Excuse me if I offended you. That was not my intention. I was simply disagreeing with your apparent assumption that this chip is a "proof of concept." Maybe it's partly a matter of words. In design engineering, proof of concept occurs very early in the life of a project. I believe the Cell project is at the "opposite" end of the time line, near final cleanup. (I do know the translation.) IBM and Sony are using this big and expensive Cell chips simply because nobody has 65 nm up and running reliably yet. It lets them complete the design before they have the final key component, and gives them a fast path project. There is a gamble here, betting that 65 nm will work out the way they think it will. At least that is my take on the thing. I did read your entire post but only wanted to comment on the single point that I quoted.
Ok. Certainly is true that proof of concept was a bad choice of words. I'm usually in such a hurry when I post I rarely get a chance to proof read what I write.
Opinion question for the group. Is it possible to have a future generation mac workstation, (with a pricetag to match) That could use future generation G5's, like the rumored upcoming dual core G5 chips to handle the grunt work, and include parts of the Cell design (not sure how this could be done, on some kind of daughter card maybe?) for specifics like HD final output renders in FCP? It would be amazing to edit in FCP or AE and have everything be realtime in and out! Or would apple be forced to use the entire Cell chip as a whole or not at all? (forgive my lack of technical terminology) If it's possible to use the best of both worlds, how would the Cell parts be added to the system? If some kind of ad on card, what inerface would be fast enough? How would a system like this compare to the best Intel/AMD have to offer? thnx or any thoughts/opinions!
Opinion question for the group. Is it possible to have a future generation mac workstation, (with a pricetag to match) That could use future generation G5's, like the rumored upcoming dual core G5 chips to handle the grunt work, and include parts of the Cell design (not sure how this could be done, on some kind of daughter card maybe?) for specifics like HD final output renders in FCP? It would be amazing to edit in FCP or AE and have everything be realtime in and out! Or would apple be forced to use the entire Cell chip as a whole or not at all? (forgive my lack of technical terminology) If it's possible to use the best of both worlds, how would the Cell parts be added to the system? If some kind of ad on card, what inerface would be fast enough? How would a system like this compare to the best Intel/AMD have to offer? thnx or any thoughts/opinions!
I believe that it was SJ himself that said not too long ago that we will be rendering in real time, and he would know of what IBM was working on. I do believe that it is possible, because of the SoC design, program the Cell and it will use the support units to execute the code. The CPUs could handle playing traffic cop/control and still have plenty of resources left to attend to other tasks. Yes why wouldn't a curtain other business that SJ is associated with be very interested in this tech.
Also Apple charges $300 to $450 for video cards, I would think that a Cell processor could be built and added for this cost or less. I believe that the MS Xbox was the only box that shipped at a loss to encourage their market. A Cell designed box would cost about $200 to $250 to build and sell, I'm pretty sure that Sony and the others are not ready to break with the working model, break even on the boxes and make the money on the games and API access. So I would guess that depending on the changes that Apple would ask for, if any, the cost could be less than $200 for a Cell based daughter card, there is my guess for the cost, which should not be a factor. I think that with everyone that could be making Cell chips that the cost will not be very high at all.
Quote:
Originally posted by Brendon
I could see Apple using this chip to compliment the 9XX processors. I could see Apple using this chip in a video board. I could also see Apple using something like this for scientific or heavy math work. I could also see Apple being interested in this from a design concept in that Apple may want IBM to make a similar chip using VMX units. The problem would be feeding the beast. A similar chip with 4 VMX cores would a great help for code that relies heavily on vector calcs. I guess I see this chip or something like it as a great support chip, while staying with the 9XX series because the cost of moving the code base to Cell. I see Apple easing into that water not jumping in. Maybe a Cell processor will be in an Apple machine soon 2006, but in the role of video accelerator, or math/heavy calculation support if the market is there.
The particular CPU that was released is not a particularly clean or intuitive fit with Macs, mostly because of the oddball PPC. On the one hand, it could replace a G4 convincingly except for the wattage—and not coincidentally, the G4 is winding up in models where that's important. On the other hand, there is one crucial area, double-precision floating point, where the 970 looks to have the Cell PPC for lunch. A step backward in this performance metric, which is particularly important to a lot of professional and scientific customers, would not do. Cell might be able to make make up the lack of 64-bit FP capability in its PPC core with a lot of cores (4-6 Cells per machine, or variants with two PPC cores and 4 APUs?), but that's not an even tradeoff. It would have to be designed carefully and measured carefully.
