You can't judge one processor by another. There are many design choices and trade offs made in every processor. Plus IBM has some tricks it hasn't yet rolled out into the 970 line of processors, at the 90 nm node.
Yes, tricks... hmm.
And btw, weren't one of those who said that we've reached the end of MHz scaling. If those tricks work...
And btw, weren't one of those who said that we've reached the end of MHz scaling. If those tricks work...
Those tricks and design choices are typically trade-offs, so having them work usually costs you something else. I was saying we've reached the end of the cheap-and-easy frequency scaling that has been going on for over a decade. IBM's process tricks might give another 0.5 GHz on the 970FX, but that is a mere 20% over the current machine -- hardly remarkable, and not repeatable. More interesting is that it drops the power/heat on the current speeds.
I was saying we've reached the end of the cheap-and-easy frequency scaling that has been going on for over a decade.
To amplify this point, does anyone remember when Intel would get so much of a clock speed boost from a process improvement that they'd be able to milk it for an entire year or even two, using incremental upgrades? I'm sure Intel remembers them fondly.
AMD taking it to Intel at the high end (of the PC market) began the end of that era, and 90nm finished it with a whimper.
Now, if someone releases an update to a chip that runs 20% faster, it's because they just found a way to make the thing go 20% faster. And that's that, until they find another way.
IBM's process tricks might give another 0.5 GHz on the 970FX, but that is a mere 20% over the current machine -- hardly remarkable, and not repeatable.
This is at 90 nm? If they do as you say, the 970fx would be going from 2 GHz to 3 GHz. That would be remarkable considering everyone elses troubles at 90 nm. Intel only did a little bit better with the Pentium 4 at 130 nm.
Quote:
More interesting is that it drops the power/heat on the current speeds.
Yes, that is an interesting aspect since it would make mobile 970fx systems possible, and we all agree IBM will have to use a combination of micro-architecture design and process tweaking to get there. I'm just surprised that a Cell processor going into a small form factor box will be manufactured on a 90 nm node. A Power derived CPU with multiple "Cell" cores? Sounds like a big chip, just as hot as 970fx, and costly.
If that have a process tweak which allows them to run at less power, through lower voltage or more efficient transistors or both, it good news for both clock rate scaling and mobile systems.
On the really interesting side, the Cell CPU will be using XDR SDRAM. I'll be interested in the performance of that compared to DDR.
This is at 90 nm? If they do as you say, the 970fx would be going from 2 GHz to 3 GHz. That would be remarkable considering everyone elses troubles at 90 nm. Intel only did a little bit better with the Pentium 4 at 130 nm.
Last time I checked the 970fx was shipping at 2.5 GHz and has been for 6 months. And the transition to 65 nm isn't expected to improve the clock rates / power curves much, if at all.
Intel went from 3.15 -> 3.8 GHz on 90nm, which is about 20% as well. Of course they didn't gain anything on the first 130->90 switch, but they had tweaked out their 130nm process more than IBM had for the 970.
Quote:
A Power derived CPU with multiple "Cell" cores? Sounds like a big chip, just as hot as 970fx, and costly.
You can't always make assumptions like that. Next week should bring answers.
Last time I checked the 970fx was shipping at 2.5 GHz and has been for 6 months. And the transition to 65 nm isn't expected to improve the clock rates / power curves much, if at all.
Intel went from 3.15 -> 3.8 GHz on 90nm, which is about 20% as well. Of course they didn't gain anything on the first 130->90 switch, but they had tweaked out their 130nm process more than IBM had for the 970.
We're arguing over what is remarkable here. Do you believe IBM could have shipped a 2.5 GHz 130-nm 970? If so, I would agree that a 3 GHz 970fx is hardly remarkable. But if you don't think so, what clock could IBM have shipped that 130-nm 970 at? If it is 2 to 2.2 GHz, then a prospective 970fx clock increase to 3 GHz is pretty good scaling for all the troubles seen at 90 nm. That's proportionately better than Intel had done or has chosen to do with Prescott.
