How GPU's works?

[novasoft]

Hi, i just wonder how a GPU work, if anyone knows because i tought about Nvidias PhysX thing and CUDA and so on. And i wonder how the cards calcuates the physics and other stuff on the cards. Is it Like this: (for exampel) GF8500GT has 16 Shader-CPU's and the Shader-clock is on 900MHz. So is it like this: 900 x 16 = 15600 = 15.6GHz.. It looks kinda much for a cheap card, but could it be that!?. (I know that GPU's don't work the same way as CPU's!)

Or does they use the Core-clock? Or everything on the Card!? And i read that PhysX needed atleast 256MB VRAM, does the RAM do much difference!?

If anyone know this and could give me an answer because i have wondered how the things works, i would appreciate it really much! Because i think this things it really interesting!

 

[TheGreatGrapeApe]

It's called Evil Kid

 

[Freeduck]

Maybe this explanation will do? http://fastra.ua.ac.be/en/index.html

 

[novasoft]

Yeah, it helped a little. Thanks.
But it wasn't enough.

 

[-Fran-]

There's an article in TH that talks about CUDA and explains very well how the GPU handles all the work loads and describes a lot of things.

That could be a little dense to read, but u can have wikipedia on a second window/tab =)

Esop!

PS: http://www.tomshardware.com/reviews/nvidia-cuda-gpu,1954.html

 

[novasoft]

Oh thanks..

Hehe.. I Didn't think about Wikipedia.. >.<
Fail on me..

 

[FrozenGpu]

Epic fail... :lol:

for the most part Wikipedia is ur friend, if ur not sure about the reliability you could always check back here on more specific information concerning specifications and compatibility, etc, etc, etc, but a site called how things work, I think its called that helps too.

 

[TheGreatGrapeApe]

Anywhoo, related to the focus of your inquiry, it's the shader clock that would matter, the physics calculations can be done in 2 ways, as stream processor workloads (likely but not necessary cause CUDA would let you access other no-GPGPU-traditional components as well), or it could use vertex and geometry components to run the physics vector calculations.

Anywhoo, both reside in the shader part of the GPU.

There isn't a whitepaper on it yet, but here's they're GDC presentation which mentions the SPUs;
http://developer.download.nvidia.com/presentations/2008/GDC/GDC08_GPUPhysics.pdf

 

[novasoft]

Ok thanks to you too..

Now i understand some of the things even better.. yay! Hehe..

 

[novasoft]

And just one more thing.. Uhmm.. If the physics uses the Shaders, then can i underclock the Core and Memory speed so the card gets cooler (if it works!?) and then Overclock the shaders more. Or will The Card get slower Because of the Core and Memory underclock? (The card will be Dedicated for only physics and stuff, so it will not use any graphical things at all.)

 

[-Fran-]

U should note that when we talk about GPGPU, the cards RAM becomes some-what standard RAM in terms of practical mapping in CUDA (and OpenCL too i think). So underclocking the GPU and RAM of the card would affect it's performance.

Now, i don't know how much it would be, and since CUDA is very programmer-optimized code, it's a really subjective point to discuss =/

And i also don't have a clear picture about optimizations in CUDA, so i could be omfg-wrong about it, lol.

Esop!

 

[TheGreatGrapeApe]

Yeah Yuka is right, the VRAM is important (although not sure how bandwidth sensitive PhysX would be) ; you could underclock the core, but you'd want to leave the memory where it is (doesn't save you much power and heat wise anyways) and also OC the shaders, but the core with the ROP and TMUs shouldn't matter for the physics component of the equation.

 

[novasoft]

Ok, great.
Thanks once again. x)