people won't try to squeeze as much performance out of a piece of hardware, they will try to thread their apps as much as possible using modern programming paradigms and they will see an increase in application performance linear to the # of cores.
I would imagine that single threaded general purpose performance won't increase for the next few years in any processor architecture.
if you require really good performance for a specific algorithm, it would be a better bet to have computers would come with FPGAs, or you would have FPGA PCI cards, and high performance applications will load programming files for critical algorithms onto these FPGAs to run them through custom logic.
One thing to note on the 3GHZ thing - IBM can currently build 970FX chips that hit 3GHZ. They couldn't build them en masse when Apple needed them for an update and Apple had concerns about the dissipation. Apple is focused on making quiet machines. The new dual-2.5GHZ machines are very quiet. 3GHZ 970FX chips are stable, but too hot for Apple to have an interest in them - at least at the moment. Because there is no market for IBM to put them in, they are not manufacturing them (though Apple does have some).
The next generation should scale past 3GHZ, even with multiple cores. IBM has gotten many of the power leakage issues under control at 90nm and will attempt to go to 65nm at some point next year - though the focus will no longer be on scalability, but rather on addition of cache and cores. The new 970s derived from the POWER5 have been specifically engineered for this and should face fewer problems in transition.
Apple may yet receive 3GHZ 970FX chips with PowerTune if they can find a way to keep them cool while also keeping the PowerMacs fairly quiet.
Also of note on the IBM campus, there are lots more Apple engineers and QA people on the Fishkill campus - and they've been working a lot of OT developing mobile chips and nextgen 970s. The increase started over three months ago - so it may confirm that Apple is abandoning Motorola.
If all goes as planned, even with a few hiccups, there would be enough G5 variations available by next summer to have all Apple products running on G5s. This includes iBooks using slower G5s than the PowerBooks (though the iBooks would still be faster than they are today).
they go up to 130 celsius with aggressive water cooling
they just melt down. your processor life is probably measured in days.
it was the same problem with the canned motorola G5s
No they do not. I know engineers up at Fishkill who work on this and that is not true. It was six months ago, but it is not the case as of my last contact with them more recently.
3GHZ 970FX chips are stable and their dissipation is under 75W. I will not release the exact numbers because it is not my place, but the 3GHZ chips do not melt down nor do they run at 130 celsius when aggressively cooled. They don't even run at 130 Fahrenheit when properly cooled. The biggest issue with Apple is noise.
Let me know where you get your info from. I get mine directly from the people who make them.
I don't get the thing about dual core designs solving power consumption problems with high clock rate processors. TANSTAAFL. A dual core processor doubles the transistor count of a processor. It will have higher power requirements. Like almost double the power requirements of a single core version at the same clock rate. It is a tradeoff like all things are.
The upcoming 65 nm and 50/45 nm nodes will present incredible challenges undoubtedly, even moreso than 90 nm, but they will eventually come. Maybe not Moore's "law" quick, but they'll be coming. Other areas where the market is behind in is power management and cooling technologies. Thermoelectric materials could be developed to cool chips at the transistor, die level, and to conserve power at the same time. Manufacturers may have to bite the bullet and move to phase-change cooling systems to cool higher clock rate systems.
I don't get the thing about dual core designs solving power consumption problems with high clock rate processors. TANSTAAFL. A dual core processor doubles the transistor count of a processor. It will have higher power requirements. Like almost double the power requirements of a single core version at the same clock rate. It is a tradeoff like all things are.
The upcoming 65 nm and 50/45 nm nodes will present incredible challenges undoubtedly, even moreso than 90 nm, but they will eventually come. Maybe not Moore's "law" quick, but they'll be coming. Other areas where the market is behind in is power management and cooling technologies. Thermoelectric materials could be developed to cool chips at the transistor, die level, and to conserve power at the same time. Manufacturers may have to bite the bullet and move to phase-change cooling systems to cool higher clock rate systems.
it solves it because it is twice the area hence easier to cool. Also, you don't have to drive the voltage up(which you have to do if you want to up the frequency instead) power is proportional to voltage ^2 so that matters a lot
you are wrong about faster processors coming. ask ibm, intel and amd
it solves it because it is twice the area hence easier to cool. Also, you don't have to drive the voltage up(which you have to do if you want to up the frequency instead) power is proportional to voltage ^2 so that matters a lot . . .
