No, there is no exact number to qualify as an absolute limit. What we're talking about is that crazy overclockers can make PPC970fx run at, say, 5GHz right now, but only in a can of liquid Nitrogen. Since IBM won't even consider shipping insanely overclocked chips, it roughly equals IBM having stuck at 2.5GHz. They may reach 3GHz or so for mass production of PPC970fx, but they are still having problems producing them even at 2.5GHz in quantities sufficient for Apple. You know what I mean: if those 3GHz chips turn out to be too hot, too few or too expensive, it will be just as though IBM haven't made a single chip running at 3GHz. Nobody gives a damn about a theoretical clock speed of a given chip - it only matters what clock speed the chip is running at on your desktop.
Yes, but the point is if Intel can get to 4 GHZ, IBM can as well. It might take a longer pipeline, but it can be done. We've not hit a wall just yet.
Well there are a number of things to consider, one is that the effort to extend clock rate really isn't scaling if you take Intels approach. That is Intel streteched out the pipelines with each (almost all) revision of the Pentium and by doing so slowed the chips down, sometimes considerably. In effect Intel took the approach that went for clock speed at the expense of efficent operation of the CPU.
There is no doubt that IBM could take this approach but there is a big question related to benefits. Intel's approach leaves the processor with a large amount of idle time as it tries to keep those pipelines filled. The thing is that Intel does make out real good when the right types of instruction streams come along, a 4 GHz chip can be twice as fast as a 2 GHz chip under ideal conditions.
The Approach on the PPC is for much shorter pipelines and reduced penalties because of the shorter pipelines. With respect to the 970FX at 2.5GHz this has really worked out well, with many apps seeing the clock rate increase. What one ends up with, with IBM's current tech, is a very hot chip. We really don't know what the top frequency is for this chip as it does seem to be thermally limited, we don't know how fast the logic is capable of operating. That is why I think it is very important to see where IBM goes with reductions in thermal demand and also see where the rest of the industry goes. There is information floating about that does seem to indicate that 90nm done right can pay off and all extending the clock rate of a 130nm design.
What I find interesting though is that my account has been offline two days and this discussion is still going on. I'm also surprised that so many still want to throw in the towel. This is barely round one, it time to get up off the matt and breath some fresh air.
If we want to give Nr9 credit for anything he may be on to the 970MP release. It may very well come out at the same clock rate as the current 970FX's, but that may be related to a number of things. The number one thing problly being the doubling of thermal load. May the number one issue being shared with the thought that IBM still doesn't have an alternative process.
Dave
Quote:
Originally posted by bunge
Yes, but the point is if Intel can get to 4 GHZ, IBM can as well. It might take a longer pipeline, but it can be done. We've not hit a wall just yet.
<snip> your probably just a 5 year old who pays forty bucks a month to get access to the ACM portal and also fail to take note of anything interesting in there. if you are in SIGARCH and haven't noticed the shift towards power consumption as a priority, the forty bucks you pay are probably not worth it. <snip>
I would think you could at least get the price right if you were a member. You can't even get a student membership for $40 and SIGARCH is another $28++ on top of that. Or start with $198 for a full access including the Portal if you're a professional member (plus the $28++ for SIGARCH), but the $40 figure crops up nowhere. The facts keep unravelling, and the rhetoric more vile. The usual for a vain salvage attempt.
I second the lockination here. Maybe a fresh and unsullied thread could take it's place if folks still wanted to hash CPU issues out, but free of this threads baggage.
<snip>If we want to give Nr9 credit for anything he may be on to the 970MP release. It may very well come out at the same clock rate as the current 970FX's, but that may be related to a number of things. The number one thing problly being the doubling of thermal load. May the number one issue being shared with the thought that IBM still doesn't have an alternative process.
Dave
If that hadn't already been discussed thoroughly in the wild since late July he might have had a chance, but two months late to the dance don't get credit for original thought, let alone justify a declaration of no more Hz--ever--for anyone.
