But if u think as Intel's Terascale having 80 cores at 3.1 ghz and giving 1 teraflop while maintaining reasonable power consumption, and AMD's on chip GPU running at 3 ghz with 48 GPU pipes while delivering 1 TFLOP. You can make an observation that this is feasible
The AMD/ATI "Fusion" CPU/GPU Discussion
- Thread starter DemoCoder
- Start date
But if u think as Intel's Terascale having 80 cores at 3.1 ghz and giving 1 teraflop while maintaining reasonable power consumption, and AMD's on chip GPU running at 3 ghz with 48 GPU pipes while delivering 1 TFLOP. You can make an observation that this is feasible
AMD hasn't proven it can get a GPU pipe to that clock speed. They're hoping it can reach 3GHz, but Intel's project isn't the same as what AMD hopes to do.
Intel has shown silicon that at least shows it can manufacture a device of that type, but those mini cores do not have the same kind of job as a GPU pipe.
Second, AMD can't do what Intel did to supply sufficient bandwidth. Intel bonded an EDRAM die directly onto die with the cores. That would be a thermal nightmare for an AMD chip that will have a full-fledged processor right next to the GPU.
The x86 portion will also take up a lot of space, so the GPU will not have the same amount of chip space as the 80 mini-cores in the Terascale example.
Not to mention 48 shader cores running at 3 GHz would have 3-4 times the computational oomph of todays top of the line GPUs. Keeping those fed would be... troublesome.
I'd imagine 4 unified shader cores at 3GHZ would be enough to saturate any given CPU socket.
Cheers
I'd imagine 4 unified shader cores at 3GHZ would be enough to saturate any given CPU socket.
Cheers
Last edited by a moderator:
does software have to be rewritten to tak advantage of the GPU on Fusion chips?
If so then AMD better dump this idea and adopt a Terascale project just like Intel's. Why, because with a Terascale chip we will be able to have 1 teraflop of performance on the CPU itself, just imagine 1 Teraflop of graphics\physics\AI on one CPU. Somebody might come up and say that here also the software will have to be rewritten. My answer is that it is better to rewrite software for something that will produce higher performance ( Terascale), other than Fusion. I am saying that a Terasclae idea would be better if and only if the software has to be rewritten to take advantage of Fusion
If so then AMD better dump this idea and adopt a Terascale project just like Intel's. Why, because with a Terascale chip we will be able to have 1 teraflop of performance on the CPU itself, just imagine 1 Teraflop of graphics\physics\AI on one CPU. Somebody might come up and say that here also the software will have to be rewritten. My answer is that it is better to rewrite software for something that will produce higher performance ( Terascale), other than Fusion. I am saying that a Terasclae idea would be better if and only if the software has to be rewritten to take advantage of Fusion
Well, I'd think not. At least not "rewritten", which implies existing software being changed. Now, new software might be written somewhat differently to take better advantage of the performance profile that Fusion presents, whatever that is. But I'd certainly hope that we aren't planning to abandon API's here, and that that level of abstraction asssures backwards compatibility.
Unless, of course, the "software" being rewritten is AMD's DX10 driver, then of course, same old same old for a new generation graphics device.
I thnk the best way to do this is to use a mobile GPU core. For example, if a CPU draws 70 W, and a mobile GPU core thay draws areound 40 W is stuck next to it, u get 110 W, which is very accepatble. However, another way will be by allowing the GPU to take more transistor count, but this can also have a negative imapct on CPU performance, since the no. of cores will have to be less.
do u think that allowing the GPU to take more transistors on Fusion chips will be better? Thus allowing for a high-implementation of of Fusion . I look at it this way. What will a gamer use his PC for other than gaming, Web browsing, e-mail checking. All these can be easily done on a very simple CPU, and don't forget that physix and AI will also be done on the GPU in future, thus decreasing the need for a powerful CPU.
A high-end GPU paired with a weak CPU increases the chance that it will be CPU limited. There aren't many games that have heavy graphics use that don't have significant CPU load, and AMD is unlikely to put its CPU in the back-seat anyway.
