Getting away from the love-in that Melgross and Sunil are having, I came across this when I was looking for any hints as to whether the system controller (U3H or a revised CP925) in the G5 would have to be revised with the transition to dual core.
The answer, according to this is yes...a new controller will be needed:
This poses an interesting question: If they have to come up with a new system controller, how long will it take and will it seek to improve upon the latency issues. If you remember back a couple of years, the system controller was the part the caused the greatest troubles in the initial G5.
Getting away from the love-in that Melgross and Sunil are having, I came across this when I was looking for any hints as to whether the system controller (U3H or a revised CP925) in the G5 would have to be revised with the transition to dual core.
The answer, according to this is yes...a new controller will be needed:
This poses an interesting question: If they have to come up with a new system controller, how long will it take and will it seek to improve upon the latency issues. If you remember back a couple of years, the system controller was the part the caused the greatest troubles in the initial G5.
You are mis-reading those posts. The current chipset can support two FSB connections, which means it can support two FSB-equipped chips. If those chips happen to have two (or more) cores hung off the same FSB connection, then this is transparent to the chipset (it just sees the one FSB which happens to be shared between cores).
My guess is that Apple will make a minor revision to their chipset and that's it... I might be wrong, however, since if the move to x86 really was decided only recently they would have been proceeding at full speed with whatever their U3 revision would have been. It is possible that they had a new U3 supporting PCI-E, DDR2, etc which was close to production, in which case they will push it out the pipeline and we'll see one or two PowerMac revisions with it before the PM goes x86 in 2007.
The datasheet has interesting figures on data consumption that underlines the misleading figures given on the press release. It's good to see they got to 1.0V 970s though, albeit for 1.2Ghz and 18 months later than they said they would.
Quote:
Originally posted by Programmer
You are mis-reading those posts. The current chipset can support two FSB connections, which means it can support two FSB-equipped chips. If those chips happen to have two (or more) cores hung off the same FSB connection, then this is transparent to the chipset (it just sees the one FSB which happens to be shared between cores).
My guess is that Apple will make a minor revision to their chipset and that's it... I might be wrong, however, since if the move to x86 really was decided only recently they would have been proceeding at full speed with whatever their U3 revision would have been. It is possible that they had a new U3 supporting PCI-E, DDR2, etc which was close to production, in which case they will push it out the pipeline and we'll see one or two PowerMac revisions with it before the PM goes x86 in 2007.
I don't know if this adds something to the discussion, but here is an IBM document where one can see that the 970MP was expected to reach 2.7 GHz. Perhaps Apple will finally get the MP chips at this clock speed. There are interesting power figures for several chips in this document.
Of interest also is the next generation MP for next year, 2.0-4.0 GHz with 2 MB L2 / core. Will IBM deliver or not is now the question.
Of interest also is the next generation MP for next year, 2.0-4.0 GHz with 2 MB L2 / core. Will IBM deliver or not is now the question.
Yeah.. but you know accrding to Nr.9 there's a industry consensus that the wall everyone ran into some years ago is never going to be breached. We will _never_ see a PowerMac with processors faster than 2.5 GHz.. Never!
These are not the GHz we're looking for. *waving my hand*
I don't know if this adds something to the discussion, but here is an IBM document where one can see that the 970MP was expected to reach 2.7 GHz. Perhaps Apple will finally get the MP chips at this clock speed. There are interesting power figures for several chips in this document.
Of interest also is the next generation MP for next year, 2.0-4.0 GHz with 2 MB L2 / core. Will IBM deliver or not is now the question.
I think that is a great link very relevant to our discussion. If you look at the graphs, notice how the MP is marked as 'limited availability' and everything beyond that is all just clearly shown as vaporware (that is, 'proposed')
I think Steve very likely saw that this kind of stuff this time last year, and started his crusade to release Tiger UniversalBinary by WWDC 2005.
Also, the power/performance curves on those graphs aren't promising as well. I (and apparently IBM at the time of making those slides) do not know what intel curves are like, but those IBM curves look very very painfully exponential past 2ghz.
