Multi core GPUs? Why not?

[taltamir]

I think his postulations is one chip that contains multi die. Cut down each die to 1/4th of the size and put four of them on a single chip. Getting better yeilds as a result...
Interestingly enough, most prvious generation cards contain a second tiny die (1/15th the size of main die) containing the HDCP logic circuitry on the same chip. Rather then it being built in to the main die.

Thanks for that article about sony... shame on them. But what can they do, 10 to 20% yields are just painful. You would think they will be on the forefront of fabrication processess because of it, making 45nm chips... but for reason not...

 

[Matrices]

SLI will never ever be mainstream.

It is by definition inherently inefficient.

You are only using the video RAM equal to that of one card, so even if the vRAM usage is distributed evenly, half the RAM is literally being wasted. That's just wasted money sitting in your computer case.

The scalability is also poor. 30 to 50 percent performance - when drivers work properly - is the norm.

And finally, excess heat, noise and power consumption. It's double that of one card while the performance and actual utilization of cards is more like 1.5.

 

[taltamir]

you forgot the fact that you have incompatibilities... you can't just plug it in on any motherboard and have it work... your board has to be either SLI or xfire... AND have two slots to begin with.

If I wanted to sell my 7900gs and get a 3870 I could... If i had 7900GS in SLI and wanted to get 3870 crossfire I Would have had to replace my motherboard.

I will probably BUY (not just consider) sli/xfire the day they promise me it works on any board, using intel/nvidia/amd/via/sis chipset... that ofcourse means that they have to agree on a single standard and either make it open or cheap enough to license... which isn't happening any time soon.

 

[Cookie Monster]

Before the discussion carries on further i would like to note that GPUs have been "multi core" for years. Ever heard about quads or clusters etc?
G80 has 8 "clusters", so simply it has 8 "cores" although its quite difficult to explain this in a non technological manner not to mention the difference between a GPU and the CPU.

Now what i see is that the major IHVs could be going for "multiple GPU" concept. Examples are the Rage Fury Maxx, 7950GX2, R680 (RV670x2), and maybe rumoured G92 GX2. Although personally i dont like this at all.

However communication of between the multiple GPUs have always been the biggest obstacle meaning scaling is always varies depending on applications. Add in limitations/restrictions such as heat, power, cost compared to a single PCB GPU high end solution, size and various other things make it hard to push such a concept without having to spend alot of R&D.

 

[taltamir]

actually it has been mentioned that the stream processors and the other things are already like little cores...

generally though what people mean is multiple die... just slap more of them together and you get more performance... stupid waste of time, money, and research efforts... should be instead developping better cards. If one of the two companies had same price cards performing 10% better then their compeditors but with no multi gpu support they will be selling alot more cards... but they wouldn't be making as much money on their platform locked chipsets... and thats the point...

Heck with fusion amd wants to lock you into a 100% amd system... with centrino (centrino = brand name for a laptop containing an intel cpu, chipset, and wireless controller of any kind) intel already has suck locks for laptops, and they are developing their own video cards so they could sell all intel pcs...

When average joe goes to buy a computer in 2008 he asks the guy at best buy "why does this computer say AMD fusion, while this one just says it uses an AMD phenom processor"
wants to bet best buy employees are gonna tell people that "oh, that computer here has fusion and this one here doesn't means that this fusion PC is using a newer, better processor"
instead of "the computer here says fusion and this one doesn't because the fusion computer uses a phenom CPU, radeon video card, and amd chipset. while the other computer utilizes a phenom cpu with an nforce/via chipset and an nvidia geforce card giving you more bang for your buck"

It is the EXACT same thing that intel did with laptops... when you buy centrino you are just paying 50-100$ extra for an intel wireless card and motherboard. rather then a different mobo (or the same) with a netgear/dlink/whatever wireless card. (and it IS a card, you can open your laptop and replace it)

 

[hans007]

stream processors are cores.


its pretty stupid to use multiple dies because there is stuff you dont need more than 1 of. you dont need 2 2d engines, or say 2 HD decoders, or 2 memory controllers etc.

 

[taltamir]

mmm... you know what that makes me think of... multiple clusters on seperate die connected on a single chip...

So for example, have a single memory controller + engine + HD decoder etc die... then 2, 3, or 4 die each made out of X stream processors, etc... all on the same CHIP. but made of different die... it will give you 2 things. 1. You will be insanely scaleable, 2 you will have very high yields..

The cost would be that you are loosing a bit of efficiency... and you are increasing the complexity of the design.

 

[hans007]

Originally posted by: taltamir
mmm... you know what that makes me think of... multiple clusters on seperate die connected on a single chip...

So for example, have a single memory controller + engine + HD decoder etc die... then 2, 3, or 4 die each made out of X stream processors, etc... all on the same CHIP. but made of different die... it will give you 2 things. 1. You will be insanely scaleable, 2 you will have very high yields..

The cost would be that you are loosing a bit of efficiency... and you are increasing the complexity of the design.

external bus interconnects would not be as fast.


the memory bandwidth of the memory controller direct to the shader cores is much higher than over any bus.

basically the only way you'd be even possibly able to do it, is something like how the opterons work.

they all have their own memory controllers to their own memory pools with hypertransport between cpus only for transferring data over NUMA.

the problem is that graphics is far more memory intensive than most CPUs and actually probably wouldnt work that well over some sort of non unified memory.

that is the reason that cpus have huge caches, and gpus dont. so if you wanted to do this basically you'd still be left in the same place, you'd need seperate memory controllers per die, and basically you could not sure memory so it'd be exactly like crossfire or SLI with the little sli bridges instead of something like hypertransport.

 

[Skunkwourk]

Originally posted by: NanoStuff

Originally posted by: taltamir
um.... You just told him "you are wrong" and then paraphrased him...

Oh good, that means we're in agreement

haha, almost sig worthy

 

[taltamir]

Originally posted by: hans007

Originally posted by: taltamir
mmm... you know what that makes me think of... multiple clusters on seperate die connected on a single chip...

So for example, have a single memory controller + engine + HD decoder etc die... then 2, 3, or 4 die each made out of X stream processors, etc... all on the same CHIP. but made of different die... it will give you 2 things. 1. You will be insanely scaleable, 2 you will have very high yields..

The cost would be that you are loosing a bit of efficiency... and you are increasing the complexity of the design.

external bus interconnects would not be as fast.


the memory bandwidth of the memory controller direct to the shader cores is much higher than over any bus.

basically the only way you'd be even possibly able to do it, is something like how the opterons work.

they all have their own memory controllers to their own memory pools with hypertransport between cpus only for transferring data over NUMA.

the problem is that graphics is far more memory intensive than most CPUs and actually probably wouldnt work that well over some sort of non unified memory.

that is the reason that cpus have huge caches, and gpus dont. so if you wanted to do this basically you'd still be left in the same place, you'd need seperate memory controllers per die, and basically you could not sure memory so it'd be exactly like crossfire or SLI with the little sli bridges instead of something like hypertransport.

That would only be the case if there are two different CHIPS not two different cores on one chip.... two cores on one chip has the same access speed of L3 cache, doesn't need separate memory controllers, and is way faster then a multi chip one card solution. But still slower then one unified core. (think about the speed difference between the early intel quad cores which had two duel core dies on one chip... vs the ones with a single die containing all four cores)