GPU client

[Devaster]

i see that in some cases is GPU application better than stock app ....

interesting ....

[Radiohead]

i see that in some cases is GPU application better than stock app ....

interesting ....

Hm...
This is 8800GTX 

[guido.man]

There are too many variables to make an accurate assessment of the true speedup of GPU client.
When I use default-515.exe as the reference client I also get Speedup Ratio  greater than 1.00 for some WU's.
Devaster what source files are you using to base your GPU client on, and what optimizations are done?
If you compiled a reference client to be used against your GPU client then most of the variables would
be known.

[Devaster]

at home by tests i am using latest optimized app from  cruncher page and i have 8500GT only ....

remember , for now aren't in GPU code optimizations (shared mem usage ,memory coalescent access by read/write, optimal thread/block scheduling against core, ). used is only partial loop unroll ...

[Devaster]

last weekend i have downloaded GUI profiler for CUDA. It has showed  many interesting things .... i will write more later - it would bigger ...

[Gecko_R7]

Hi Mimo,

Just saw this recent update on the Cuda developer forum.

http://forums.nvidia.com/index.php?showtopic=34241

Any benefit to your efforts?

[Devaster]

SETI@HOME uses 1d FFT and this is batched naturally in CUFFT library ....

[Gecko_R7]

SETI@HOME uses 1d FFT and this is batched naturally in CUFFT library ....

SETI@HOME uses 1d FFT and this is batched naturally in CUFFT library ....

Sorry  .  My inquiry/interest was related to the batching, not the 2d fft itself since as you pointed-out, Seti uses 1d fft single precision complex ^2.
Didn't realize the batching was already done in the CUFFT library however.
Just ignore "the little man from behind the curtain".....

[Devaster]

try

[attachment deleted by admin]

[Radiohead]

What has changed?

[Radiohead]

New data...

[attachment deleted by admin]

[popandbob]

my first run...

8600GTS

Cpu and vid card not OC'ed.

~BoB

[attachment deleted by admin]

[Devaster]

- changed data alignment to multiples of two (float2,float4) -some speedup
 

[Devaster]

my first run...

8600GTS

Cpu and vid card not OC'ed.

~BoB

more multiprocessors and more MHz as me and better results ...

[Batstat]

My result, open attachment

Result      : weakly similar.   
Result      : weakly similar.   
Result      : DIFFERENT. 
Result      : weakly similar.   
Result      : DIFFERENT. 
Result      : weakly similar.   
Result      : weakly similar.   
Speedup: 5.88%, Ratio: 1.06 x
Speedup: 9.83%, Ratio: 1.11 x
Speedup: 11.48%, Ratio: 1.13 x
Speedup: -31.30%, Ratio: 0.76 x
Speedup: 10.21%, Ratio: 1.11 x
Speedup: 24.52%, Ratio: 1.32 x
Speedup: 28.01%, Ratio: 1.39 x

[attachment deleted by admin]