Steve did a bit of dy/dx action and probably saw that even if intel's power/performance curves was somewhat exponential, it was much more promising than this IBM *ahem* garbage**
I think that is a great link very relevant to our discussion. If you look at the graphs, notice how the MP is marked as 'limited availability' and everything beyond that is all just clearly shown as vaporware (that is, 'proposed')
Well, it was done almost a year ago. But even so you are right, Intel have solid product roadmaps into 2007, as do AMD. For IBM to provide a roadmap into 2005, not 2006 or 2007 probably got Jobs very worried.
Not to mention that slide there that clearly said that Intel was better performing at a certain range, that just so happened to include all of the notebook wattages. Maybe they should have ported the 750GX to 90nm (with a VMX unit too) and stuck two of them on a die, lol.
Quote:
Also, the power/performance curves on those graphs aren't promising as well. I (and apparently IBM at the time of making those slides) do not know what intel curves are like, but those IBM curves look very very painfully exponential past 2ghz.
This is typical, after a certain stage you have to ramp up the voltages to get the chip to operate at the higher speed, and also the processor draws more current at a higher speed. P=VI, so hence the exponential curve. Notice the straight areas - that's constant voltage with increasing current, the jumps are probably jumps in voltage of between 0.05 and 0.1 V. This document shows that the 970FX operates ideally underneath 2GHz. I'm sure that since the creation of this document the process has improved for maybe a 0.1 V drop across the board though, making the processor fine below 2.3GHz. IBM just supplies those 105C processors to Apple that can run at 2.7GHz at 0.2 V or so higher than standard.
Now make that dual-core. That just means 2x the power at a certain clockspeed. The total processor power usage Pp is2 x Ic x Vc. If your core is using under 50W at 2.3GHz and below (the straight bit of the earlier graph) then the power hit is merely 2x that straight bit. But in the exponential part of the curve it gets a bit nasty, and you don't want power consumptions that high - the core temperature isn't usually an issue, it's the power consumption and removal of a similar core temperature over twice the die area.
From Intel's dual-core Pentium D we can see their sweet spot for dual-cores is 3.2GHz and below. From AMD's we can see their sweet spot is 2.4GHz and below. Namely at these speeds a single core uses under 40-50W, so you can safely make a dual-core to fit inside the platform's TDP specification. Above that however the exponential part of the power consumption curve comes into play.
IBM unveiled their roadmap for 2006 on their site:
After the single core PPC970GX and its 1 MB L2, the dual core PPC970MP and its 2 x 1 MB dedicated L2 (2005), come in 2006 the dualcore PPC970MX and its 2 MB shared L2, 3+ GHz.
We know also that after 970MX, the next gen dualcore will have 2 MB L2 per core, and will perform between 2 and 4 GHz. But then I think Intel will be way beyond that (more than quadcore).
There is also comparison charts with Intel Pentium M Dothan and upcoming Freescale e600, about the "performance per watt" ratio. Interesting.
Something that's really interessting is that 970GX was generally available a year ago. So.. where is it?!
Where did you see this? I see availability for this year only (in second page). As for the 970MX, it is under the "product proposal" label, so maybe it is already dead. Furthermore, the 970MP at 1.5-2.7 GHz should be available some time between 2004 and now. Yet, we don't know when it will go into production.
And what is this next generation MP, 2.0-4.0 GHz? The alleged Power5 derivative a la PowerPC 970?
And what is this next generation MP, 2.0-4.0 GHz? The alleged Power5 derivative a la PowerPC 970?
Looks like a cache increase to me. And remember the Power5 core is just a Power4 core with SMT added. No big deal. Now if they added a high performance backside L3 and on-chip memory controller, that would be a pretty good win for 970 performance.
Nice -- it is clearly marked IBM confidential, and it shows IBM admitting that Intel is better in the power range Apple cares about. And remember, IBM is not privvy to Intel's roadmap so they are comparing to current offerings, not whatever magic Intel has coming.
The 970MX and later clock rate numbers (3-4 GHz) are targets, not guarantees. Judging from the Cell design they can deliver high clock rates at high power levels by ripping out most of the complex OoOE logic. This might mean the processors aren't actually any faster on typical code than they are on a current 970 at a much lower clock rate. MHz Myth applies within a processor family too.