Quote:
You can't always make assumptions like that. Next week should bring answers.
The session description says Cell is a "64 bit Power architecture." Once again, it's a question of nomenclature. I'm only taking the terms as they are known to be, and Power means either Power 3, Power 4, or Power 5, and Power 5 is mostly an SMT extension of the Power 4. So, this prospective Cell is a single core Power 4 with auxillary SIMD (streaming) units onboard, I/O and an integrated memory controller fabricated at 90 nm. Sounds very much like a 970fx type of processor in terms of power consumption and die size to me.
If it is a brand new Power ISA microarchitecture, then of course, all bets are off. If it is a Power chip as we know it, than I think it is good news all around. Whether it is a process tweak or a better design for power consumption, I'm sure they can make their way into processors for Apple.
[B]We're arguing over what is remarkable here. Do you believe IBM could have shipped a 2.5 GHz 130-nm 970? If so, I would agree that a 3 GHz 970fx is hardly remarkable. But if you don't think so, what clock could IBM have shipped that 130-nm 970 at? If it is 2 to 2.2 GHz, then a prospective 970fx clock increase to 3 GHz is pretty good scaling for all the troubles seen at 90 nm. That's proportionately better than Intel had done or has chosen to do with Prescott.
I doubt a 2.5 GHz 130 nm 970 was practical to ship, especially given Apple's desire to keep its machines quiet. It may not have even functioned if there was a specific problem with the design.
IBM's 90 nm performance does look like one of the best, but remember that they took one crack at 130 nm and then switched to 90 nm. Intel did quite a bit of refinement on 130 nm, plus they were at a higher clock rate already and the relationship between power and frequency is highly non-linear. Intel has also shown a willingness to ship really hot chips. As a result its hard to compare the two. Compared to the last 10 years, however, a further 20% increase isn't particuarly noteworthy.
Quote:
The session description says Cell is a "64 bit Power architecture." Once again, it's a question of nomenclature. I'm only taking the terms as they are known to be, and Power means either Power 3, Power 4, or Power 5, and Power 5 is mostly an SMT extension of the Power 4. So, this prospective Cell is a single core Power 4 with auxillary SIMD (streaming) units onboard, I/O and an integrated memory controller fabricated at 90 nm. Sounds very much like a 970fx type of processor in terms of power consumption and die size to me.
That is what Power architecture means. Wait until next week and you'll see what I mean.
Quote:
If it is a brand new Power ISA microarchitecture, then of course, all bets are off. If it is a Power chip as we know it, than I think it is good news all around. Whether it is a process tweak or a better design for power consumption, I'm sure they can make their way into processors for Apple.
There are shades of grey between the options you outline. Processors are big and complex things, and there are a lot of options even within a single design.
Of course, Apple being Apple, if it does happen then they're going to want a veil of absolute silence over it.
So IBM might well show off a variant of the Cell architecture that has no business in a Mac, but since Cell is highly modular, all bets are off in the general case.
First of all, down the road Apple could be in a fine position to take advantage of such tech advances. Doubtful if Intel, MS, and Dell could be so nimble.
As an aside, I wonder if it would be the approach for TV's to render raw digital data on the fly. That would circumvent half of the current workflow of content creation, computer rendering, compression, and delivery.
First of all, down the road Apple could be in a fine position to take advantage of such tech advances. Doubtful if Intel, MS, and Dell could be so nimble.
As an aside, I wonder if it would be the approach for TV's to render raw digital data on the fly. That would circumvent half of the current workflow of content creation, computer rendering, compression, and delivery.
Wow indeed mugwump!
Sony and Toshiba are already positioning the Cell into their TVs
Cell would seem to be able to create mini render farms on the workstation which would, as you say, reduce by half the effort of content creation. (with perhaps game coding aside)
Rendering and compression are ideal Cell tasks and should show dramatic improvements over current practices. Imagine Core Video being processed by the Cell in everything from the user interface to FCP.