...but two months late to the dance don't get credit for original thought, let alone justify a declaration of no more Hz--ever--for anyone.
It is one thing to be late to the dance, it is another thing to come with new information. I would hope that we as a group here would continue to be receptive to information from any number of sources related to things like the 970MP. There has to be more information floating about with respect to that chip that I would like to hear about.
Nr9's decleration that there will be no more Hz increases though is problematic. I'm not sure he understands how this statement so completely underminds everything else he states. Maybe the first release of the MP won't be that much faster than current devices, but that does not mean that the end of future speed increases is here. This may be where Nr9 is confused.
Yes, but the point is if Intel can get to 4 GHZ, IBM can as well. It might take a longer pipeline, but it can be done. We've not hit a wall just yet.
Ugh. Are you ignoring me, or just not getting it? If IBM has decided that the tradeoffs to increase clock rate (power, longer pipelines) aren't worth it, then they have effectively hit a "wall". It is a fuzzy and vaguely defined one, but it is the outer envelope of the processor performance space made possible by the available technology... and the technologies that they have on their roadmap. We won't really know if Nr9 is right until they revise their roadmap and the POWER6 is no longer a 5-6 GHz monster, but is instead a massively multi-core one.
We won't really know if Nr9 is right until they revise their roadmap and the POWER6 is no longer a 5-6 GHz monster, but is instead a massively multi-core one.
The Power6 was intended to be massively multicore al along.. At least according to 2 year old roadmaps. Morpherus (at AppleNova) hints that the 6 GHz target is still on schedule for Power6, and all your talk about higher clockspeeds for small Cell cores might fit perfectly with this. No? If Power6 is Cell based that is.
Ugh. Are you ignoring me, or just not getting it? If IBM has decided that the tradeoffs to increase clock rate (power, longer pipelines) aren't worth it, then they have effectively hit a "wall".
Never!
When was IBM introducing the 90nm-Process? In October 2003, maybe.
And nine month after the introduction of a new manufacturing process they hit the wall
Nr9's decleration that there will be no more Hz increases though is problematic.
I'm not defending Nr9, but it's possible and quite believable. I can imagine the situation like this:
When IBM got 2.5GHz chips and realized that they're too hot (they may still be acceptably hot for IBM, but Steve said he's not buying chips any hotter), this instantly became a limitation #1. IBM engineers then had 2 big headaches: low yields and high power dissipation. They worked on these and managed to improve yields somehow. However, better yields are money today while better power consumption is money tomorrow. They tried and tried and tried and yes, they have 3+GHz chips right now, which are either too few for a customer like Apple, or too hot for showmen like Steve, or just too unreliable in the long term. IBM can tweak the process more, of course, to squeeze a couple of MHz without sacrificing the far-from-excellent yields, but... The dual-core version got so much closer during this struggle that it's probably more economically effective to launch PPC970MP now, than to delay it until they perfect the process so that you can stick a 3.5GHz PPC970FX into a pocket PC.
This does not mean we'll never see a PowerPC running at more than 2.5GHz. This only means that we may see a dual-core before the painful transition to 90nm culminates in higher clock rates. This only means that IBM engineers may find it easier and quicker to launch a dual-core 970 than to mass-produce any PPC at, say, 3GHz with their current technology. And this does not mean IBM suddenly stops maturing the process in favour of dual-core designs, because this same process (if I'm not totally off) will help them make faster both single and dual cores until they switch to 65nm. And if they do switch to 65nm in the near future is an 'if'.
It isn't being argued that IBM have hit a wall. Clearly there is still room to grow on 90nm, but the headroom from starting a new process, to reaching the wall of the progress given the architecture of the CPU is quite clearly evident if it is mearured in terms of clock speed increase.
Chips will get faster for sure, but parallelism is the new clockspeed.
I can only clarify what I meant by "hitting a wall," because I did in fact use that phrase: I meant that what they were doing abruptly stopped yielding results. So IBM, and everyone else, is changing course to route around the wall they ran into.