Looks like a cache increase to me. And remember the Power5 core is just a Power4 core with SMT added. No big deal.
The absence of the 970 moniker in this future MP makes me think that it is more than the obvious cache increase. And Power5, if I am not mistaken, has solved the latency problems of the 970 generation.
The absence of the 970 moniker in this future MP makes me think that it is more than the obvious cache increase.
Well, like Programmer said, it could just be a PPE (from Cell/Xenon) based multicore chip. Who knows. But if it is "Power5" core based, then it's just a PPC 970 core with SMT and more registers, possibly more L1 cache. There really wasn't anything big that happened to the Power4 core to make it a Power5 core.
Ruminating:
970mp is a 90 nm chip with 1 MB L2 per core
970mx is a 65 nm chip with 2 MB shared L2
next gen MP has 2 MB L2 per core at 65 nm
Hmm, I'm thinking that with the 970mx being a shared 2 MB L2 dual core and the "next gen MP" being a 2 MB L2 per core, hmm maybe this next gen MP is just a dual-dual: 2 970mx cores on a chip.
Quote:
And Power5, if I am not mistaken, has solved the latency problems of the 970 generation.
The Power5's decreased latency over the Power4 by implementing a better memory subsystem. Hence, a 970 with backside L3 cache and on-chip memory controller would improve its performance pretty well.
Well, we don?t know yet all the facts about the 970MP and 970FX (availability, heat, price, etc...), but I believe 2 things:
- no way Apple will use the FX series in PowerBook or iMacs, they are just too slow even a little overclocked. Apple would have better use of Freescale 7448 for the next revision of PowerBooks (before the switch to Intel) as it won?t change the architecture of the motherboards (it still a 32-bit CPU, as Yonah, if it?s really the Intel?s CPU that will be used by Apple for portables and entry-level desktops).
- I hope that Apple will use the 970MP in PowerMacs and iMacs (if the price is right) as soon as they became available (january 2006?). Even at the low speeds they?re being announced (1.4 to 2.5GHz), they can provide nearly twice the processing than the actual models.
iMac G5 (currently at 1.8, 2.0GHz) maybe speedbumped to 2x1.4 and 2x1.6, that would meant around 1.5 more performance.
PowerMac G5 (currently at 2x2.0, 2x2.3, 2x2.7GHz) maybe speedbumped to 4x1.6 (dual dual-core), 4x1.8 and 4x2.0, that would meant again around 1.5 more performance.
Doing so, there is still room for another update (late 2006) up to 4x2.5Ghz for PowerMacs and maybe 2x2.0GHz for iMacs (another 1.25 gain in performance) before switching to Intel?s CONROE CPU around june 2007 (?).
Well, we don?t know yet all the facts about the 970MP and 970FX (availability, heat, price, etc...), but I believe 2 things:
- no way Apple will use the FX series in PowerBook or iMacs, they are just too slow even a little overclocked. Apple would have better use of Freescale 7448 for the next revision of PowerBooks (before the switch to Intel) as it won?t change the architecture of the motherboards (it still a 32-bit CPU, as Yonah, if it?s really the Intel?s CPU that will be used by Apple for portables and entry-level desktops).
- I hope that Apple will use the 970MP in PowerMacs and iMacs (if the price is right) as soon as they became available (january 2006?). Even at the low speeds they?re being announced (1.4 to 2.5GHz), they can provide nearly twice the processing than the actual models.
iMac G5 (currently at 1.8, 2.0GHz) maybe speedbumped to 2x1.4 and 2x1.6, that would meant around 1.5 more performance.
PowerMac G5 (currently at 2x2.0, 2x2.3, 2x2.7GHz) maybe speedbumped to 4x1.6 (dual dual-core), 4x1.8 and 4x2.0, that would meant again around 1.5 more performance.
Doing so, there is still room for another update (late 2006) up to 4x2.5Ghz for PowerMacs and maybe 2x2.0GHz for iMacs (another 1.25 gain in performance) before switching to Intel?s CONROE CPU around june 2007 (?).
The problem with that idea is that many applications will take a performance hit.
Comments
Originally posted by melgross
Hey, didn't you say that already?