No and yes. If a chip has twice the area it is easier to cool only if in has the same circuitry spread over that area, so the heat density is less. However, when it also has twice the circuitry the power is double. The heat density is not changed and there is twice the heat to dissipate.
You are correct about increasing voltage to achieve higher clock frequency. Power increases rapidly, not only because of the voltage squared relationship, but due to leakage current. The leakage current is in fact the principal problem at 90nm I believe.
No and yes. If a chip has twice the area it is easier to cool only if in has the same circuitry spread over that area, so the heat density is less. However, when it also has twice the circuitry the power is double. The heat density is not changed and there is twice the heat to dissipate.
You are correct about increasing voltage to achieve higher clock frequency. Power increases rapidly, not only because of the voltage squared relationship, but due to leakage current. The leakage current is in fact the principal problem at 90nm I believe.
eh, of course dual cores are going to be more power than single cores
but dual core are going to be a lot less power and a lot easier to cool than single cores that are running twice as fast. That's the motivation here. Single core CPUs already have heat densities reaching the limit.
you are wrong about faster processors coming. ask ibm, intel and amd
Actually, they really don't know. Based on current technology and knowledge, it seems like the clockrate wall is just about hit - but there are theoretical increases out there based on developing technology.
IBM engineers are publicly going to be very skeptical - everyone is. Behind closed doors, they are still working to drive clockrates up and ultimately find a balance. As I said, the 3GHZ 970FX does exist. The 130nm 3GHZ original 970 exists as well, though it is superhot and not plausible as a desktop chip. It was pushed to that limit by IBM engineers to test the chip. They tried on the 970 to go higher, but it did result in flame outs.
The 970FX did originally flame out at 3GHZ due to powerloss much higher than anticipated at the 90nm fab. However, IBM was dealing with new technology that they still had to refine. They have worked for nearly a year on the problems and have made quite a bit of headway. Their yields are improving and they've reduced power leakage by quite a bit, enabling the 3GHZ to get under control. Early in the run even the 2.5GHZ had problems and this affected the whole lineup. Problems with the fab became apparent and IBM engineers had to put in a lot of overtime, but things are calming down. They are now working quite a bit on anticipated problems at 65nm - something they are now committing more resources to due to the problems they encountered during the 90nm transition.
you seem to be making this up. 2.5GHz is the limit.
Quote:
Originally posted by Mr. MacPhisto
Actually, they really don't know. Based on current technology and knowledge, it seems like the clockrate wall is just about hit - but there are theoretical increases out there based on developing technology.
IBM engineers are publicly going to be very skeptical - everyone is. Behind closed doors, they are still working to drive clockrates up and ultimately find a balance. As I said, the 3GHZ 970FX does exist. The 130nm 3GHZ original 970 exists as well, though it is superhot and not plausible as a desktop chip. It was pushed to that limit by IBM engineers to test the chip. They tried on the 970 to go higher, but it did result in flame outs.
The 970FX did originally flame out at 3GHZ due to powerloss much higher than anticipated at the 90nm fab. However, IBM was dealing with new technology that they still had to refine. They have worked for nearly a year on the problems and have made quite a bit of headway. Their yields are improving and they've reduced power leakage by quite a bit, enabling the 3GHZ to get under control. Early in the run even the 2.5GHZ had problems and this affected the whole lineup. Problems with the fab became apparent and IBM engineers had to put in a lot of overtime, but things are calming down. They are now working quite a bit on anticipated problems at 65nm - something they are now committing more resources to due to the problems they encountered during the 90nm transition.
It doesn't really matter if the 3 GHz 970 runs or not -- the point is that clock rate scaling has reached its limit, at least for the currently understood technologies. Its possible something will come out of the blue to allow additional jumps, but right now it looks like increasing clock rate is too inefficient in terms of performance gain per unit of power consumed and the simple destructive effects of really high operating temperatures.