Put in less dramatic terms, they squeezed all they could out of one approach, and now they're turning to another. It's happened many times before ("the Pentium will never reach 200MHz!"). It's not the end of the world.
It is funny, though, that the assumptions behind the PowerPC alliance are just now beginning to come true, years after AIM disintegrated: Power consumption finally did hit a wall; small, efficient cores finally are an advantage; an inherently 64-bit platform finally has paid off; and dual cores, which Motorola was investigating when the G4 was still a design candidate codenamed "Max," have finally become economical (two decades after speculation about the limits of single-core designs spawned research into actor languages!).
Well, you know, better late than never. Speaking as a software guy, I think it'll be a healthy switch overall. It'll force a lot of uuuuuugly legacy C and C++ code to get properly factored and cleaned up and threaded, and it'll attach a limited but significant material advantage to proper design.
It is one thing to be late to the dance, it is another thing to come with new information. I would hope that we as a group here would continue to be receptive to information from any number of sources related to things like the 970MP. There has to be more information floating about with respect to that chip that I would like to hear about.
Agreed, I just don't see anything new yet.
Quote:
Nr9's decleration that there will be no more Hz increases though is problematic. I'm not sure he understands how this statement so completely underminds everything else he states. Maybe the first release of the MP won't be that much faster than current devices, but that does not mean that the end of future speed increases is here. This may be where Nr9 is confused.
Dave [/B]
Well in answer to this and Nr9, Morpheus (over at AppleNova) posted this morning:
Quote:
"... and shouldn't be long till we see Altair-based 2.7Ghz PMs...
Morpheus"
He started the thread off a few weeks ago with [my bold]:
Quote:
"In the begining there was the chip known as GPUL aka PPC970. Next the GPUL 10S aka Altair aka PPC970FX. Next? An****s aka PPC9****!"
Nick from Thinksecret vouches for him as the real deal, as do a few others who know his posting. He has consistently painted a conservative but progressing picture and was the one who broke the 970MP (Antares) openly and widely into the rumor wild. That first post pre-dates everything else on the MP I have seen firsthand or seen discussed. So far he has stood up to 2 1/2 months of scrutiny with no dents in the proverbial armor.
We have been hearing of the "Wall" for decades in just about every engineering endeavor man has undertaken, and I can't yet think of a single instance where this wall was proven just to be anything other than the next solved engineering challenge. To think we hit the wall 6 months ago while industry still has plans to go 45nm and smaller over the next 2+ process improvements is just too Chicken Little to me. Especially in light of the many technical opinions put forth otherwise, which do accept difficulties, but not outright complete and total future failure to move forward. AND information from vouched for sources that directly contradict dire predictions from an already failed predictor.
A quote from the article which was written September 8th(over a month ago)
Quote:
"They have been reported to be qualifying low-k during the last few months, and we are expecting to see low-k product soon. We want to compare it with the other low-k processes that we have analyzed."
Granted, the "low-k product" Mr. Dick James, Chipworks' senior technology analyst, was refering to may be the dual core G5 MP, but my bet is that it isn't. The current 970Fx using a more aggressive low-k dialectric would be a drop in replacement with the same pin layout. Same motherboards, same controllers etc. Heck, Apple/IBM may start using them in the current line up with NO CHANGES in speed until a full ramp up would allow them to announce new G5's @ higher speeds and possibly a laptop sometime near the end of the year/beginning of next year. Anyway, that's what I would do.
It'll force a lot of uuuuuugly legacy C and C++ code to get properly factored and cleaned up and threaded, and it'll attach a limited but significant material advantage to proper design
Nah, applying the super new bodges is much easier ! (vectorisation! gcc 4! Converting XML to (sensible, object-like) C structures then to legacy C structures we can't edit!))
Comments
Originally posted by Programmer
Yes, but those that can do and those that can't, teach. Plus teacher's pay sucks.