(It's only 11:00 P.M. here. Wait 'till it gets late.)
Originally posted by sunilraman
It's still not late.
The answer, according to this is yes...a new controller will be needed:
http://www-128.ibm.com/developerwork...m=496#13716219
This poses an interesting question: If they have to come up with a new system controller, how long will it take and will it seek to improve upon the latency issues. If you remember back a couple of years, the system controller was the part the caused the greatest troubles in the initial G5.
Originally posted by a j stev
Getting away from the love-in that Melgross and Sunil are having, I came across this when I was looking for any hints as to whether the system controller (U3H or a revised CP925) in the G5 would have to be revised with the transition to dual core.
The answer, according to this is yes...a new controller will be needed:
http://www-128.ibm.com/developerwork...m=496#13716219
This poses an interesting question: If they have to come up with a new system controller, how long will it take and will it seek to improve upon the latency issues. If you remember back a couple of years, the system controller was the part the caused the greatest troubles in the initial G5.
You are mis-reading those posts. The current chipset can support two FSB connections, which means it can support two FSB-equipped chips. If those chips happen to have two (or more) cores hung off the same FSB connection, then this is transparent to the chipset (it just sees the one FSB which happens to be shared between cores).
My guess is that Apple will make a minor revision to their chipset and that's it... I might be wrong, however, since if the move to x86 really was decided only recently they would have been proceeding at full speed with whatever their U3 revision would have been. It is possible that they had a new U3 supporting PCI-E, DDR2, etc which was close to production, in which case they will push it out the pipeline and we'll see one or two PowerMac revisions with it before the PM goes x86 in 2007.
In passing, IBM has cleared out its inventory of 970FX documents. It seems like they having nothing to hide now. They can be seen here:
http://www-306.ibm.com/chips/techlib...icroprocessors
The datasheet has interesting figures on data consumption that underlines the misleading figures given on the press release. It's good to see they got to 1.0V 970s though, albeit for 1.2Ghz and 18 months later than they said they would.
Originally posted by Programmer
You are mis-reading those posts. The current chipset can support two FSB connections, which means it can support two FSB-equipped chips. If those chips happen to have two (or more) cores hung off the same FSB connection, then this is transparent to the chipset (it just sees the one FSB which happens to be shared between cores).
My guess is that Apple will make a minor revision to their chipset and that's it... I might be wrong, however, since if the move to x86 really was decided only recently they would have been proceeding at full speed with whatever their U3 revision would have been. It is possible that they had a new U3 supporting PCI-E, DDR2, etc which was close to production, in which case they will push it out the pipeline and we'll see one or two PowerMac revisions with it before the PM goes x86 in 2007.
Of interest also is the next generation MP for next year, 2.0-4.0 GHz with 2 MB L2 / core. Will IBM deliver or not is now the question.
Originally posted by PB
Of interest also is the next generation MP for next year, 2.0-4.0 GHz with 2 MB L2 / core. Will IBM deliver or not is now the question.
Yeah.. but you know accrding to Nr.9 there's a industry consensus that the wall everyone ran into some years ago is never going to be breached. We will _never_ see a PowerMac with processors faster than 2.5 GHz.. Never!
These are not the GHz we're looking for. *waving my hand*
Originally posted by Henriok
.....These are not the GHz we're looking for. *waving my hand*
Originally posted by PB
I don't know if this adds something to the discussion, but here is an IBM document where one can see that the 970MP was expected to reach 2.7 GHz. Perhaps Apple will finally get the MP chips at this clock speed. There are interesting power figures for several chips in this document.
Of interest also is the next generation MP for next year, 2.0-4.0 GHz with 2 MB L2 / core. Will IBM deliver or not is now the question.
I think that is a great link very relevant to our discussion. If you look at the graphs, notice how the MP is marked as 'limited availability' and everything beyond that is all just clearly shown as vaporware (that is, 'proposed')
I think Steve very likely saw that this kind of stuff this time last year, and started his crusade to release Tiger UniversalBinary by WWDC 2005.
Also, the power/performance curves on those graphs aren't promising as well. I (and apparently IBM at the time of making those slides) do not know what intel curves are like, but those IBM curves look very very painfully exponential past 2ghz.