I believe that the Cell's PPE does have a full FPU, including a double precision capability. IBM has said it is a full Power implementation, and that, by definition, includes double precision support. At 4+ GHz this Power core should perform roughly like a 970 of about half the clock rate (my estimate and is conservative).
Note that the SPEs have double precision capability added but it is not pipelined, so is "only" in the range of 25-30 GFLOPS for the entire chip. This compares well to a quad 970 arrangement. For most of the expected uses for Cell (and for most uses Apple would put it to), single precision is far more important -- OpenGL, Quartz2D, CoreImage, CoreAudio, etc. And on that score the Cell leaves all existing processors sitting on their asses wondering what the hell hit them.
I believe that the Cell's PPE does have a full FPU, including a double precision capability. IBM has said it is a full Power implementation, and that, by definition, includes double precision support. At 4+ GHz this Power core should perform roughly like a 970 of about half the clock rate (my estimate and is conservative).
Note that the SPEs have double precision capability added but it is not pipelined, so is "only" in the range of 25-30 GFLOPS for the entire chip. This compares well to a quad 970 arrangement. For most of the expected uses for Cell (and for most uses Apple would put it to), single precision is far more important -- OpenGL, Quartz2D, CoreImage, CoreAudio, etc. And on that score the Cell leaves all existing processors sitting on their asses wonder what the hell hit them.
Wow is that boldly stated. Nobody has ever seen this processor in a demonstration of use have they? I mean to say it's putting everything on their asses, and all that already is like. Whooof. A lot to handle. I remember when before the 970 came out, everybody said the same thing, and Apple did that shifty demo at mac world, or WWDC. They hyped the 970's capabilities a lot more than this. And it really tuned out it really wasn't the x86 killer they made it out to be.
Wow is that boldly stated. Nobody has ever seen this processor in a demonstration of use have they? I mean to say it's putting everything on their asses, and all that already is like. Whooof. A lot to handle. I remember when before the 970 came out, everybody said the same thing, and Apple did that shifty demo at mac world, or WWDC. They hyped the 970's capabilities a lot more than this. And it really tuned out it really wasn't the x86 killer they made it out to be.
Like the man said, in single precision floating point operations on stream-able data, everything else is just so 20th century. Outside that narrow regime it is a decent chip, but not a world beater.
Wow is that boldly stated. Nobody has ever seen this processor in a demonstration of use have they? I mean to say it's putting everything on their asses, and all that already is like. Whooof. A lot to handle. I remember when before the 970 came out, everybody said the same thing, and Apple did that shifty demo at mac world, or WWDC. They hyped the 970's capabilities a lot more than this. And it really tuned out it really wasn't the x86 killer they made it out to be.
When the 970 came out I figured it would turn in excellent integer, FP, and vector performance compared to the G4. It would do okay on integer against the x86 crowd, quite well at floating point, and very well at vector. That has pretty much held up. Steve's promise of 3 GHz in a year I accepted because that was before it became clear that 90nm brought with it new challenges (which have impacted Intel & AMD just as much). The 970 was the first completely new PowerPC in about 7 years, and I feel it delivered well on what I expected.
The Cell is more like the GPU revolution that has been going on for the last 10 years or so. The GPU-set has been riding the wave of multi-threaded, deep pipeline vector hardware, but the CPU-set has been coasting on process technology and rapid scaling. The Cell is a sudden leap (well, not actually that sudden considering the 4-5 year development time) in the direction the GPUs have been going for some time plus pushing the process envelope to its limits. On problems with solutions that can take advantage of it, this thing will be amazing. It remains to be seen how many solutions can take advantage of it.
I'm sure somebody somewhere has seen a demo. IBM has told us they exist, and they even had one running 5.6 GHz (360 GFLOPS! 180 watts ).
When the 970 came out I figured it would turn in excellent integer, FP, and vector performance compared to the G4. It would do okay on integer against the x86 crowd, quite well at floating point, and very well at vector. That has pretty much held up. Steve's promise of 3 GHz in a year I accepted because that was before it became clear that 90nm brought with it new challenges (which have impacted Intel & AMD just as much). The 970 was the first completely new PowerPC in about 7 years, and I feel it delivered well on what I expected.