THE VISION for the Cell processor as proclaimed by IBM, Sony, and Toshiba.
"Breakthrough microprocessor architecture that puts broadband communications right on the chip."
I'm doing a bit of connect the dots here, but I have this sort of unifying theory that mugwump has started with his post.
I'll take Apple's low profile with Power.org as a further manifestation of Apple's obsessive desire for holding their cards close to their chest. To continue the analogy, Apple also has an ace up their sleeve on this one that they will play soon. Core services in Tiger are the smoking gun here for me.
The turn on a dime from all their professed views as shown with the new mini's may also be indicative of how fast Apple could embrace the forgotten spoke to their digital hub - TV and not just any old TV, HDTV!
Get ready for the Apple channel! What does Al Gore know that would cause him to spend 70 million on a cable TV channel?
Quote:
"Among those on the company's advisory board guiding the growth of INdTV is Apple CEO Steve Jobs. The company's COO Mark Goldman was formerly with News Corp. He built a satellite venture in Latin America. Goldman was the president and COO of Sky Latin America Partners prior to teaming up with Gore's venture."
Core is a important clue. Routing this stuff through the GPU is a revolutionary change in the computer/software model. Apple is doing it now. Most importantly, they are doing it without breaking all the existing apps. Wouldn't that model work well for Cell? Hand off Quicktime, OpenGL, Core and other intensive stuff. Using the power where you really need it, my friends, is the future.
Comments
Originally posted by Programmer
You can't judge one processor by another. There are many design choices and trade offs made in every processor. Plus IBM has some tricks it hasn't yet rolled out into the 970 line of processors, at the 90 nm node.
Yes, tricks... hmm.
And btw, weren't one of those who said that we've reached the end of MHz scaling. If those tricks work...
Originally posted by THT
Yes, tricks... hmm.
And btw, weren't one of those who said that we've reached the end of MHz scaling. If those tricks work...
Those tricks and design choices are typically trade-offs, so having them work usually costs you something else. I was saying we've reached the end of the cheap-and-easy frequency scaling that has been going on for over a decade. IBM's process tricks might give another 0.5 GHz on the 970FX, but that is a mere 20% over the current machine -- hardly remarkable, and not repeatable. More interesting is that it drops the power/heat on the current speeds.
Originally posted by Programmer
I was saying we've reached the end of the cheap-and-easy frequency scaling that has been going on for over a decade.
To amplify this point, does anyone remember when Intel would get so much of a clock speed boost from a process improvement that they'd be able to milk it for an entire year or even two, using incremental upgrades? I'm sure Intel remembers them fondly.
AMD taking it to Intel at the high end (of the PC market) began the end of that era, and 90nm finished it with a whimper.
Now, if someone releases an update to a chip that runs 20% faster, it's because they just found a way to make the thing go 20% faster. And that's that, until they find another way.
Originally posted by Programmer
IBM's process tricks might give another 0.5 GHz on the 970FX, but that is a mere 20% over the current machine -- hardly remarkable, and not repeatable.
This is at 90 nm? If they do as you say, the 970fx would be going from 2 GHz to 3 GHz. That would be remarkable considering everyone elses troubles at 90 nm. Intel only did a little bit better with the Pentium 4 at 130 nm.
More interesting is that it drops the power/heat on the current speeds.
Yes, that is an interesting aspect since it would make mobile 970fx systems possible, and we all agree IBM will have to use a combination of micro-architecture design and process tweaking to get there. I'm just surprised that a Cell processor going into a small form factor box will be manufactured on a 90 nm node. A Power derived CPU with multiple "Cell" cores? Sounds like a big chip, just as hot as 970fx, and costly.
If that have a process tweak which allows them to run at less power, through lower voltage or more efficient transistors or both, it good news for both clock rate scaling and mobile systems.