Mozart taught! He wasn't paid very much however.
Originally posted by costique
No, there is no exact number to qualify as an absolute limit. What we're talking about is that crazy overclockers can make PPC970fx run at, say, 5GHz right now, but only in a can of liquid Nitrogen. Since IBM won't even consider shipping insanely overclocked chips, it roughly equals IBM having stuck at 2.5GHz. They may reach 3GHz or so for mass production of PPC970fx, but they are still having problems producing them even at 2.5GHz in quantities sufficient for Apple. You know what I mean: if those 3GHz chips turn out to be too hot, too few or too expensive, it will be just as though IBM haven't made a single chip running at 3GHz. Nobody gives a damn about a theoretical clock speed of a given chip - it only matters what clock speed the chip is running at on your desktop.
Yes, but the point is if Intel can get to 4 GHZ, IBM can as well. It might take a longer pipeline, but it can be done. We've not hit a wall just yet.
There is no doubt that IBM could take this approach but there is a big question related to benefits. Intel's approach leaves the processor with a large amount of idle time as it tries to keep those pipelines filled. The thing is that Intel does make out real good when the right types of instruction streams come along, a 4 GHz chip can be twice as fast as a 2 GHz chip under ideal conditions.
The Approach on the PPC is for much shorter pipelines and reduced penalties because of the shorter pipelines. With respect to the 970FX at 2.5GHz this has really worked out well, with many apps seeing the clock rate increase. What one ends up with, with IBM's current tech, is a very hot chip. We really don't know what the top frequency is for this chip as it does seem to be thermally limited, we don't know how fast the logic is capable of operating. That is why I think it is very important to see where IBM goes with reductions in thermal demand and also see where the rest of the industry goes. There is information floating about that does seem to indicate that 90nm done right can pay off and all extending the clock rate of a 130nm design.
What I find interesting though is that my account has been offline two days and this discussion is still going on. I'm also surprised that so many still want to throw in the towel. This is barely round one, it time to get up off the matt and breath some fresh air.
If we want to give Nr9 credit for anything he may be on to the 970MP release. It may very well come out at the same clock rate as the current 970FX's, but that may be related to a number of things. The number one thing problly being the doubling of thermal load. May the number one issue being shared with the thought that IBM still doesn't have an alternative process.
Dave
Originally posted by bunge
Yes, but the point is if Intel can get to 4 GHZ, IBM can as well. It might take a longer pipeline, but it can be done. We've not hit a wall just yet.
Originally posted by Nr9
<snip> your probably just a 5 year old who pays forty bucks a month to get access to the ACM portal and also fail to take note of anything interesting in there. if you are in SIGARCH and haven't noticed the shift towards power consumption as a priority, the forty bucks you pay are probably not worth it. <snip>
I would think you could at least get the price right if you were a member. You can't even get a student membership for $40 and SIGARCH is another $28++ on top of that. Or start with $198 for a full access including the Portal if you're a professional member (plus the $28++ for SIGARCH), but the $40 figure crops up nowhere. The facts keep unravelling, and the rhetoric more vile. The usual for a vain salvage attempt.
I second the lockination here. Maybe a fresh and unsullied thread could take it's place if folks still wanted to hash CPU issues out, but free of this threads baggage.
Originally posted by wizard69
<snip>If we want to give Nr9 credit for anything he may be on to the 970MP release. It may very well come out at the same clock rate as the current 970FX's, but that may be related to a number of things. The number one thing problly being the doubling of thermal load. May the number one issue being shared with the thought that IBM still doesn't have an alternative process.
Dave
If that hadn't already been discussed thoroughly in the wild since late July he might have had a chance, but two months late to the dance don't get credit for original thought, let alone justify a declaration of no more Hz--ever--for anyone.
Originally posted by Programmer
Yes, but those that can do and those that can't, teach.
Originally posted by AirSluf
...but two months late to the dance don't get credit for original thought, let alone justify a declaration of no more Hz--ever--for anyone.