The Cell is more like the GPU revolution that has been going on for the last 10 years or so. The GPU-set has been riding the wave of multi-threaded, deep pipeline vector hardware, but the CPU-set has been coasting on process technology and rapid scaling. The Cell is a sudden leap (well, not actually that sudden considering the 4-5 year development time) in the direction the GPUs have been going for some time plus pushing the process envelope to its limits. On problems with solutions that can take advantage of it, this thing will be amazing. It remains to be seen how many solutions can take advantage of it.
I'm sure somebody somewhere has seen a demo. IBM has told us they exist, and they even had one running 5.6 GHz (360 GFLOPS! 180 watts ).
Don't you think that this new architecture, will recquiere a lot of software optimisation, in order to be able to take advantage of it.
I fear that the currents app, will be dog slow on the cell chip without the proper recompilation.
Don't you think that this new architecture, will recquiere a lot of software optimisation, in order to be able to take advantage of it.
I fear that the currents app, will be dog slow on the cell chip without the proper recompilation.
If Apple is going to buy into the Cell thingie, they will most likely use it to give the GFX subsystem a boost, starting with Quartz and going up to CoreVideo. Any app using those layers will profit.
OTOH, only few developers will really tax the power the specialized cores will provide - writing software cells might be not exactly trivial.
The 90nm part shown will be produced to create the workstations needed to program for PS3.
IBM and it's partners plan to step this Cell design down to 65nm and eventually 45nm.
I would expect that the PS3 will ship with the 65nm part, as to do otherwise would be to costly in terms of wafer use. At 65nm the Cell is appropriate to use as a coprocessor in a high end Apple graphics workstation.
A Powermac or xServe with the Cell as coprocessor would make complete sense in some markets, especially once the dual core G5's start showing up.
It should be cheaper from a chip cost to put a Cell processor in the second processor slot, than a second 970xx.
How about a build to order option where you could opt for one multi core G5 (for a dual), dual multi core G5's (for a quad), and one multi core G5 and one multi core Cell processor?
Choice is nice, and those with the need for what the Cell will offer can have it wrapped in OSX niceness.
Opinion question for the group. Is it possible to have a future generation mac workstation, (with a pricetag to match) That could use future generation G5's, like the rumored upcoming dual core G5 chips to handle the grunt work, and include parts of the Cell design (not sure how this could be done, on some kind of daughter card maybe?) for specifics like HD final output renders in FCP?
Yes, it is possible, with lots of work. No daughtercard needed because the Cell architecture is modular with the element interconnect bus (EIB). Everything would be on one die, system-on-a-chip style. If you've seen the Cell architecture diagrams, a 970 or 2 970s would replace the in-order PowerPC core as the powerpc processing element or processor element. The EIB would bridge the 970 core(s) with the SPEs, memory controller and I/O bus.
It would have consequences though. It won't clock as high as the Cell and it would consume more power at the same time. At 90 nm, with 2 970fx cores and 4 SPEs clocked at 2.5 GHz running at 1.1V, it'll be a 100+ watt chip.
I believe that we were talking about a daughter card, maybe a big daughter card, but still. What would the size be? Oh 8.7inches squared @90nm, or 4.5inches squared @65nm.
At 65nm daughter card is no problem, but 90nm chip is huge and I would think that supporting it and cooling it would be difficult unless you have this water cooling thing down
Yes, it is possible, with lots of work. No daughtercard needed because the Cell architecture is modular with the element interconnect bus (EIB). Everything would be on one die, system-on-a-chip style. If you've seen the Cell architecture diagrams, a 970 or 2 970s would replace the in-order PowerPC core as the powerpc processing element or processor element. The EIB would bridge the 970 core(s) with the SPEs, memory controller and I/O bus.
It would have consequences though. It won't clock as high as the Cell and it would consume more power at the same time. At 90 nm, with 2 970fx cores and 4 SPEs clocked at 2.5 GHz running at 1.1V, it'll be a 100+ watt chip.
hmm, which would be better; A dual processor dual core set-up or a single processor dual-core with cell attributes attached...DP dual core sounds a little redundant...
I believe that the Cell's PPE does have a full FPU, including a double precision capability. IBM has said it is a full Power implementation, and that, by definition, includes double precision support. At 4+ GHz this Power core should perform roughly like a 970 of about half the clock rate (my estimate and is conservative).