The CPU still has to draw up the command lists and run the drivers, even more so if this is some kind of SLI configuration.
I also don't expect GPUs to completely take on AI or Physics. There are certain portions of Physics that can be run on a GPU pretty well, as well as some (mostly low-level) AI.
In more complex AI especially, the way GPUs work currently would be stupendously wasteful.
Let's not forget that the GPU is also supposed to be drawing frames during all this.
More AMD slides from [H] http://www.hardocp.com/news.html?news=MjI0OTUsLCxobmV3cywsLDE=
http://www.hardocp.com/image.html?image=MTE2MzY5MTQ0MUJ4UnByVWtEV1pfMV8xM19sLmdpZg== is a bit interesting, tho this early in the design phase I don't know how literal I would take it.
Some other high points: an acknowledgement that memory demands will limit performance. Pointing at the process advantage of being their own fabs (whereas initially they were at pains to not spook TSMC, apparently), and another promise that they won't abandon high-end discrete, here http://www.hardocp.com/image.html?image=MTE2MzY5MTQ0MUJ4UnByVWtEV1pfMV8xNF9sLmdpZg==
http://www.hardocp.com/image.html?image=MTE2MzY5MTQ0MUJ4UnByVWtEV1pfMV8xM19sLmdpZg== is a bit interesting, tho this early in the design phase I don't know how literal I would take it.
Some other high points: an acknowledgement that memory demands will limit performance. Pointing at the process advantage of being their own fabs (whereas initially they were at pains to not spook TSMC, apparently), and another promise that they won't abandon high-end discrete, here http://www.hardocp.com/image.html?image=MTE2MzY5MTQ0MUJ4UnByVWtEV1pfMV8xNF9sLmdpZg==
Due to a bit of Keystone cops-ery, we missed out on the conference call. However, we've received the slide deck now, and it appears that [H] did not find the final section on AMD's Process roadmap to be interesting enough to include.
Well, we think our kind of B3D geeks would be appalled by that decision, so for your geekery pleasure. . .
EDIT: Brent from [H] dropped me a line that they did handle these slides in a subsequent news piece posted just a little later than the original upstream which I missed. Happy to make the correction. . .
Well, we think our kind of B3D geeks would be appalled by that decision, so for your geekery pleasure. . .
EDIT: Brent from [H] dropped me a line that they did handle these slides in a subsequent news piece posted just a little later than the original upstream which I missed. Happy to make the correction. . .
I think there is one side sorely missing from this debate: the simple urge to build more powerful processor chips, and have them sold aplenty. Because that is the technological bottom line: build better (faster/more powerful/cheaper) chips, that people want.
Just dumping more x86 cores on a die is not going to work. Neither is simply putting the GPU on the CPU die. And, there is this huge x86 legacy to take into account. The chips have to run any ancient x86 program faster. Or, do they?
First is virtualization. If you can have a single chip run multiple computers, is that a win? You might have a good argument to sell those chips, but if you sell less of them (because they can power multiple workspaces at the same time), you're not going to sell more of them. But at the same time, that is unavoidable. And that requires integrating the chipset before the GPU.
Then again, if you limit the amount of I/O devices you can hook up, you can still run all those Virtual Machines, but you need a chip at every worker's location. That's good. And, you have different requirements at every location.
Mass production of silicon chips functioning as it does, it makes more sense to produce a single chip that can do it all, and sell the ones that have local defects as cheaper versions, or scale back the design than making a lot of different designs to fill all the niches.
But, let's get back to making better processors. What, exactly, is a candidate to embed and deliver a better consumer satisfaction? Hm. CPUs are pretty much fast enough as they are, some marginal markets excluded. And the last generation(s) of GPUs don't only offer a benefit in reduced costs and consumer satisfaction, but they can actually make CPUs much more powerful, while retaining backwards compatibility: by using all the stream processors as floating point, vertex, MMX, SSE, and (of course
) 3D graphics addons.