On the really interesting side, the Cell CPU will be using XDR SDRAM. I'll be interested in the performance of that compared to DDR.
Originally posted by THT
This is at 90 nm? If they do as you say, the 970fx would be going from 2 GHz to 3 GHz. That would be remarkable considering everyone elses troubles at 90 nm. Intel only did a little bit better with the Pentium 4 at 130 nm.
Last time I checked the 970fx was shipping at 2.5 GHz and has been for 6 months. And the transition to 65 nm isn't expected to improve the clock rates / power curves much, if at all.
Intel went from 3.15 -> 3.8 GHz on 90nm, which is about 20% as well. Of course they didn't gain anything on the first 130->90 switch, but they had tweaked out their 130nm process more than IBM had for the 970.
A Power derived CPU with multiple "Cell" cores? Sounds like a big chip, just as hot as 970fx, and costly.
You can't always make assumptions like that. Next week should bring answers.
http://www.blachford.info/computer/Cells/Cell4.html
Originally posted by Programmer
Last time I checked the 970fx was shipping at 2.5 GHz and has been for 6 months. And the transition to 65 nm isn't expected to improve the clock rates / power curves much, if at all.
Intel went from 3.15 -> 3.8 GHz on 90nm, which is about 20% as well. Of course they didn't gain anything on the first 130->90 switch, but they had tweaked out their 130nm process more than IBM had for the 970.
We're arguing over what is remarkable here. Do you believe IBM could have shipped a 2.5 GHz 130-nm 970? If so, I would agree that a 3 GHz 970fx is hardly remarkable. But if you don't think so, what clock could IBM have shipped that 130-nm 970 at? If it is 2 to 2.2 GHz, then a prospective 970fx clock increase to 3 GHz is pretty good scaling for all the troubles seen at 90 nm. That's proportionately better than Intel had done or has chosen to do with Prescott.
You can't always make assumptions like that. Next week should bring answers.
The session description says Cell is a "64 bit Power architecture." Once again, it's a question of nomenclature. I'm only taking the terms as they are known to be, and Power means either Power 3, Power 4, or Power 5, and Power 5 is mostly an SMT extension of the Power 4. So, this prospective Cell is a single core Power 4 with auxillary SIMD (streaming) units onboard, I/O and an integrated memory controller fabricated at 90 nm. Sounds very much like a 970fx type of processor in terms of power consumption and die size to me.
If it is a brand new Power ISA microarchitecture, then of course, all bets are off. If it is a Power chip as we know it, than I think it is good news all around. Whether it is a process tweak or a better design for power consumption, I'm sure they can make their way into processors for Apple.
Originally posted by THT
[B]We're arguing over what is remarkable here. Do you believe IBM could have shipped a 2.5 GHz 130-nm 970? If so, I would agree that a 3 GHz 970fx is hardly remarkable. But if you don't think so, what clock could IBM have shipped that 130-nm 970 at? If it is 2 to 2.2 GHz, then a prospective 970fx clock increase to 3 GHz is pretty good scaling for all the troubles seen at 90 nm. That's proportionately better than Intel had done or has chosen to do with Prescott.
I doubt a 2.5 GHz 130 nm 970 was practical to ship, especially given Apple's desire to keep its machines quiet. It may not have even functioned if there was a specific problem with the design.
IBM's 90 nm performance does look like one of the best, but remember that they took one crack at 130 nm and then switched to 90 nm. Intel did quite a bit of refinement on 130 nm, plus they were at a higher clock rate already and the relationship between power and frequency is highly non-linear. Intel has also shown a willingness to ship really hot chips. As a result its hard to compare the two. Compared to the last 10 years, however, a further 20% increase isn't particuarly noteworthy.