I never doubted that. The G4 has full double-precision FP support, too. But the 970 has two full FP units. The question was not, can Cell do 64-bit FP, but can it muster the performance of a 970 at 64-bit FP. (See below before answering.)
If you want to talk capability, they're all still catching up to the 68040's 80-bit FP capability.
Quote:
Note that the SPEs have double precision capability added but it is not pipelined, so is "only" in the range of 25-30 GFLOPS for the entire chip. This compares well to a quad 970 arrangement. For most of the expected uses for Cell (and for most uses Apple would put it to), single precision is far more important -- OpenGL, Quartz2D, CoreImage, CoreAudio, etc. And on that score the Cell leaves all existing processors sitting on their asses wondering what the hell hit them.
I wasn't aware that the SPEs (or APUs, whatever—the little buggers hanging off the bus) had 64-bit FP at all. That does change things somewhat, although the need to have 1 thread per FPU makes it an elliptical answer at best to the 970's dual FPUs.
I'm not even going to try to contest that 32-bit FP performance will own everything. That would be true even if Cell only had its 4GHz AltiVec unit to fall back on.
It would have consequences though. It won't clock as high as the Cell and it would consume more power at the same time. At 90 nm, with 2 970fx cores and 4 SPEs clocked at 2.5 GHz running at 1.1V, it'll be a 100+ watt chip.
Speaking of 100+ watts, I wonder how far along IBM is with their Strained Silicon Directly on Insulator technology? Heck, for that matter, I wonder how far along they are in implementing SOI, the last mention of it was Chipwork's statement that they were expecting samples "soon", whatever that means?
Speaking of 100+ watts, I wonder how far along IBM is with their Strained Silicon Directly on Insulator technology? Heck, for that matter, I wonder how far along they are in implementing SOI, the last mention of it was Chipwork's statement that they were expecting samples "soon", whatever that means?
Are you only referring to Cell? Because I thought that the 970 was SOI.
Comments
Originally posted by snoopy
Excuse me if I offended you. That was not my intention. I was simply disagreeing with your apparent assumption that this chip is a "proof of concept." Maybe it's partly a matter of words. In design engineering, proof of concept occurs very early in the life of a project. I believe the Cell project is at the "opposite" end of the time line, near final cleanup. (I do know the translation.) IBM and Sony are using this big and expensive Cell chips simply because nobody has 65 nm up and running reliably yet. It lets them complete the design before they have the final key component, and gives them a fast path project. There is a gamble here, betting that 65 nm will work out the way they think it will. At least that is my take on the thing. I did read your entire post but only wanted to comment on the single point that I quoted.
Ok.
Originally posted by mikenap
Opinion question for the group. Is it possible to have a future generation mac workstation, (with a pricetag to match) That could use future generation G5's, like the rumored upcoming dual core G5 chips to handle the grunt work, and include parts of the Cell design (not sure how this could be done, on some kind of daughter card maybe?) for specifics like HD final output renders in FCP? It would be amazing to edit in FCP or AE and have everything be realtime in and out! Or would apple be forced to use the entire Cell chip as a whole or not at all? (forgive my lack of technical terminology) If it's possible to use the best of both worlds, how would the Cell parts be added to the system? If some kind of ad on card, what inerface would be fast enough? How would a system like this compare to the best Intel/AMD have to offer? thnx or any thoughts/opinions!
I believe that it was SJ himself that said not too long ago that we will be rendering in real time, and he would know of what IBM was working on. I do believe that it is possible, because of the SoC design, program the Cell and it will use the support units to execute the code. The CPUs could handle playing traffic cop/control and still have plenty of resources left to attend to other tasks. Yes why wouldn't a curtain other business that SJ is associated with be very interested in this tech.
Also Apple charges $300 to $450 for video cards, I would think that a Cell processor could be built and added for this cost or less. I believe that the MS Xbox was the only box that shipped at a loss to encourage their market. A Cell designed box would cost about $200 to $250 to build and sell, I'm pretty sure that Sony and the others are not ready to break with the working model, break even on the boxes and make the money on the games and API access. So I would guess that depending on the changes that Apple would ask for, if any, the cost could be less than $200 for a Cell based daughter card, there is my guess for the cost, which should not be a factor. I think that with everyone that could be making Cell chips that the cost will not be very high at all.