Ok, so far, so good. But, as all of this runs in VMs, that can use all the legacy APIs, you can also expose those new units directly. And offering a vastly improved CPU speed while doing so.
4 nVidia 8800 cards in SLI are close to supercomputing territoria. When using the direct programming mode.
And that's about the only way you can offer vastly more powerful (and desirable) processors. They're going to do just that, IMHO. Because the general purpose CPU is more than powerfull enough for all intents and purposes, and won't scale anymore in either case.
Or, does anyone have a better way to increase the desireability of CPUs? Because, market hype will only get you so far. As soon as the majority of customers sees a definite lack of speed increase with their quad core CPU, you need to be able to offer something better. And those users already have an integrated GPU. Do they care where it resides?
Just dumping more x86 cores on a die is not going to work. Neither is simply putting the GPU on the CPU die. And, there is this huge x86 legacy to take into account. The chips have to run any ancient x86 program faster. Or, do they?
First is virtualization. If you can have a single chip run multiple computers, is that a win? You might have a good argument to sell those chips, but if you sell less of them (because they can power multiple workspaces at the same time), you're not going to sell more of them. But at the same time, that is unavoidable. And that requires integrating the chipset before the GPU.
Then again, if you limit the amount of I/O devices you can hook up, you can still run all those Virtual Machines, but you need a chip at every worker's location. That's good. And, you have different requirements at every location.
Mass production of silicon chips functioning as it does, it makes more sense to produce a single chip that can do it all, and sell the ones that have local defects as cheaper versions, or scale back the design than making a lot of different designs to fill all the niches.
But, let's get back to making better processors. What, exactly, is a candidate to embed and deliver a better consumer satisfaction? Hm. CPUs are pretty much fast enough as they are, some marginal markets excluded. And the last generation(s) of GPUs don't only offer a benefit in reduced costs and consumer satisfaction, but they can actually make CPUs much more powerful, while retaining backwards compatibility: by using all the stream processors as floating point, vertex, MMX, SSE, and (of course
) 3D graphics addons.Ok, so far, so good. But, as all of this runs in VMs, that can use all the legacy APIs, you can also expose those new units directly. And offering a vastly improved CPU speed while doing so.
4 nVidia 8800 cards in SLI are close to supercomputing territoria. When using the direct programming mode.
And that's about the only way you can offer vastly more powerful (and desirable) processors. They're going to do just that, IMHO. Because the general purpose CPU is more than powerfull enough for all intents and purposes, and won't scale anymore in either case.
Or, does anyone have a better way to increase the desireability of CPUs? Because, market hype will only get you so far. As soon as the majority of customers sees a definite lack of speed increase with their quad core CPU, you need to be able to offer something better. And those users already have an integrated GPU. Do they care where it resides?
Flawless execution? 65nm only by "late" Q4? Full conversion only by mid 2007? Intel has been shipping 65nm Core since January, and is practically flooding sales channels with 65nm products. AMD is a year late to the party. I mean, I certainly expect the K8L to be a hit, but AMD needs to get its shit in gear with respect to keeping up with Intel in process technology. Intel made a huge blunder with NetBurst, and at that time, being on a large process was less of a disadvantage because your competitors architecture had a perf/clock and perf/watt sucked. But now Intel's perf/clock and perf/watt doesn't suck, and so a year head start on process tech is much more dangerous.
AMD has been good for Intel and us consumers. Had AMD not existed, we would not have seen so much innovation from Intel. Linux + x86 ISA has practically killed off competiting chip designs in the server market (PA-RISC, SPARC, MIPS, PPC, et al) commodifying everything in its wake, so Intel would not have needed to be so aggressive about rearchitecting failed designs.
IMHO, AMD is in a dangerous position right now. They are small compared to Intel and need all the advantages they can get. With NetBurst, they had Intel by the balls, and should have used that time to put a larger distance between themselves and Intel. Instead, it seems they frittered it away, and now are in serious danger of being overwhelmed by a giant who has woken up and smelled the coffee. Their design wins and partnerships with major PC manufacturers can't be depended on to sustain them if they can't maintain tech parity. AMD should have been FAR more aggressive on new microarchitectures and probably should have used their assets to get more aggressive on process, vs going on a buying spree/aquisition. Their core competency and focus is CPUs, and diversification won't save them if they can't hang in that market in the first place.
