GPU Folding

So, would you?

  • Yes

    Votes: 75 65.2%

  • No (why?)

    Votes: 40 34.8%
  • Total voters
    115
I'd rather not stress my graphics card more than i need too. That's what my overclocked X2-3800+ is for ;)
 
I just wanted to say that I am pretty sure that it can work just fine on nvidia hardware.

I have played with their (nvidia's) Gelato software, and a 7800 level GPU can render a scene about as fast as a QUAD core opteron workstation in MAX.

it's really neat to watch a scene come together in gelato, and it gives you an idea of how fast a GPU can be as a general purpose processing unit.
 
Had I card that could, I would. However, I would try to run the card at 50% so I could maintain some semblance of silence and to extend the life of the card a bit. I don't think running a high-end card at 100% 24x7 would be a good thing. If a X1300 was at stake here, I'd already have one warmed up for Folding.
 
Well, that is what the testing phase of this project is for. Stanford is seeing how well these cards hold up to 24/7 use. Pending those results, I would assume we'd see a client soon.

Edit: Oh and thanks for the english lesson, but I used moot properly. :p
 
Yashu said:
I just wanted to say that I am pretty sure that it can work just fine on nvidia hardware.

I have played with their (nvidia's) Gelato software, and a 7800 level GPU can render a scene about as fast as a QUAD core opteron workstation in MAX.

it's really neat to watch a scene come together in gelato, and it gives you an idea of how fast a GPU can be as a general purpose processing unit.
Click to expand...
A GPU can render a scene faster then 10 CPU's. WE're talking precision. This is why PPU's cannot be used and this is also what bogs down current nVIDIA GPU performance.

I really would love to see this in all it's glory. Finally we'd be able to use the x1K's Dedicated Branching unit for truly precise computing.

One thing does bug me though... the electrical costs. Running a CPU at 100% isn't nearly as bad as running two x1K's at 100%. I wouldn't be surprised if this could be used as a replacement for a home furnace during the cold winters up north.
 
Yeah, I didnt know of the alternate meanings...

I agree, its a pretty stupid word.

Oh well, I stand corrected :)
 
A GPU can render a scene faster then 10 CPU's. WE're talking precision.
Click to expand...

I was not speaking of rendering in the same way you play half life...

read up on gelato if you don't understand, gelato is a non-real time renderer... it does work like a software renderer, kindof (max/povray, that kind) except on the gpu. with all the precision too.

I think if the folding guys know what they are doing, then there shouldn't be any reason nvidia hardware would be slow vs ati hardware... However, if they are not integrating the work tightly with the CPU then who knows.
 
Yashu said:
I was not speaking of rendering in the same way you play half life...

read up on gelato if you don't understand, gelato is a non-real time renderer... it does work like a software renderer, kindof (max/povray, that kind) except on the gpu. with all the precision too.

I think if the folding guys know what they are doing, then there shouldn't be any reason nvidia hardware would be slow vs ati hardware... However, if they are not integrating the work tightly with the CPU then who knows.
Click to expand...
I'm sure they know what they're doing. As I've stated previously. The issue arises with the fact that the 7x00 seris GPU's don't have a dedicated branch execution unit. Dynamic branching is used for physics processing as well as complicated precision processing. It allows for dyanamic branch prediction processing.

I think it's great that they've decided to use a DX10 level feature on today's VPU/GPU's.

Xbitmark_x18.gif


x1800XT is twice as fast as a 7800GTX when Dynamic branching is concerned. x1900XT and XTX are about 2.5 - 3 times faster then the x1800XT.

x1800 level hardware was designed with a default thread size of 16x16.. but it can calculate with EXTREME precision at 4x4 thread size (with a small performance penalty of course).

So I believe Stanford opted to use the High precision power of teh x1K architecture in this manner. Pixel shader units impact branching performance.. not simple clock rates.. so a 7900GTX should score close to the 7800GTX number shown in the graph.

Oh I found a graph that highlights this...

Look at the 7900GTX vs x1900XTX numbers. (The last 3 are Branching indicators.. the guy who made the benchmark program has stated that the x1900 series are almost on idle during teh shader tests... in other words you're not seeing the x1900's true shader potential.. in texturing tests.. it's a definite win for the nVIDIA cards.. 24 vs 16 Texturing units).
xbitmark.gif

Xbit Labs

This is the power Stanford tapped into for Folding. I mean at 4 times the power of a 7900GTX the x1900XTX definatly flexes it's muscle when computing with high precision.
 
I understand what you are saying, but that is why I meantioned tightly integrating with the CPU.

you are assuming that *everything* is going to be done on the GPU.


any of this is speculation, however, what will happen will depend on what kind of SDK ati or nvidia provides.

I cited Gelato as an example because it is already here and being used. It is the first real use of a GPU as a general processor... it is in the wild now and it works.

why would they want to limit this to ATI hardware anyway? they wouldn't. it is in their interest to get this working on as many platforms as they can.
 
Yashu said:
I understand what you are saying, but that is why I meantioned tightly integrating with the CPU.

you are assuming that *everything* is going to be done on the GPU.


any of this is speculation, however, what will happen will depend on what kind of SDK ati or nvidia provides.

I sited Gelato as an example because it is already here and being used. It is the first real use of a GPU as a general processor... it is in the wild now and it works.

why would they want to limit this to ATI hardware anyway? they wouldn't. it is in their interest to get this working on as many platforms as they can.
Click to expand...

They're not limiting it to ATi.. as the original posts said.. they're working with both nVIDIA and ATi. But folding requires EXTREMELY precise calculations (something Gelato does not). Hence why Folding takes forever on a CPU.

This is where the Dedicated Branch preditor unit comes in. It offers EXTREME precision. Stanford cannot just throw in some low precision calculations and take teh results seriously.. this is, afterall, scientific research.

If you preffer nVIDIA cards.. don't sweat it.. G80 will come with a dedicated branch preditor unit. 7x00 hardware will run it... just much slower then a comparable system running an x1900. It's an architectural difference.. not an SDK difference. Odds are it will be written in generic Direct3D.
 
how precise are we talking... 64bit? 128bit? 256bit?

I read somewheres that the x1X00 uses 512bit internal math... that is pretty precise. is this why you think the client will favor ATI?

I am not a fan of either camp more then the other... I am thinking of consumer penetration; there are just more high end nvidia cards out there. with tight integration with the CPU, both chipsets will be fast and stanford wins.

what about cross platforms... I checked the F@H site, and there are several cross platform options... the only way to keep compatable with mac and linux would be to NOT use direcX, do ATI's advantages apply outside of directX?
 
Negatory, a gpu is the one thing I cant live without. Folding would put my gpu at 100% load and decrease it's life. Most GPUs have a shorter ETOL (estimated time of life) than a cpu. I'll stick to cpu folding. (boinc) :D
 
You must log in or register to reply here.
Back
Top