Originally posted by Brendon
I could see Apple using this chip to compliment the 9XX processors. I could see Apple using this chip in a video board. I could also see Apple using something like this for scientific or heavy math work. I could also see Apple being interested in this from a design concept in that Apple may want IBM to make a similar chip using VMX units. The problem would be feeding the beast. A similar chip with 4 VMX cores would a great help for code that relies heavily on vector calcs. I guess I see this chip or something like it as a great support chip, while staying with the 9XX series because the cost of moving the code base to Cell. I see Apple easing into that water not jumping in. Maybe a Cell processor will be in an Apple machine soon 2006, but in the role of video accelerator, or math/heavy calculation support if the market is there.
Originally posted by Amorph
The particular CPU that was released is not a particularly clean or intuitive fit with Macs, mostly because of the oddball PPC. On the one hand, it could replace a G4 convincingly except for the wattage—and not coincidentally, the G4 is winding up in models where that's important. On the other hand, there is one crucial area, double-precision floating point, where the 970 looks to have the Cell PPC for lunch. A step backward in this performance metric, which is particularly important to a lot of professional and scientific customers, would not do. Cell might be able to make make up the lack of 64-bit FP capability in its PPC core with a lot of cores (4-6 Cells per machine, or variants with two PPC cores and 4 APUs?), but that's not an even tradeoff. It would have to be designed carefully and measured carefully.
I believe that the Cell's PPE does have a full FPU, including a double precision capability. IBM has said it is a full Power implementation, and that, by definition, includes double precision support. At 4+ GHz this Power core should perform roughly like a 970 of about half the clock rate (my estimate and is conservative).
Note that the SPEs have double precision capability added but it is not pipelined, so is "only" in the range of 25-30 GFLOPS for the entire chip. This compares well to a quad 970 arrangement. For most of the expected uses for Cell (and for most uses Apple would put it to), single precision is far more important -- OpenGL, Quartz2D, CoreImage, CoreAudio, etc. And on that score the Cell leaves all existing processors sitting on their asses wondering what the hell hit them.
Originally posted by Programmer
I believe that the Cell's PPE does have a full FPU, including a double precision capability. IBM has said it is a full Power implementation, and that, by definition, includes double precision support. At 4+ GHz this Power core should perform roughly like a 970 of about half the clock rate (my estimate and is conservative).
Note that the SPEs have double precision capability added but it is not pipelined, so is "only" in the range of 25-30 GFLOPS for the entire chip. This compares well to a quad 970 arrangement. For most of the expected uses for Cell (and for most uses Apple would put it to), single precision is far more important -- OpenGL, Quartz2D, CoreImage, CoreAudio, etc. And on that score the Cell leaves all existing processors sitting on their asses wonder what the hell hit them.
Wow is that boldly stated. Nobody has ever seen this processor in a demonstration of use have they? I mean to say it's putting everything on their asses, and all that already is like. Whooof. A lot to handle. I remember when before the 970 came out, everybody said the same thing, and Apple did that shifty demo at mac world, or WWDC. They hyped the 970's capabilities a lot more than this. And it really tuned out it really wasn't the x86 killer they made it out to be.
Originally posted by onlooker
Wow is that boldly stated. Nobody has ever seen this processor in a demonstration of use have they? I mean to say it's putting everything on their asses, and all that already is like. Whooof. A lot to handle. I remember when before the 970 came out, everybody said the same thing, and Apple did that shifty demo at mac world, or WWDC. They hyped the 970's capabilities a lot more than this. And it really tuned out it really wasn't the x86 killer they made it out to be.
Like the man said, in single precision floating point operations on stream-able data, everything else is just so 20th century. Outside that narrow regime it is a decent chip, but not a world beater.
Originally posted by onlooker
Wow is that boldly stated. Nobody has ever seen this processor in a demonstration of use have they? I mean to say it's putting everything on their asses, and all that already is like. Whooof. A lot to handle. I remember when before the 970 came out, everybody said the same thing, and Apple did that shifty demo at mac world, or WWDC. They hyped the 970's capabilities a lot more than this. And it really tuned out it really wasn't the x86 killer they made it out to be.
