GPU crunching question

[citroja]

  There is some serious debate about using a GPU to crunch SETI.  From what I understand there are works in progress to enable the nvidia 8800 cards to crunch....

  I currently have a 7800GTX card (obsolete) in my machine and I was talking to someone about upgrading.  They said that as long as I stick with a PCIe card I could just add any other nvidia card to the other PCIe slot (7300, 7600, 7900, 8800)does anyone know if this is true???

  I was seriously considering adding another card and if I can I will go for the 8800 so that I can crunch with it.  The only problem that i have thought of is that the 8800 is a directX 10 card and the 7800 is a directX 9 card.

  if anyone has any thought please let me know.

[EDIT]
Here is the S@H forum posts about GPU crunching

http://setiathome.berkeley.edu/forum_thread.php?id=36242
[/EDIT]

-citroja

[pepperammi]

From what I read you need a pretty heafty CPU to feed that 8800 and get the best out of it.
Also take note if your Powersupply can feed those monsters too and if you have room for a card that uses an exhaust that need an extra slot.

I'd love one myself but so expensive 

[citroja]

From what I read you need a pretty heafty CPU to feed that 8800 and get the best out of it.
Also take note if your Powersupply can feed those monsters too and if you have room for a card that uses an exhaust that need an extra slot.

I'd love one myself but so expensive 

I have an AMD X2 4800+, 600W PSU, and a free place for exhaust (but it won't be needed with a liquid cooling unit).  What is more important for me to figure out is if:

a) my 7800GTX will be able to crunch
b) my 7800GTX is compatible with any of the current nvidia PCIe card (haven't found anything online yet)
c) my 7800GTX will be compatible with the new DirectX 10 card (8800 and beyond)

but I suppose that this is all null and void until there is a proven GPU cruncher available.  If anyone needs help testing (7800 series) let me know....

-citroja

[citroja]

After some long research i found that as of right now you CANNOT mix and match SLI cards by type (i.e 7800GTX must be paired with another 7800GTX) it doesn't matter if one is overclocked or not.  Theoretically (and with some patching) the same cards with different memory (256 vs. 512) can be paired to run at the lower settings but it is not recommended.

I have not found anything that said you can't have a 7800 and say a 7900 in the same system, from what I can tell they just can't be SLI configured (at least as of right now).

For more info this is from the nvidia site:
http://www.slizone.com/page/slizone_faq.html

For those of you with only a SINGLE (obsolete) GPU if you want a match look for it on ebay....especially with the new DirectX 10 cards coming...people (read as 'gamers') will begin to upgrade their rigs and dump the older cards.

-citroja

[Devaster]

i have modified sah application to use GPU for FFT , power spectrum and chirping using BrookGPU. i think i may release this on next weekend....

[Gecko_R7]

i have modified sah application to use GPU for FFT , power spectrum and chirping using BrookGPU. i think i may release this on next weekend....

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

[Devaster]

first benefit is this that a GPU and CPU version can run pararel with a small performance hit

[BenHer]

Here's some thoughts...

If you had an SLI machine, could you run 1 copy CPU, and 1 copy of GPU for each card?

Is the GPU kept busy enough by the version, or is the GPU idle some percentage of the time?
If somewhat idle, on a multi-core or multi-cpu system can multiple GPU copies be run?

Weee!

[Devaster]

sli mode is based on driver if is enabled i think it would running but there is no a way to do running one prog per GPU separate...

by gpu version is all heavy calc doing on gpu therefore is cpu freed from workload. this free time can be used by cpu version....

multicore or multicpu has nothing with this  you use a multiple gpus as one ,all distributing per gpu is done via driver !!!!

read more at http://gpgpu.org/

[Devaster]

ok for now some link problems ....

btw. wanna see part of  ps30 shader code ? :

Code:

namespace {
using namespace ::brook::desc;
static const gpu_kernel_desc __DFTX_ps30_desc = gpu_kernel_desc()
.technique( gpu_technique_desc()
.output_address_translation()
.input_address_translation()
.pass( gpu_pass_desc(
"    ps_3_0\n"
"    def c26, 0, 0.5, 1, 2\n"
"    def c27, -1, 1, 0, 0\n"
"    dcl_texcoord1 v0.xy\n"
"    dcl_2d s0\n"
"    dcl_2d s1\n"
"    dcl_2d s2\n"
"    frc r0.xy, v0\n"
"    add r0.xy, -r0, v0\n"
"    mov r1.xy, c26\n"
"    dp2add r0.z, r0, c20, r1.y\n"
"    dp2add r0.x, r0, c20, r1.x\n"
"    mul r2, r0.z, c22\n"
"    frc r3, r2\n"
"    add r2, r2, -r3\n"
"    mad r0, r2, -c21, r0.x\n"
"    add r0, r0, c26.y\n"
"    mov r2, c23\n"
"    mad r0, r0, r2, -c24\n"
"    frc r2, r0\n"
"    add r0, r0, -r2\n"
"    cmp r2, r0, c26.x, c26.z\n"
"    dp4 r1.x, r2, r2\n"
"    cmp r1.x, -r1.x, c26.x, c26.z\n"
"    mov r2, -r1.x\n"
"    texkill r2\n"
"    add r2, r0, -c25\n"
"    cmp r2, r2, c26.z, c26.x\n"
"    dp4 r1.x, r2, r2\n"
"    cmp r1.x, -r1.x, c26.x, c26.z\n"
"    mov r2, -r1.x\n"
"    texkill r2\n"
"    mad r2, r0, c0, r1.y\n"
"    mul r2, r2, c1\n"

 

nice isnt ? (whole code is about 8000 lines long)

[Devaster]

Is the GPU kept busy enough by the version, or is the GPU idle some percentage of the time?

for now i dont have tested it i must install nvperfhud for doing it ....

[Simon]

Good luck, sounds interesting!

Keep us posted.

Regards,
Simon.

[citroja]

need any test machines (or cards)??

7800GTX ready and waiting

-citroja

[Devaster]

for now changing from brook to rapidmind backend ... be patient

[Devaster]

here are numbers from fft test under rapidmind :

RapidMind 2D FFT Benchmark
===============================================
Size: 256 x 256 = 2^8 x 2^8
Radix: 4 = 2^2
Total number of floating point operations: 5.24288e+006

Run timings, to and from host (in ms):

Average execution time: 11.8233ms
Overall average execution time: 11.8287ms
Minimum execution time: 10.8905ms
Average Mflops: 443.437
Peak Mflops: 481.419

Run timings, GPU-local (in ms):

Average execution time: 9.78332ms
Overall average execution time: 9.7893ms
Minimum execution time: 9.24811ms
Average Mflops: 535.9
Peak Mflops: 566.913

nice isnt ? (nvidia6600 300/600 MHz)

If anything simply cannot go wrong, it will anyway.

- Murphy's Law