Posts by HA-SOFT, s.r.o.

61) (Message 2349)
Posted 2 Jan 2014 by Profile HA-SOFT, s.r.o.
Post:
Thanks for your answer - I wonder how the problem will be solved.

Nevertheless, I'm surprised because the effect described above did NOT occur with the foregoing cuda45 wu's, and it is NOT occuring with all other GPU-appliances like Einstein-, Seti-GPU-wu's or movies using the GPU.

The effect of extreme slow reactions occures ONLY with Boinc-Cuda55-Asteroid-wu's.

Regards, Piquadrat


As I look at your wu's, the slowest kernel takes about 70 ms which should be ok. I think the lag is related to memory controller load. Will see with next version. I have moved some arrays to texture memory.
62) (Message 2346)
Posted 2 Jan 2014 by Profile HA-SOFT, s.r.o.
Post:
On these slow cards like GT4xx there is a lag experience. We know about it

We are investigating posibilities how to add parameter to limit card usage (and of course increase total runtime).
63) (Message 2345)
Posted 2 Jan 2014 by Profile HA-SOFT, s.r.o.
Post:
I have not tried mine here yet but I ALWAYS leave a cpu core free when using my gpu's, unless I see in the very low cpu % usage while the gpu is crunching. For instance my 7970, on another project, is using 0.84% cpu and I do have a cpu core free just to keep it fed and running as fast as possible. If I change it to use all cpu cores for crunching my gpu crunch times go up.


I do this for GPUGRID only. Other project like DistrRTGen and Asteroids can live without dedicated CPU core.
64) (Message 2344)
Posted 2 Jan 2014 by Profile HA-SOFT, s.r.o.
Post:
I got much higher times on 560TI 384 cores then written down here. Does it need a free CPU Core? It ran at 99% Load, but i know this doesnt mean much ;)


It depends on wu and some users have overclocked cards.
65) (Message 2336)
Posted 31 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
GTS450 ~ 14217.13s
Averaged over three results.
Standard deviation 4.33032


Thanks for info.
66) (Message 2334)
Posted 31 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
To all GPUs fans.

This is first public version and our development do not stop with this version.

Next version update will:

1. Use less memory. (example Titan 1100MB -> 688MB). Little speedup.
2. I will try to move some memory arrays to textures with int2 texels. This may speedup app.

PS: It looks like we made to much fast CPU apps :-D
67) (Message 2333)
Posted 31 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
The time you reported for your Titan has a note in parentheses beside it... (recommended CUDA double precision disabled). That confuses me because I thought A@H GPU tasks would benefit from DP but it appears you recommend disabling it. I was planning on hacking my GTX 660Ti to unlock it's DP capability, as was discussed about a month ago, but now it seems like that would be a waste of time.


It's because the time of calculation is divided half-and-half between memory load and dp calculations. The second point is that we mostly use multiply-add ops which are not so agresive.
68) (Message 2332)
Posted 31 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
Lots of diplomatic talk about GPU vs. CPU credit.....
To the point:

1. The GPU credit is ridiculously low (by 5x to 10x).
2. You are missing a very big opportunity (due to low GPU credit)to use the new GPU app to attract the big GPU crunchers into an important project.

xyzzy


1. I agree with Dagorath. Same work = Sam credit. Yes I know that power consumption of GPUs in compare to credit gain is disadvantageous.

2. I think if we improve GPU app the things will be better, but we can never beat DistrRtGen or GPUGrid in credit.
69) (Message 2325)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
The performance on a highend card is also a bit disappointing. My GTX Titan seems to need nearly one hour for a workunit.

That's why i'm looking forward to the OpenCL version with the very much stronger AMD GPUs.


I'm affraid it will be the same problem. Only 79xx card looks promising.
70) (Message 2321)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
I see, nicely overclocked. Just curious: Is this reference design cooling? (mine is)


Yes. Reference design from Zotac.
71) (Message 2319)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
Těžko říct, nepřehřívá se?
72) (Message 2317)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:


EDIT: Not quite. 46 min. @ 980 MHz GPU and ~ 74°C now.


My at 1094Mhz ~60 °C fan full speed

see

http://asteroidsathome.net/boinc/result.php?resultid=27987109

But there may be different wu also.
73) (Message 2312)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
Times of Titan are somewhere about 36 min. Not to much comparing to Intel CPU's even when we have 100% card usage.
74) (Message 2310)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:

Need smaller data chunks for lesser screen lags, or make it adjustable like on MilkyWay.


Grid size is calculated from number of SMX. You can look into stderr.txt, there are times of the largest kernel.

Times for GTS450 are about 100 ms:

CUDA Device: GeForce GTS 450
Compute capability: 2.1
Multiprocessors: 4
Grid dim: 32 = 4*8
Block dim: 128
mrqcof1 curve1 kernel time: 102 ms
mrqcof1 curve1 kernel time: 32 ms


100 ms should be relatively ok.
75) (Message 2309)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
So I discontinue this project if you don't bother about mid-range cards.


Simply, It's not our decision. It's hw which limits our posibilities.
76) (Message 2307)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
31.12.2013 0:36:37 |  | CUDA: NVIDIA GPU 0: GeForce GTS 450 (driver version 327.23, CUDA version 5.5, compute capability 2.1, 1024MB, 826MB available, 637 GFLOPS peak)



It's very slow card and we have high memmory controller load (cca 40-60%) also. This may be the cause of lags.
77) (Message 2306)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
Nice to hear all is ok.
78) (Message 2302)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
It works! Deleted the project and added it again.
But with strong screen lags...


Which is your card?
79) (Message 2300)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
No this mean your client do not request them.
80) (Message 2298)
Posted 30 Dec 2013 by Profile HA-SOFT, s.r.o.
Post:
Please, try again. Kyong has made some changes on server side.


Previous 20 · Next 20