If you've been in this industry for awhile, you'll remember lots of once great CPUs and manufacturers, who lost their lead, their mindshare, and tanked overnight, or just plain got bought out due to execution failures. Only Sun and IBM are really hanging on by a thread, Sun is hoping Niagara will save them, otherwise, they'll just be a seller of high-end x86 boxes with Linux, and from SGI and others, we know where this route takes them. With IBM, PPC (and their own architectures) survives because of niche contracts in embedded/HPC space, as well as plain old pathdependence and legacy of all of the mainframes and minicomputers out there, with gazillions of old business apps running on IBM operating systems.
If AMD doesn't get its shit in gear, our future of Intel with a complete monopoly in the desktop with be very bleak.
AMD has been good for Intel and us consumers. Had AMD not existed, we would not have seen so much innovation from Intel. Linux + x86 ISA has practically killed off competiting chip designs in the server market (PA-RISC, SPARC, MIPS, PPC, et al) commodifying everything in its wake, so Intel would not have needed to be so aggressive about rearchitecting failed designs.
IMHO, AMD is in a dangerous position right now. They are small compared to Intel and need all the advantages they can get. With NetBurst, they had Intel by the balls, and should have used that time to put a larger distance between themselves and Intel. Instead, it seems they frittered it away, and now are in serious danger of being overwhelmed by a giant who has woken up and smelled the coffee. Their design wins and partnerships with major PC manufacturers can't be depended on to sustain them if they can't maintain tech parity. AMD should have been FAR more aggressive on new microarchitectures and probably should have used their assets to get more aggressive on process, vs going on a buying spree/aquisition. Their core competency and focus is CPUs, and diversification won't save them if they can't hang in that market in the first place.
If you've been in this industry for awhile, you'll remember lots of once great CPUs and manufacturers, who lost their lead, their mindshare, and tanked overnight, or just plain got bought out due to execution failures. Only Sun and IBM are really hanging on by a thread, Sun is hoping Niagara will save them, otherwise, they'll just be a seller of high-end x86 boxes with Linux, and from SGI and others, we know where this route takes them. With IBM, PPC (and their own architectures) survives because of niche contracts in embedded/HPC space, as well as plain old pathdependence and legacy of all of the mainframes and minicomputers out there, with gazillions of old business apps running on IBM operating systems.
If AMD doesn't get its shit in gear, our future of Intel with a complete monopoly in the desktop with be very bleak.
AMD's ramp is delayed from when they originally hoped, but it has always been behind.
It is doubtful they can ever truly catch up with Intel, the larger company is just more capable of pushing fast transitions. It has more engineers and more fabs for test runs. AMD has to overlap transitions on a handful of production fabs, and it is still much smaller than Intel.
Intel can bleed much more cash. AMD is not in a position to take losses, which with Conroe, Cloverton, and others encroaching even in servers is likely to happen again.
Only Intel has the resources to do what it does, every other manufacturer has been forced to collaborate with others, something Intel doesn't do much (yet).
AMD's best hope is that the difficulty in scaling geometries further is becoming so great that even Intel will be forced to extend the length of process generations. Unfortunately, it's not going to get any easier for AMD when this happens.
It is doubtful they can ever truly catch up with Intel, the larger company is just more capable of pushing fast transitions. It has more engineers and more fabs for test runs. AMD has to overlap transitions on a handful of production fabs, and it is still much smaller than Intel.
Intel can bleed much more cash. AMD is not in a position to take losses, which with Conroe, Cloverton, and others encroaching even in servers is likely to happen again.
Only Intel has the resources to do what it does, every other manufacturer has been forced to collaborate with others, something Intel doesn't do much (yet).