It is one thing to be late to the dance, it is another thing to come with new information. I would hope that we as a group here would continue to be receptive to information from any number of sources related to things like the 970MP. There has to be more information floating about with respect to that chip that I would like to hear about.
Nr9's decleration that there will be no more Hz increases though is problematic. I'm not sure he understands how this statement so completely underminds everything else he states. Maybe the first release of the MP won't be that much faster than current devices, but that does not mean that the end of future speed increases is here. This may be where Nr9 is confused.
Dave
Originally posted by wilco
Yes, but those that can do and those that can't, teach.
What, have you guys never heard that quote before? I thought it was funny...
Originally posted by bunge
Yes, but the point is if Intel can get to 4 GHZ, IBM can as well. It might take a longer pipeline, but it can be done. We've not hit a wall just yet.
Ugh. Are you ignoring me, or just not getting it? If IBM has decided that the tradeoffs to increase clock rate (power, longer pipelines) aren't worth it, then they have effectively hit a "wall". It is a fuzzy and vaguely defined one, but it is the outer envelope of the processor performance space made possible by the available technology... and the technologies that they have on their roadmap. We won't really know if Nr9 is right until they revise their roadmap and the POWER6 is no longer a 5-6 GHz monster, but is instead a massively multi-core one.
Originally posted by Programmer
We won't really know if Nr9 is right until they revise their roadmap and the POWER6 is no longer a 5-6 GHz monster, but is instead a massively multi-core one.
The Power6 was intended to be massively multicore al along.. At least according to 2 year old roadmaps. Morpherus (at AppleNova) hints that the 6 GHz target is still on schedule for Power6, and all your talk about higher clockspeeds for small Cell cores might fit perfectly with this. No? If Power6 is Cell based that is.
Originally posted by Programmer
Ugh. Are you ignoring me, or just not getting it? If IBM has decided that the tradeoffs to increase clock rate (power, longer pipelines) aren't worth it, then they have effectively hit a "wall".
Never!
When was IBM introducing the 90nm-Process? In October 2003, maybe.
And nine month after the introduction of a new manufacturing process they hit the wall
I don't think so....
Originally posted by wizard69
Nr9's decleration that there will be no more Hz increases though is problematic.
I'm not defending Nr9, but it's possible and quite believable. I can imagine the situation like this:
When IBM got 2.5GHz chips and realized that they're too hot (they may still be acceptably hot for IBM, but Steve said he's not buying chips any hotter), this instantly became a limitation #1. IBM engineers then had 2 big headaches: low yields and high power dissipation. They worked on these and managed to improve yields somehow. However, better yields are money today while better power consumption is money tomorrow. They tried and tried and tried and yes, they have 3+GHz chips right now, which are either too few for a customer like Apple, or too hot for showmen like Steve, or just too unreliable in the long term. IBM can tweak the process more, of course, to squeeze a couple of MHz without sacrificing the far-from-excellent yields, but... The dual-core version got so much closer during this struggle that it's probably more economically effective to launch PPC970MP now, than to delay it until they perfect the process so that you can stick a 3.5GHz PPC970FX into a pocket PC.
This does not mean we'll never see a PowerPC running at more than 2.5GHz. This only means that we may see a dual-core before the painful transition to 90nm culminates in higher clock rates. This only means that IBM engineers may find it easier and quicker to launch a dual-core 970 than to mass-produce any PPC at, say, 3GHz with their current technology. And this does not mean IBM suddenly stops maturing the process in favour of dual-core designs, because this same process (if I'm not totally off) will help them make faster both single and dual cores until they switch to 65nm. And if they do switch to 65nm in the near future is an 'if'.
What do you think?
Chips will get faster for sure, but parallelism is the new clockspeed.
Originally posted by MarcUK
It isn't being argued that IBM have hit a wall. Clearly there is still room to grow on 90nm,
That's exactly what's being argued and it's not clearly at all that there's room to grow on 90 nm. Not to some at least.