The session description says Cell is a "64 bit Power architecture." Once again, it's a question of nomenclature. I'm only taking the terms as they are known to be, and Power means either Power 3, Power 4, or Power 5, and Power 5 is mostly an SMT extension of the Power 4. So, this prospective Cell is a single core Power 4 with auxillary SIMD (streaming) units onboard, I/O and an integrated memory controller fabricated at 90 nm. Sounds very much like a 970fx type of processor in terms of power consumption and die size to me.
http://www.power.org
That is what Power architecture means. Wait until next week and you'll see what I mean.
If it is a brand new Power ISA microarchitecture, then of course, all bets are off. If it is a Power chip as we know it, than I think it is good news all around. Whether it is a process tweak or a better design for power consumption, I'm sure they can make their way into processors for Apple.
There are shades of grey between the options you outline. Processors are big and complex things, and there are a lot of options even within a single design.
I never said Apple couldn't use them.
Originally posted by mikenap
wow, powerful article about cell and Apples potential in using it. an intereting read, please post feedback...
http://www.blachford.info/computer/Cells/Cell4.html
This looks to me as if we were about to break some laws of physics (performance-wise). Can anyone in the know tell something on this article?
Originally posted by Programmer
http://www.power.org
That is what Power architecture means. Wait until next week and you'll see what I mean.
Anybody have any ideas why Apple is still absent from those on board on IBM's power list?
So IBM might well show off a variant of the Cell architecture that has no business in a Mac, but since Cell is highly modular, all bets are off in the general case.
You know, for Sky-Net?
Originally posted by PB
This looks to me as if we were about to break some laws of physics (performance-wise). Can anyone in the know tell something on this article?
That article is crap -- you don't need to know anything about Cell to tell that.
First of all, down the road Apple could be in a fine position to take advantage of such tech advances. Doubtful if Intel, MS, and Dell could be so nimble.
As an aside, I wonder if it would be the approach for TV's to render raw digital data on the fly. That would circumvent half of the current workflow of content creation, computer rendering, compression, and delivery.
Originally posted by Programmer
That article is crap -- you don't need to know anything about Cell to tell that.
Thanks, for a moment I thought that my common sense detectors were completely off.
Originally posted by mugwump
First of all, down the road Apple could be in a fine position to take advantage of such tech advances. Doubtful if Intel, MS, and Dell could be so nimble.
As an aside, I wonder if it would be the approach for TV's to render raw digital data on the fly. That would circumvent half of the current workflow of content creation, computer rendering, compression, and delivery.
Wow indeed mugwump!
Sony and Toshiba are already positioning the Cell into their TVs
Cell would seem to be able to create mini render farms on the workstation which would, as you say, reduce by half the effort of content creation. (with perhaps game coding aside)
Rendering and compression are ideal Cell tasks and should show dramatic improvements over current practices. Imagine Core Video being processed by the Cell in everything from the user interface to FCP.
delivery delivery? delivery!
MPEG4 - H264 part 10
IPTV - TV over broadband
INdTV - Al Gore's new venture
1394 - Firewire
THE VISION for the Cell processor as proclaimed by IBM, Sony, and Toshiba.
"Breakthrough microprocessor architecture that puts broadband communications right on the chip."
I'm doing a bit of connect the dots here, but I have this sort of unifying theory that mugwump has started with his post.
I'll take Apple's low profile with Power.org as a further manifestation of Apple's obsessive desire for holding their cards close to their chest. To continue the analogy, Apple also has an ace up their sleeve on this one that they will play soon. Core services in Tiger are the smoking gun here for me.
The turn on a dime from all their professed views as shown with the new mini's may also be indicative of how fast Apple could embrace the forgotten spoke to their digital hub - TV and not just any old TV, HDTV!
Get ready for the Apple channel! What does Al Gore know that would cause him to spend 70 million on a cable TV channel?
"Among those on the company's advisory board guiding the growth of INdTV is Apple CEO Steve Jobs. The company's COO Mark Goldman was formerly with News Corp. He built a satellite venture in Latin America. Goldman was the president and COO of Sky Latin America Partners prior to teaming up with Gore's venture."