AMD's best hope is that the difficulty in scaling geometries further is becoming so great that even Intel will be forced to extend the length of process generations. Unfortunately, it's not going to get any easier for AMD when this happens.
Well, it's certainly true, but I don't think AMD has been smart with its resources as it could have been. One can only hope that the $5billion invested in ATI will yield them an earnings stream to pay down debt and raise capital for fab improvements.
Ironically, there is one company that has the capital and expertise necessarily to rescue AMD and go up against Intel, and that's IBM. In fact, from a basic research point of view, IBM stands above Intel (I used to work at IBM T.J. Watson research, so I am biased). They've got 10 billion in cash, hugely diverse revenue streams, and massive credit worthiness.
So if AMD eventually flouders, IMHO, the best fit is an IBM acquisition, because as improving process becomes more and more expensive and utilizes more exotic techniques, no other company outside Intel, except IBM, IMHO, has the scale, resources, expertise, to break out.
Unfortunately, IBM may take a pass. They make far more money off software and support and reselling Intel chips, there is little need to get into the CPU business, when they are divesting themselves slowly of capital intensive activities. With the exception of PowerPC, which IMHO, is going to eventually disappear outside of specialized HPC contracts and support for legacy IBM customers, it's a lot of money to spend for thinner margins.
"Fabless" companies have proven a workable model, one with higher returns, since you can license IP instead of building stuff. From IBM's point of view, any new radical process techniques they come up with, they'll just patent, and then license the patent to people crazy enough to invest in fabs, which IMHO, will increasingly be Asian financed/state subsidized activities. TSMC is just the tip of the iceberg, wait until mainland China goes full-bore with fab construction, some of it backed by bank loans from communist affiliated banks. ROI simply won't be as much of a concern, IMHO, it will be the era of "Airbus"-like joint-ventures with foreign investors and government backing to reduce the riskiness of massive capital investment.
Ironically, there is one company that has the capital and expertise necessarily to rescue AMD and go up against Intel, and that's IBM. In fact, from a basic research point of view, IBM stands above Intel (I used to work at IBM T.J. Watson research, so I am biased). They've got 10 billion in cash, hugely diverse revenue streams, and massive credit worthiness.
So if AMD eventually flouders, IMHO, the best fit is an IBM acquisition, because as improving process becomes more and more expensive and utilizes more exotic techniques, no other company outside Intel, except IBM, IMHO, has the scale, resources, expertise, to break out.
Unfortunately, IBM may take a pass. They make far more money off software and support and reselling Intel chips, there is little need to get into the CPU business, when they are divesting themselves slowly of capital intensive activities. With the exception of PowerPC, which IMHO, is going to eventually disappear outside of specialized HPC contracts and support for legacy IBM customers, it's a lot of money to spend for thinner margins.
"Fabless" companies have proven a workable model, one with higher returns, since you can license IP instead of building stuff. From IBM's point of view, any new radical process techniques they come up with, they'll just patent, and then license the patent to people crazy enough to invest in fabs, which IMHO, will increasingly be Asian financed/state subsidized activities. TSMC is just the tip of the iceberg, wait until mainland China goes full-bore with fab construction, some of it backed by bank loans from communist affiliated banks. ROI simply won't be as much of a concern, IMHO, it will be the era of "Airbus"-like joint-ventures with foreign investors and government backing to reduce the riskiness of massive capital investment.
The very-delayed process node AMD is struggling to transition to is the product of an IBM-AMD collaboration. They've been working together on process technology for some time.
They both are still lagging behind Intel. IBM's service segment does very well, but it has less than stellar luck with hardware. If it weren't for the full control of the software stack and system platform, IBM's chips would have been outclassed.
IBM's core competencies are being challenged, and there is some doubt about the long-term growth prospects of its mainframe division.
They both are still lagging behind Intel. IBM's service segment does very well, but it has less than stellar luck with hardware. If it weren't for the full control of the software stack and system platform, IBM's chips would have been outclassed.
IBM's core competencies are being challenged, and there is some doubt about the long-term growth prospects of its mainframe division.
Similar threads