Adding transistors increases power consumption linearly and increases area at the same time. Increasing clock rate doesn't increase the area, and it increases power consumption in a super-linear fashion. Also, to accomplish higher clock rates pipelines have been getting longer which has resulted in significantly less than linear performance increases relative to clock rate. This means that performance has been increasing very slowly compared to power consumption, which is not a sustainable situation. The writing has been on the wall for years -- the future is multi-core, massively parallel, highly superscalar, and SIMD. This is what Cell is all about. Look at Sun's new SPARCs. It is also the direction GPUs have been going.
Get used to it. If you're a software developer, its time to change how you write software.
It'll be interesting to see the next IBM POWER roadmap to see if Nr9 is right about their POWER6 plans.
From what I've been told, all dual-core designs for desktops are POWER5 based. The only dual-cores based off the current design (with modifications) are the "mobile solution" versions that we'd see in PowerBooks.
The POWER5 derivatives were meant for the end of this year, but with all the delays IBM has seen the latest estimates would put their availability sometime near summer.
Mr. MacPhisto,
Not many weeks ago the table of contents for a 970MP was circulating around, a dual core 970FX type chip from what I could tell. How does this fit with your information? Would this be the dual core PowerBook chip you speak of? Most folks believe it is for the next Power Mac upgrade, getting to a quad with two such chips. So far, your posts on IBM chips seem to be the most trustworthy. Are you able to shed any light on this 970MP thing?
Not many weeks ago the table of contents for a 970MP was circulating around, a dual core 970FX type chip from what I could tell. How does this fit with your information? Would this be the dual core PowerBook chip you speak of? Most folks believe it is for the next Power Mac upgrade, getting to a quad with two such chips. So far, your posts on IBM chips seem to be the most trustworthy. Are you able to shed any light on this 970MP thing?
The way I understand it, there are two multi-core chip designs underway. There is one based off the low-voltage 970 chip that is being specifically worked on for mobile use. There are some design enhancements over the standard desktop 970FX. I believe the major change is L2 increase. I don't believe there is significant architectural overhaul on the chip, but it is slated for mobile use, as stated. We'll first see a low-voltage single core, though a dual-core version may be on the market sometime during the second half of next year, barring any major hitches (and also depending largely on the 65nm shrink and how successful it is - I don't believe the mobile multi-core design is planned at 90nm). This is based off the current 970FX. I do not believe it is the 970MP and I think the 970MP designation is external from IBM. The Antares project, however, is POWER5 based and really a new generation. This would be the chip that Thinksecret reported on. I personally suspect we will see it in January, as ThinkSecret reported. Information I have received on the project indicates that each core will also have multi-threading capability. There are other refinements, but I'm not a chip guru nor do I want to put too much out there.
The mobile 970 is based off the low-voltage FX, as stated, and will come after the desktop multi-core chips. IBM is also working with Apple on mobile versions of Antares, though neither party is sure when those will be ready or what they will be like. The good news is that both parties are focusing on portable specific chips that are more L2 heavy. Have not asked if the pipeline length is the same as the 970FX, though I assume it is.
The Following is my opinion based on what I know:
I believe Apple will go 100% G5 come January, though there is a possibility the iBook will keep the G4 a bit longer, becoming the last to jump just as it was with the G3. I think a mobile G5 solution will be ready in January and that it will be placed in the PowerBooks. Lower clocked G5s could find their way into te iBooks coupled with slower memory, lesser GPU, etc.
The iMac is already G5, though I think we will see the new iMac sport dual-core G5s with the PowerMac have dual CPUs featuring the dual-core.
I think the eMac may get one more revision as a G4, though it may get discontinued altogether. We may finally see a box. I would favor the same configs as the iMac G5, minus the display cost (only a few hundred bucks). More expandability from the box, but you end up paying more if you want a new display and you miss out on the elegance. The new iMac actually gives some real advantage to the AIO because of minimal clutter.
If the iBook keeps the G4 (going up to 1.5GHZ maybe?) then I see it getting the G5 if the dual-core mobile version arrive by late summer, early fall (about a year from now).
Granted, there are still some "ifs". If IBM can deliver the desktop dual-core G5s for January shipment in quantity then Apple would experience a strong performance leap and jump to the head of the industry in chip technology, for a brief period. G5 PowerBooks would also help increase sales.
