typical dual 970, however, would most likely have a dual ported companion chip and the only thing that the 970's would not share would be the FSB. The same is probably true of quad processors. More than 4 ports would result in an unruly number of pins on the chipset, however, and the chipset would have to be a very fast memory controller in order to satisfy the appetite of 4 970s. Beyond this level the system would have to start using something like RapidIO, PCI-X, or some other system/board level bus to exchange data and this would cause a performance drop on anything they share. They could all still connect to the same I/O subsystem, however. The next step is from systems to clusters and the interconnects become Ethernet, FibreChannel, FireWire, or something along those lines. At that point only the communications channel is shared and each machine has its own I/O and even power supply.
VSD isn't some magic solution, but it does allow the easier development of clustered applications.</strong><hr></blockquote>
How about the companion chip having 2 ports to each CPU plus a RIO port to another companion chip. This would require seperate memory spaces for each pair of processors, but it would allow for a low pin count on the companion chips.
<strong>How about the companion chip having 2 ports to each CPU plus a RIO port to another companion chip. This would require seperate memory spaces for each pair of processors, but it would allow for a low pin count on the companion chips.</strong><hr></blockquote>
Heh, I wasn't trying to develop and exhaustive list.
<strong>
About cost: Many people assume that the cost of the IBM 970 will preclude it from being offered in dual, much less quad or octo configurations.
[ 01-26-2003: Message edited by: Aphelion ]</strong><hr></blockquote>
As far as I understood from previous discussions, it is not the technical side that prevents Apple from going quad or octo but patent-issues. It cost Apple half a fortune to buy back the rights for dual-configurations from one of the former Mac-clone manufacturers.
<strong>
As far as I understood from previous discussions, it is not the technical side that prevents Apple from going quad or octo but patent-issues. It cost Apple half a fortune to buy back the rights for dual-configurations from one of the former Mac-clone manufacturers.</strong><hr></blockquote>
You have some reference for this?
I can't see that this has anything to do with patent, because it would preclude *any* dual/quad/octo/etc processor machines - and that's clearly not the case.
Possibly Apple granted someone an exclusive licence to develop multiple processor machines - ie they couldn't do it themselves, under contract terms. But again this sounds unlikely to me.
AFAIR the only quad processor Macs were produced by Daystar Digital.
I believe that there have been, and probably are still, multiple vendors for multiple processors PPC boards/machines (running *nix). Easiest example is the upgrade market where you can easily find multiple vendors for dual processor Mac upgrades.
<strong>
Possibly Apple granted someone an exclusive licence to develop multiple processor machines - ie they couldn't do it themselves, under contract terms.
</strong><hr></blockquote>
You are right, this was it! (Sorry, I did not know how to express that in English).
As far as I can remember this license dates back to the times of the PPC 604.
BTW: Did you ever wonder why no third-party supplier since that time offered quad- or octo-configurations? This could have been a beautiful niche for them!
[ 02-07-2003: Message edited by: Gulliver ]</p>
<strong>
You are right, this was it! (Sorry, I did not know how to express that in English).
As far as I can remember this license dates back to the times of the PPC 604.
BTW: Did you ever wonder why no third-party supplier since that time offered quad- or octo-configurations? This could have been a beautiful niche for them!
[ 02-07-2003: Message edited by: Gulliver ]</strong><hr></blockquote>
I may be off here, but if I remember correctly Daystar developed the Dual cards, and the system extensions to utalize them. Apple licensed this technology from Daystar to build their dual computers. Apple designed a mother board with 2 daughter card slots that Daystar used to build Quads (2 dual processor daughter cards).
The reason that no one has since built a Quad system is that Apple stoped building a motherboard that would support 2 daughter cards. I'm not a hardware engineer, but I would be willing to bet that there are hardware limitations that prohibit putting 2 G4's on one daughter card. Also, from what I have read on these boards, issues like system bus speed and memory speeds would make this "upgrade" a bad value.
One last comment that I only have a laymans grasp of the legal issues. Apple can get around any patent issues with the hardware by engineering their own solution, or using one developed by IBM to get Quad configurations. One the software side, the extensions were for OS 7/8/9... OS X has a new core which natively supports MP systems so they would not have to worry about paying licensing fees to Daystar for MP support once they leave OS 9 complealy behhind.
There will be a couple of sessions in the two-day Digital Video Production Workshop on Final Cut Pro and DVD Studio Pro, but no Apple employees are scheduled to be involved.
This, of course, doesn't obviate Apple from doing an intro on the vendor floor, but it seems unlikely for a product this significant.
I'm not saying anything about how great it would or wouldn't be, or whether they can or can't do it for technical or legal reasons. I'm just saying that they won't.