GPU crunching question

[citroja]

Any idea where I can get a list of those cards???

or even just the series of cards???

-citroja

[Devaster]

nvidia : 6k,7k and G80 series,Quadros with ShaderModel 3.0
ATI : all1x00

[Devaster]

interesting thing:
i have try replace a FFTGPU with a rapidmind implementation .... result ?

by standard test wu has seti crashed every time on a memory write error . it consumes a 1,5 GB (one and half gigabyte) of memory and want more .... 
why ?
there is a function generating a twiddle arrays - this arrays are with every step greater and greater (exponencially)  max size of fft on my machine (2 GB with vista) is 2 upto  20 ....
so i must stay on FFTGPU ...

[Devaster]

Next performance issue : power spectrum calculated on GPU is (i dont now why) VERY slow ...
Rapidmind perf issue ? maybe ....

[zac]

Very interesting.
i wanna know What kind of results are you seeing vs. CPU-only performance?

[Devaster]

optimized CPU versus my GPU
GPU powerspectrum is veeeeeeeeeery slow , but  i think i have found why : context switching.

i initialize rapidmind before client graphics and i am running under vista with aero ....
so i think there is doing next :
rapidmind must upload all data ,calculate and download all data. when awro need GPU and graphics memory then in middle of calculation again - stop calc, data download, aero upload ,calc,aero download, data upload and continue calc .....

[Simon]

Ouch!

Good thought though, sounds very plausible, especially since Vista likes to poll ALL hardware every few microseconds. Have you tried with XP/XP64?

Regards,
Simon.

[Devaster]

i havent installed it 
but i think i will return to old good wxp...

[Vyper]

Switch of the Aero interface and the Gfx is running in 2d mode (I know i've checked)
..

I have modified a 8800 bios with throttling and 2d mode, it works and i've checked
that atitool reports lower mem/gpu clocks and if it does it runs in 2d only mode leaving all GPU cycles to your program..

Try again before reinstalling

Kind Reg Vyper

[Devaster]

maybe another reason : MS in Vista has fu***d off any opengl support and rapidmind is running on opengl ....

i compile a brook version with all four backends (cpu,opengl,dx9,ctm) and i will see ....

[Vyper]

Yea do so, btw i see u like Duke nukem..

If u search the Mule network for duke nukem in a rar package u can find the duke nukem midis that i have personally  retuned with real soundbanks to download if u want some flashbacks from that time and game

Duke Nukem Midi Collection Complete in Mp3.rar

//Vyper

P.S Opengl support works its albeit abit slower than in XP.. D.S

[ThierryH]

Hello all

I found a way to divide by 2 or 3 seti wu crunch time with CUDA. I didn't code since more than 15 years. It could take me very long time to do it. So, I'm searching interesting persons to work with me on it.

The way to do:
Like Devaster saw it, there's no gain to expect to try GPU crunching on "regular" functions where we generally concentred our optimizing effort. But there's still one function where we done a very poor job bacause it's probabely not possible to do better: find_pulse. With very good job made on other functions, find_pulse is now running more than half time. The goal is to parallelize this function on CUDA to excecute it 128 times simultaneously and divide his excecution time by 128, or more because this kind of calculations is normally more efficient on GPU. Do to that, we have to find a new way to call the function, and a way to call ReportPulseEvent with results inside GPU.

Hope some of you will follow me in this attempt to dramaticaly reduce find_pulse excecution time

[Vyper]

Hi

Yea Cuda api on the 8800 series have 128 shader units on the GTX and 96 on the TS..

But!! Cuda api only has 16 available GPU units on the GTX and 12 on the GTS, so yes there is a possibility to divide up the workload on 12 vs. 16 GPU units.

But as u said, the programming is intense and nowadays i cant program as well, it was different back then i programmed O/S friendly assembly programming on the Amiga O/S

Kind Reg Vyper

[currey]

i dunno if its possible but what about a tool that emulates all gpu cores to windows.
i know gpu cores dont have those things like sse eg. and are more simple than a normal cpu, but with that kinda tool you dont have to code a different code for every software gpu crunching is used for.

But the whole idea of gpu crunching is awesome. all boinc projects have a combined power of 502.201 TeraFLOPS by 250k active user. if only 20k use a fast gpu and are able to crunch with it these 20k create a gpu power of 750.000 TeraFLOPS (estimated to app. 30x speed up) AWESOME for distributive computing.

looking forward for your works, if ya need someone running test on a 8800gts, just let me know.

[ThierryH]

i dunno if its possible but what about a tool that emulates all gpu cores to windows.
i know gpu cores dont have those things like sse eg. and are more simple than a normal cpu, but with that kinda tool you dont have to code a different code for every software gpu crunching is used for.

It's the goal of CUDA. But it's a nVidia tool...