Steve did a bit of dy/dx action and probably saw that even if intel's power/performance curves was somewhat exponential, it was much more promising than this IBM *ahem* garbage**
**just being facetious here
Originally posted by sunilraman
I think that is a great link very relevant to our discussion. If you look at the graphs, notice how the MP is marked as 'limited availability' and everything beyond that is all just clearly shown as vaporware (that is, 'proposed')
Well, it was done almost a year ago. But even so you are right, Intel have solid product roadmaps into 2007, as do AMD. For IBM to provide a roadmap into 2005, not 2006 or 2007 probably got Jobs very worried.
Not to mention that slide there that clearly said that Intel was better performing at a certain range, that just so happened to include all of the notebook wattages. Maybe they should have ported the 750GX to 90nm (with a VMX unit too) and stuck two of them on a die, lol.
Also, the power/performance curves on those graphs aren't promising as well. I (and apparently IBM at the time of making those slides) do not know what intel curves are like, but those IBM curves look very very painfully exponential past 2ghz.
This is typical, after a certain stage you have to ramp up the voltages to get the chip to operate at the higher speed, and also the processor draws more current at a higher speed. P=VI, so hence the exponential curve. Notice the straight areas - that's constant voltage with increasing current, the jumps are probably jumps in voltage of between 0.05 and 0.1 V. This document shows that the 970FX operates ideally underneath 2GHz. I'm sure that since the creation of this document the process has improved for maybe a 0.1 V drop across the board though, making the processor fine below 2.3GHz. IBM just supplies those 105C processors to Apple that can run at 2.7GHz at 0.2 V or so higher than standard.
Now make that dual-core. That just means 2x the power at a certain clockspeed. The total processor power usage Pp is2 x Ic x Vc. If your core is using under 50W at 2.3GHz and below (the straight bit of the earlier graph) then the power hit is merely 2x that straight bit. But in the exponential part of the curve it gets a bit nasty, and you don't want power consumptions that high - the core temperature isn't usually an issue, it's the power consumption and removal of a similar core temperature over twice the die area.
From Intel's dual-core Pentium D we can see their sweet spot for dual-cores is 3.2GHz and below. From AMD's we can see their sweet spot is 2.4GHz and below. Namely at these speeds a single core uses under 40-50W, so you can safely make a dual-core to fit inside the platform's TDP specification. Above that however the exponential part of the power consumption curve comes into play.
After the single core PPC970GX and its 1 MB L2, the dual core PPC970MP and its 2 x 1 MB dedicated L2 (2005), come in 2006 the dualcore PPC970MX and its 2 MB shared L2, 3+ GHz.
We know also that after 970MX, the next gen dualcore will have 2 MB L2 per core, and will perform between 2 and 4 GHz. But then I think Intel will be way beyond that (more than quadcore).
There is also comparison charts with Intel Pentium M Dothan and upcoming Freescale e600, about the "performance per watt" ratio. Interesting.
Here is the PowerPoint.
Originally posted by Henriok
Something that's really interessting is that 970GX was generally available a year ago. So.. where is it?!
Where did you see this? I see availability for this year only (in second page). As for the 970MX, it is under the "product proposal" label, so maybe it is already dead. Furthermore, the 970MP at 1.5-2.7 GHz should be available some time between 2004 and now. Yet, we don't know when it will go into production.
And what is this next generation MP, 2.0-4.0 GHz? The alleged Power5 derivative a la PowerPC 970?
Originally posted by Henriok
Something that's really interessting is that 970GX was generally available a year ago. So.. where is it?! We could surely use some of that L2 cache..
*All future dates and frequencies are estimations and are subject to change without notice.
Originally posted by PB
And what is this next generation MP, 2.0-4.0 GHz? The alleged Power5 derivative a la PowerPC 970?
Looks like a cache increase to me. And remember the Power5 core is just a Power4 core with SMT added. No big deal. Now if they added a high performance backside L3 and on-chip memory controller, that would be a pretty good win for 970 performance.
The 970MX and later clock rate numbers (3-4 GHz) are targets, not guarantees. Judging from the Cell design they can deliver high clock rates at high power levels by ripping out most of the complex OoOE logic. This might mean the processors aren't actually any faster on typical code than they are on a current 970 at a much lower clock rate. MHz Myth applies within a processor family too.