When the 970 came out I figured it would turn in excellent integer, FP, and vector performance compared to the G4. It would do okay on integer against the x86 crowd, quite well at floating point, and very well at vector. That has pretty much held up. Steve's promise of 3 GHz in a year I accepted because that was before it became clear that 90nm brought with it new challenges (which have impacted Intel & AMD just as much). The 970 was the first completely new PowerPC in about 7 years, and I feel it delivered well on what I expected.
The Cell is more like the GPU revolution that has been going on for the last 10 years or so. The GPU-set has been riding the wave of multi-threaded, deep pipeline vector hardware, but the CPU-set has been coasting on process technology and rapid scaling. The Cell is a sudden leap (well, not actually that sudden considering the 4-5 year development time) in the direction the GPUs have been going for some time plus pushing the process envelope to its limits. On problems with solutions that can take advantage of it, this thing will be amazing. It remains to be seen how many solutions can take advantage of it.
I'm sure somebody somewhere has seen a demo. IBM has told us they exist, and they even had one running 5.6 GHz (360 GFLOPS! 180 watts
Originally posted by Programmer
When the 970 came out I figured it would turn in excellent integer, FP, and vector performance compared to the G4. It would do okay on integer against the x86 crowd, quite well at floating point, and very well at vector. That has pretty much held up. Steve's promise of 3 GHz in a year I accepted because that was before it became clear that 90nm brought with it new challenges (which have impacted Intel & AMD just as much). The 970 was the first completely new PowerPC in about 7 years, and I feel it delivered well on what I expected.
The Cell is more like the GPU revolution that has been going on for the last 10 years or so. The GPU-set has been riding the wave of multi-threaded, deep pipeline vector hardware, but the CPU-set has been coasting on process technology and rapid scaling. The Cell is a sudden leap (well, not actually that sudden considering the 4-5 year development time) in the direction the GPUs have been going for some time plus pushing the process envelope to its limits. On problems with solutions that can take advantage of it, this thing will be amazing. It remains to be seen how many solutions can take advantage of it.
I'm sure somebody somewhere has seen a demo. IBM has told us they exist, and they even had one running 5.6 GHz (360 GFLOPS! 180 watts
Don't you think that this new architecture, will recquiere a lot of software optimisation, in order to be able to take advantage of it.
I fear that the currents app, will be dog slow on the cell chip without the proper recompilation.
Originally posted by Brendon
I could see Apple using this chip to compliment the 9XX processors.
Compliment? It'd take up half the motherboard.
Originally posted by Powerdoc
Don't you think that this new architecture, will recquiere a lot of software optimisation, in order to be able to take advantage of it.
I fear that the currents app, will be dog slow on the cell chip without the proper recompilation.
If Apple is going to buy into the Cell thingie, they will most likely use it to give the GFX subsystem a boost, starting with Quartz and going up to CoreVideo. Any app using those layers will profit.
OTOH, only few developers will really tax the power the specialized cores will provide - writing software cells might be not exactly trivial.
Originally posted by BenRoethig
... It'd take up half the motherboard.
The 90nm part shown will be produced to create the workstations needed to program for PS3.
IBM and it's partners plan to step this Cell design down to 65nm and eventually 45nm.
I would expect that the PS3 will ship with the 65nm part, as to do otherwise would be to costly in terms of wafer use. At 65nm the Cell is appropriate to use as a coprocessor in a high end Apple graphics workstation.
A Powermac or xServe with the Cell as coprocessor would make complete sense in some markets, especially once the dual core G5's start showing up.
It should be cheaper from a chip cost to put a Cell processor in the second processor slot, than a second 970xx.
How about a build to order option where you could opt for one multi core G5 (for a dual), dual multi core G5's (for a quad), and one multi core G5 and one multi core Cell processor?
Choice is nice, and those with the need for what the Cell will offer can have it wrapped in OSX niceness.
Originally posted by mikenap
Opinion question for the group. Is it possible to have a future generation mac workstation, (with a pricetag to match) That could use future generation G5's, like the rumored upcoming dual core G5 chips to handle the grunt work, and include parts of the Cell design (not sure how this could be done, on some kind of daughter card maybe?) for specifics like HD final output renders in FCP?
Yes, it is possible, with lots of work. No daughtercard needed because the Cell architecture is modular with the element interconnect bus (EIB). Everything would be on one die, system-on-a-chip style. If you've seen the Cell architecture diagrams, a 970 or 2 970s would replace the in-order PowerPC core as the powerpc processing element or processor element. The EIB would bridge the 970 core(s) with the SPEs, memory controller and I/O bus.