If IBM can deliver in quantity (not as big an "if" as it was 9 months ago, though still substantial in size) for the PowerBooks, iMacs, iBooks, and PowerMacs in time for a January/February release, Apple would hit record sales numbers. IBM would also hit records for its chip fabbing business. Both companies have strong incentives to do so.
Comments
it was inelegant and unportable
people won't try to squeeze as much performance out of a piece of hardware, they will try to thread their apps as much as possible using modern programming paradigms and they will see an increase in application performance linear to the # of cores.
I would imagine that single threaded general purpose performance won't increase for the next few years in any processor architecture.
if you require really good performance for a specific algorithm, it would be a better bet to have computers would come with FPGAs, or you would have FPGA PCI cards, and high performance applications will load programming files for critical algorithms onto these FPGAs to run them through custom logic.
The next generation should scale past 3GHZ, even with multiple cores. IBM has gotten many of the power leakage issues under control at 90nm and will attempt to go to 65nm at some point next year - though the focus will no longer be on scalability, but rather on addition of cache and cores. The new 970s derived from the POWER5 have been specifically engineered for this and should face fewer problems in transition.
Apple may yet receive 3GHZ 970FX chips with PowerTune if they can find a way to keep them cool while also keeping the PowerMacs fairly quiet.
Also of note on the IBM campus, there are lots more Apple engineers and QA people on the Fishkill campus - and they've been working a lot of OT developing mobile chips and nextgen 970s. The increase started over three months ago - so it may confirm that Apple is abandoning Motorola.
If all goes as planned, even with a few hiccups, there would be enough G5 variations available by next summer to have all Apple products running on G5s. This includes iBooks using slower G5s than the PowerBooks (though the iBooks would still be faster than they are today).
they go up to 130 celsius with aggressive water cooling
they just melt down. your processor life is probably measured in days.
it was the same problem with the canned motorola G5s
Originally posted by Nr9
the 3GHz chips can't run at all
they go up to 130 celsius with aggressive water cooling
they just melt down. your processor life is probably measured in days.
it was the same problem with the canned motorola G5s
No they do not. I know engineers up at Fishkill who work on this and that is not true. It was six months ago, but it is not the case as of my last contact with them more recently.
3GHZ 970FX chips are stable and their dissipation is under 75W. I will not release the exact numbers because it is not my place, but the 3GHZ chips do not melt down nor do they run at 130 celsius when aggressively cooled. They don't even run at 130 Fahrenheit when properly cooled. The biggest issue with Apple is noise.
Let me know where you get your info from. I get mine directly from the people who make them.
Originally posted by Nr9
i make them
You make them up?
Originally posted by tadunne
You make them up?
Yes. Then he places it in his crack pipe and smokes it.
Originally posted by Nr9
i make them
I am a name not a number. If you are number 9 who is number 1?
The upcoming 65 nm and 50/45 nm nodes will present incredible challenges undoubtedly, even moreso than 90 nm, but they will eventually come. Maybe not Moore's "law" quick, but they'll be coming. Other areas where the market is behind in is power management and cooling technologies. Thermoelectric materials could be developed to cool chips at the transistor, die level, and to conserve power at the same time. Manufacturers may have to bite the bullet and move to phase-change cooling systems to cool higher clock rate systems.
Originally posted by THT
I don't get the thing about dual core designs solving power consumption problems with high clock rate processors. TANSTAAFL. A dual core processor doubles the transistor count of a processor. It will have higher power requirements. Like almost double the power requirements of a single core version at the same clock rate. It is a tradeoff like all things are.
The upcoming 65 nm and 50/45 nm nodes will present incredible challenges undoubtedly, even moreso than 90 nm, but they will eventually come. Maybe not Moore's "law" quick, but they'll be coming. Other areas where the market is behind in is power management and cooling technologies. Thermoelectric materials could be developed to cool chips at the transistor, die level, and to conserve power at the same time. Manufacturers may have to bite the bullet and move to phase-change cooling systems to cool higher clock rate systems.
it solves it because it is twice the area hence easier to cool. Also, you don't have to drive the voltage up(which you have to do if you want to up the frequency instead) power is proportional to voltage ^2 so that matters a lot
you are wrong about faster processors coming. ask ibm, intel and amd
Originally posted by Nr9
it solves it because it is twice the area hence easier to cool. Also, you don't have to drive the voltage up(which you have to do if you want to up the frequency instead) power is proportional to voltage ^2 so that matters a lot . . .