Put in less dramatic terms, they squeezed all they could out of one approach, and now they're turning to another. It's happened many times before ("the Pentium will never reach 200MHz!"). It's not the end of the world.
It is funny, though, that the assumptions behind the PowerPC alliance are just now beginning to come true, years after AIM disintegrated: Power consumption finally did hit a wall; small, efficient cores finally are an advantage; an inherently 64-bit platform finally has paid off; and dual cores, which Motorola was investigating when the G4 was still a design candidate codenamed "Max," have finally become economical (two decades after speculation about the limits of single-core designs spawned research into actor languages!).
Well, you know, better late than never.
Originally posted by wizard69
It is one thing to be late to the dance, it is another thing to come with new information. I would hope that we as a group here would continue to be receptive to information from any number of sources related to things like the 970MP. There has to be more information floating about with respect to that chip that I would like to hear about.
Agreed, I just don't see anything new yet.
Nr9's decleration that there will be no more Hz increases though is problematic. I'm not sure he understands how this statement so completely underminds everything else he states. Maybe the first release of the MP won't be that much faster than current devices, but that does not mean that the end of future speed increases is here. This may be where Nr9 is confused.
Dave [/B]
Well in answer to this and Nr9, Morpheus (over at AppleNova) posted this morning:
"... and shouldn't be long till we see Altair-based 2.7Ghz PMs...
Morpheus"
He started the thread off a few weeks ago with [my bold]:
"In the begining there was the chip known as GPUL aka PPC970. Next the GPUL 10S aka Altair aka PPC970FX. Next? An****s aka PPC9****!"
Nick from Thinksecret vouches for him as the real deal, as do a few others who know his posting. He has consistently painted a conservative but progressing picture and was the one who broke the 970MP (Antares) openly and widely into the rumor wild. That first post pre-dates everything else on the MP I have seen firsthand or seen discussed. So far he has stood up to 2 1/2 months of scrutiny with no dents in the proverbial armor.
We have been hearing of the "Wall" for decades in just about every engineering endeavor man has undertaken, and I can't yet think of a single instance where this wall was proven just to be anything other than the next solved engineering challenge. To think we hit the wall 6 months ago while industry still has plans to go 45nm and smaller over the next 2+ process improvements is just too Chicken Little to me. Especially in light of the many technical opinions put forth otherwise, which do accept difficulties, but not outright complete and total future failure to move forward. AND information from vouched for sources that directly contradict dire predictions from an already failed predictor.
Originally posted by snoopy
That may be sticking your neck out, unless you have inside information. There is some evidence that IBM has a dual core 970MP in the works.
I couldn't remember the link, however, THT graciously supplied them in the New Powerbook thread.
CHIPWORKS DISCOVERS IBM STRAIN AT 90NM, DELAY IN LOW-K
A quote from the article which was written September 8th(over a month ago)
"They have been reported to be qualifying low-k during the last few months, and we are expecting to see low-k product soon. We want to compare it with the other low-k processes that we have analyzed."
Granted, the "low-k product" Mr. Dick James, Chipworks' senior technology analyst, was refering to may be the dual core G5 MP, but my bet is that it isn't. The current 970Fx using a more aggressive low-k dialectric would be a drop in replacement with the same pin layout. Same motherboards, same controllers etc. Heck, Apple/IBM may start using them in the current line up with NO CHANGES in speed until a full ramp up would allow them to announce new G5's @ higher speeds and possibly a laptop sometime near the end of the year/beginning of next year. Anyway, that's what I would do.
It'll force a lot of uuuuuugly legacy C and C++ code to get properly factored and cleaned up and threaded, and it'll attach a limited but significant material advantage to proper design
Nah, applying the super new bodges is much easier ! (vectorisation! gcc 4! Converting XML to (sensible, object-like) C structures then to legacy C structures we can't edit!))
Or how about C to OO?
C++ to XML?
How you want it?