It would have consequences though. It won't clock as high as the Cell and it would consume more power at the same time. At 90 nm, with 2 970fx cores and 4 SPEs clocked at 2.5 GHz running at 1.1V, it'll be a 100+ watt chip.
Originally posted by BenRoethig
Compliment? It'd take up half the motherboard.
I believe that we were talking about a daughter card, maybe a big daughter card, but still. What would the size be? Oh 8.7inches squared @90nm, or 4.5inches squared @65nm.
At 65nm daughter card is no problem, but 90nm chip is huge and I would think that supporting it and cooling it would be difficult unless you have this water cooling thing down
Originally posted by THT
Yes, it is possible, with lots of work. No daughtercard needed because the Cell architecture is modular with the element interconnect bus (EIB). Everything would be on one die, system-on-a-chip style. If you've seen the Cell architecture diagrams, a 970 or 2 970s would replace the in-order PowerPC core as the powerpc processing element or processor element. The EIB would bridge the 970 core(s) with the SPEs, memory controller and I/O bus.
It would have consequences though. It won't clock as high as the Cell and it would consume more power at the same time. At 90 nm, with 2 970fx cores and 4 SPEs clocked at 2.5 GHz running at 1.1V, it'll be a 100+ watt chip.
hmm, which would be better; A dual processor dual core set-up or a single processor dual-core with cell attributes attached...DP dual core sounds a little redundant...
Originally posted by Powerdoc
Don't you think that this new architecture, will recquiere a lot of software optimisation, in order to be able to take advantage of it.
I fear that the currents app, will be dog slow on the cell chip without the proper recompilation.
Sure, it might cost you half your clock rate. Oh, right, that's about where we are now anyhow...
replacing the G5 I'd rather see it incorporated like the xRaid. Apple
creates a device with the xRaids dimensions & have it loaded with cell
graphics cards to be specifically used for rendering pixar quality
animation.
xGraphics - Price ?
Originally posted by Programmer
I believe that the Cell's PPE does have a full FPU, including a double precision capability. IBM has said it is a full Power implementation, and that, by definition, includes double precision support. At 4+ GHz this Power core should perform roughly like a 970 of about half the clock rate (my estimate and is conservative).
I never doubted that. The G4 has full double-precision FP support, too. But the 970 has two full FP units. The question was not, can Cell do 64-bit FP, but can it muster the performance of a 970 at 64-bit FP. (See below before answering.)
If you want to talk capability, they're all still catching up to the 68040's 80-bit FP capability.
Note that the SPEs have double precision capability added but it is not pipelined, so is "only" in the range of 25-30 GFLOPS for the entire chip. This compares well to a quad 970 arrangement. For most of the expected uses for Cell (and for most uses Apple would put it to), single precision is far more important -- OpenGL, Quartz2D, CoreImage, CoreAudio, etc. And on that score the Cell leaves all existing processors sitting on their asses wondering what the hell hit them.
I wasn't aware that the SPEs (or APUs, whatever—the little buggers hanging off the bus) had 64-bit FP at all. That does change things somewhat, although the need to have 1 thread per FPU makes it an elliptical answer at best to the 970's dual FPUs.
I'm not even going to try to contest that 32-bit FP performance will own everything. That would be true even if Cell only had its 4GHz AltiVec unit to fall back on.
I've got some reading to do this weekend.
Originally posted by THT
...
It would have consequences though. It won't clock as high as the Cell and it would consume more power at the same time. At 90 nm, with 2 970fx cores and 4 SPEs clocked at 2.5 GHz running at 1.1V, it'll be a 100+ watt chip.
Speaking of 100+ watts, I wonder how far along IBM is with their Strained Silicon Directly on Insulator technology? Heck, for that matter, I wonder how far along they are in implementing SOI, the last mention of it was Chipwork's statement that they were expecting samples "soon", whatever that means?
Originally posted by rickag
Speaking of 100+ watts, I wonder how far along IBM is with their Strained Silicon Directly on Insulator technology? Heck, for that matter, I wonder how far along they are in implementing SOI, the last mention of it was Chipwork's statement that they were expecting samples "soon", whatever that means?
Are you only referring to Cell? Because I thought that the 970 was SOI.