Originally posted by THT
Looks like a cache increase to me. And remember the Power5 core is just a Power4 core with SMT added. No big deal.
The absence of the 970 moniker in this future MP makes me think that it is more than the obvious cache increase. And Power5, if I am not mistaken, has solved the latency problems of the 970 generation.
Originally posted by PB
The absence of the 970 moniker in this future MP makes me think that it is more than the obvious cache increase.
Well, like Programmer said, it could just be a PPE (from Cell/Xenon) based multicore chip. Who knows. But if it is "Power5" core based, then it's just a PPC 970 core with SMT and more registers, possibly more L1 cache. There really wasn't anything big that happened to the Power4 core to make it a Power5 core.
Ruminating:
970mp is a 90 nm chip with 1 MB L2 per core
970mx is a 65 nm chip with 2 MB shared L2
next gen MP has 2 MB L2 per core at 65 nm
Hmm, I'm thinking that with the 970mx being a shared 2 MB L2 dual core and the "next gen MP" being a 2 MB L2 per core, hmm maybe this next gen MP is just a dual-dual: 2 970mx cores on a chip.
And Power5, if I am not mistaken, has solved the latency problems of the 970 generation.
The Power5's decreased latency over the Power4 by implementing a better memory subsystem. Hence, a 970 with backside L3 cache and on-chip memory controller would improve its performance pretty well.
- no way Apple will use the FX series in PowerBook or iMacs, they are just too slow even a little overclocked. Apple would have better use of Freescale 7448 for the next revision of PowerBooks (before the switch to Intel) as it won?t change the architecture of the motherboards (it still a 32-bit CPU, as Yonah, if it?s really the Intel?s CPU that will be used by Apple for portables and entry-level desktops).
- I hope that Apple will use the 970MP in PowerMacs and iMacs (if the price is right) as soon as they became available (january 2006?). Even at the low speeds they?re being announced (1.4 to 2.5GHz), they can provide nearly twice the processing than the actual models.
iMac G5 (currently at 1.8, 2.0GHz) maybe speedbumped to 2x1.4 and 2x1.6, that would meant around 1.5 more performance.
PowerMac G5 (currently at 2x2.0, 2x2.3, 2x2.7GHz) maybe speedbumped to 4x1.6 (dual dual-core), 4x1.8 and 4x2.0, that would meant again around 1.5 more performance.
Doing so, there is still room for another update (late 2006) up to 4x2.5Ghz for PowerMacs and maybe 2x2.0GHz for iMacs (another 1.25 gain in performance) before switching to Intel?s CONROE CPU around june 2007 (?).
Originally posted by mjteix
Well, we don?t know yet all the facts about the 970MP and 970FX (availability, heat, price, etc...), but I believe 2 things:
- no way Apple will use the FX series in PowerBook or iMacs, they are just too slow even a little overclocked. Apple would have better use of Freescale 7448 for the next revision of PowerBooks (before the switch to Intel) as it won?t change the architecture of the motherboards (it still a 32-bit CPU, as Yonah, if it?s really the Intel?s CPU that will be used by Apple for portables and entry-level desktops).
- I hope that Apple will use the 970MP in PowerMacs and iMacs (if the price is right) as soon as they became available (january 2006?). Even at the low speeds they?re being announced (1.4 to 2.5GHz), they can provide nearly twice the processing than the actual models.
iMac G5 (currently at 1.8, 2.0GHz) maybe speedbumped to 2x1.4 and 2x1.6, that would meant around 1.5 more performance.
PowerMac G5 (currently at 2x2.0, 2x2.3, 2x2.7GHz) maybe speedbumped to 4x1.6 (dual dual-core), 4x1.8 and 4x2.0, that would meant again around 1.5 more performance.
Doing so, there is still room for another update (late 2006) up to 4x2.5Ghz for PowerMacs and maybe 2x2.0GHz for iMacs (another 1.25 gain in performance) before switching to Intel?s CONROE CPU around june 2007 (?).
The problem with that idea is that many applications will take a performance hit.