No and yes. If a chip has twice the area it is easier to cool only if in has the same circuitry spread over that area, so the heat density is less. However, when it also has twice the circuitry the power is double. The heat density is not changed and there is twice the heat to dissipate.
You are correct about increasing voltage to achieve higher clock frequency. Power increases rapidly, not only because of the voltage squared relationship, but due to leakage current. The leakage current is in fact the principal problem at 90nm I believe.
Originally posted by snoopy
No and yes. If a chip has twice the area it is easier to cool only if in has the same circuitry spread over that area, so the heat density is less. However, when it also has twice the circuitry the power is double. The heat density is not changed and there is twice the heat to dissipate.
You are correct about increasing voltage to achieve higher clock frequency. Power increases rapidly, not only because of the voltage squared relationship, but due to leakage current. The leakage current is in fact the principal problem at 90nm I believe.
eh, of course dual cores are going to be more power than single cores
but dual core are going to be a lot less power and a lot easier to cool than single cores that are running twice as fast. That's the motivation here. Single core CPUs already have heat densities reaching the limit.
Originally posted by Nr9
you are wrong about faster processors coming. ask ibm, intel and amd
Actually, they really don't know. Based on current technology and knowledge, it seems like the clockrate wall is just about hit - but there are theoretical increases out there based on developing technology.
IBM engineers are publicly going to be very skeptical - everyone is. Behind closed doors, they are still working to drive clockrates up and ultimately find a balance. As I said, the 3GHZ 970FX does exist. The 130nm 3GHZ original 970 exists as well, though it is superhot and not plausible as a desktop chip. It was pushed to that limit by IBM engineers to test the chip. They tried on the 970 to go higher, but it did result in flame outs.
The 970FX did originally flame out at 3GHZ due to powerloss much higher than anticipated at the 90nm fab. However, IBM was dealing with new technology that they still had to refine. They have worked for nearly a year on the problems and have made quite a bit of headway. Their yields are improving and they've reduced power leakage by quite a bit, enabling the 3GHZ to get under control. Early in the run even the 2.5GHZ had problems and this affected the whole lineup. Problems with the fab became apparent and IBM engineers had to put in a lot of overtime, but things are calming down. They are now working quite a bit on anticipated problems at 65nm - something they are now committing more resources to due to the problems they encountered during the 90nm transition.
Originally posted by Mr. MacPhisto
Actually, they really don't know. Based on current technology and knowledge, it seems like the clockrate wall is just about hit - but there are theoretical increases out there based on developing technology.
IBM engineers are publicly going to be very skeptical - everyone is. Behind closed doors, they are still working to drive clockrates up and ultimately find a balance. As I said, the 3GHZ 970FX does exist. The 130nm 3GHZ original 970 exists as well, though it is superhot and not plausible as a desktop chip. It was pushed to that limit by IBM engineers to test the chip. They tried on the 970 to go higher, but it did result in flame outs.
The 970FX did originally flame out at 3GHZ due to powerloss much higher than anticipated at the 90nm fab. However, IBM was dealing with new technology that they still had to refine. They have worked for nearly a year on the problems and have made quite a bit of headway. Their yields are improving and they've reduced power leakage by quite a bit, enabling the 3GHZ to get under control. Early in the run even the 2.5GHZ had problems and this affected the whole lineup. Problems with the fab became apparent and IBM engineers had to put in a lot of overtime, but things are calming down. They are now working quite a bit on anticipated problems at 65nm - something they are now committing more resources to due to the problems they encountered during the 90nm transition.
Originally posted by Nr9
i make them
You're 19 years old and working at IBM? Er... Not.
Adding transistors increases power consumption linearly and increases area at the same time. Increasing clock rate doesn't increase the area, and it increases power consumption in a super-linear fashion. Also, to accomplish higher clock rates pipelines have been getting longer which has resulted in significantly less than linear performance increases relative to clock rate. This means that performance has been increasing very slowly compared to power consumption, which is not a sustainable situation. The writing has been on the wall for years -- the future is multi-core, massively parallel, highly superscalar, and SIMD. This is what Cell is all about. Look at Sun's new SPARCs. It is also the direction GPUs have been going.
Get used to it. If you're a software developer, its time to change how you write software.
It'll be interesting to see the next IBM POWER roadmap to see if Nr9 is right about their POWER6 plans.
Originally posted by Mr. MacPhisto
From what I've been told, all dual-core designs for desktops are POWER5 based. The only dual-cores based off the current design (with modifications) are the "mobile solution" versions that we'd see in PowerBooks.
The POWER5 derivatives were meant for the end of this year, but with all the delays IBM has seen the latest estimates would put their availability sometime near summer.
Mr. MacPhisto,
Not many weeks ago the table of contents for a 970MP was circulating around, a dual core 970FX type chip from what I could tell. How does this fit with your information? Would this be the dual core PowerBook chip you speak of? Most folks believe it is for the next Power Mac upgrade, getting to a quad with two such chips. So far, your posts on IBM chips seem to be the most trustworthy. Are you able to shed any light on this 970MP thing?
Originally posted by snoopy
Mr. MacPhisto,
Not many weeks ago the table of contents for a 970MP was circulating around, a dual core 970FX type chip from what I could tell. How does this fit with your information? Would this be the dual core PowerBook chip you speak of? Most folks believe it is for the next Power Mac upgrade, getting to a quad with two such chips. So far, your posts on IBM chips seem to be the most trustworthy. Are you able to shed any light on this 970MP thing?
The way I understand it, there are two multi-core chip designs underway. There is one based off the low-voltage 970 chip that is being specifically worked on for mobile use. There are some design enhancements over the standard desktop 970FX. I believe the major change is L2 increase. I don't believe there is significant architectural overhaul on the chip, but it is slated for mobile use, as stated. We'll first see a low-voltage single core, though a dual-core version may be on the market sometime during the second half of next year, barring any major hitches (and also depending largely on the 65nm shrink and how successful it is - I don't believe the mobile multi-core design is planned at 90nm). This is based off the current 970FX. I do not believe it is the 970MP and I think the 970MP designation is external from IBM. The Antares project, however, is POWER5 based and really a new generation. This would be the chip that Thinksecret reported on. I personally suspect we will see it in January, as ThinkSecret reported. Information I have received on the project indicates that each core will also have multi-threading capability. There are other refinements, but I'm not a chip guru nor do I want to put too much out there.
The mobile 970 is based off the low-voltage FX, as stated, and will come after the desktop multi-core chips. IBM is also working with Apple on mobile versions of Antares, though neither party is sure when those will be ready or what they will be like. The good news is that both parties are focusing on portable specific chips that are more L2 heavy. Have not asked if the pipeline length is the same as the 970FX, though I assume it is.
The Following is my opinion based on what I know:
I believe Apple will go 100% G5 come January, though there is a possibility the iBook will keep the G4 a bit longer, becoming the last to jump just as it was with the G3. I think a mobile G5 solution will be ready in January and that it will be placed in the PowerBooks. Lower clocked G5s could find their way into te iBooks coupled with slower memory, lesser GPU, etc.
The iMac is already G5, though I think we will see the new iMac sport dual-core G5s with the PowerMac have dual CPUs featuring the dual-core.
I think the eMac may get one more revision as a G4, though it may get discontinued altogether. We may finally see a box. I would favor the same configs as the iMac G5, minus the display cost (only a few hundred bucks). More expandability from the box, but you end up paying more if you want a new display and you miss out on the elegance. The new iMac actually gives some real advantage to the AIO because of minimal clutter.
If the iBook keeps the G4 (going up to 1.5GHZ maybe?) then I see it getting the G5 if the dual-core mobile version arrive by late summer, early fall (about a year from now).
Granted, there are still some "ifs". If IBM can deliver the desktop dual-core G5s for January shipment in quantity then Apple would experience a strong performance leap and jump to the head of the industry in chip technology, for a brief period. G5 PowerBooks would also help increase sales.
If IBM can deliver in quantity (not as big an "if" as it was 9 months ago, though still substantial in size) for the PowerBooks, iMacs, iBooks, and PowerMacs in time for a January/February release, Apple would hit record sales numbers. IBM would also hit records for its chip fabbing business. Both companies have strong incentives to do so.