no offense, but i like nvidia cards so much better. i had trouble with ati drivers... haven't had trouble with 190.38 driver and i deduce that luck to the fact that i bought a BFG card!!!!!!
i have a free 2600 XT card laying in a box.... the fan on it is weak, the heatsink sucks and it gets so damn hot when running, it could melt plastic!!!!
currently, their are projects that are starting to look at ati cards; why u might ask, the more cards and cpu's; the better for all of us.
no offense, but i like nvidia cards so much better.
Howdy ZPM,
Hey, that is perfectly understandable. I do not ‘Hate’ Nvidia cards. The company currently makes some great GPU’s !
I am sure I would be happy with a GTX 260, 275, 285, etc.
That said, I have only owned Two Nvidia cards in the last 12 years.
In 1998, I bought an original TNT 1 Board ! Which I liked. It was a
great card for it’s day.
In 2005 I bought a GS 7200 which, I did not like.
Other than those Two, I have used mostly ATI boards for the last 10 years.
This can become like a ‘Ford’ vs ‘Chevy’ debate. With one preferring one brand over the other.
So__ will I ever personally ‘change sides’ and start buying Nvidia in the future. ( quite possibly )
I have heard that their new GT 300 chip
which has just taped out, will be a powerful distributed computing cruncher !
Does Einstein@home use double precision crunching?
If not the double precision crunching in hadware capability
of the ATI card will have no advantage.
Given that the primary operation is a very long FFT of low-bit noisy data, I would think ABP crunching could easily work with single precision; I think even the folding for the test binary orbits wouldn't suffer since the total duration of the beam pointing is fairly short.
The LIGO signals may be different. Again, one is looking for something weak buried in noise so the quantization noise from single precision math shouldn't hurt much. The trickier math may again be the time transformations for the trial sky location being searched. That might require double precision or other precision-extending gymnastics.
"Better is the enemy of the good." - Voltaire (should be memorized by every requirements lead)
Does Einstein@home use double precision crunching?
If not the double precision crunching in hadware capability
of the ATI card will have no advantage.
Given that the primary operation is a very long FFT of low-bit noisy data, I would think ABP crunching could easily work with single precision; I think even the folding for the test binary orbits wouldn't suffer since the total duration of the beam pointing is fairly short.
The LIGO signals may be different. Again, one is looking for something weak buried in noise so the quantization noise from single precision math shouldn't hurt much. The trickier math may again be the time transformations for the trial sky location being searched. That might require double precision or other precision-extending gymnastics.
Exactly. The ABP1 is already running in beta test on single precision CUDA hardware, so that's not a problem. For the GW app, it's indeed trickier to eliminate double precision stuff but there's hope this can be done as well without loosing to much of the GPU-gain you would expect. Stay tuned.
i've got 2 9400 gt's that will be here to crunch if they are allowed(unlike milkyway).
Milkyway needs double precision in the CUDA code so that's probably the reason why you can't use those cards there. The ABP1 CUDA beta apps do not require this and should run fine on your boards .... of course you need a CUDA capable driver installed.
I think crunching in single precision is only a temporary solution. The accuracy of calculation should be as high as possible and IMO, that could not be reached with single precision.
I believe Einstein@H will come to the same conclusion like Milkyway befor. From this point nVidia-cards lower than a GT200 are a no-go and as long as i read, all actually AMD/ATI-cards supports double precision.
I think crunching in single precision is only a temporary solution. The accuracy of calculation should be as high as possible and IMO, that could not be reached with single precision.
I believe Einstein@H will come to the same conclusion like Milkyway befor. From this point nVidia-cards lower than a GT200 are a no-go and as long as i read, all actually AMD/ATI-cards supports double precision.
I would not agree with this. Obviously it depends on the math that is required for the individual project, but there is no general rule that double precision HAS to yield better results at all. If you are dealing with input data that itself has low precision because of the wy it was collected and the algorithms that do the processing are insensitive to small perturbations, double precision won't give you anything. If you are lucky, you can also split the computation into parts that can be done in single porecision and those that really require double precision, and if you are really lucky it turns out that the DP part can be run on the CPU and the SP part on the GPU in a way that still gives a good speedup. So I think SP in scientific computations will not go extinct anytime soon.
I read some time ago somewhere about a scientist who had been asked about single- and double precision. He stated, that the double precision requirements are mostly there due to lazyness to make the formula for computation exact enough.
I think I have a very very vague idea what he mean whith that. In nautical school one teacher did write a stability program in excel. We had the task to let excel calculate a formula discribing a graph. Don't remember the type of formula anymore. The max what excel can do on that is to the fifth grade. For some parts that was not enough. Then I increased the decimal places on my part, and then it was enough. So, instead of doing the 6th grade of the formula by hand, we increased the precision by adding more decimal places.
The scientist said, that probably only 10% of all the computation requiering actualy double precision realy need it. The others need a higher grade of the formula and are after that happy with single precision. That's at least, how I understand all that.
I read some time ago somewhere about a scientist ..... at least, how I understand all that.
He'd be referring to one function approximating another, like a polynomial expansion calculating a sine. To give a closer answer ( to the true sine value ) one could keep the same number of polynomial terms but use increased precision for each OR add more of the lower precision terms. It all depends on how the method has been matched to the problem. Double precision might only achieve more decimal places on a wrong answer, and terms don't necessarily contribute with the same sign.
For the GW WU's at E@H I guess the question would be whether the F-statistic is sufficiently sensitive to precision to be worried about. Thus if a precision change makes a 25.34 become a 25.33, that's quite different to a 25.34 becoming a 12.34.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
no offense, but i like nvidia
)
no offense, but i like nvidia cards so much better. i had trouble with ati drivers... haven't had trouble with 190.38 driver and i deduce that luck to the fact that i bought a BFG card!!!!!!
i have a free 2600 XT card laying in a box.... the fan on it is weak, the heatsink sucks and it gets so damn hot when running, it could melt plastic!!!!
currently, their are projects that are starting to look at ati cards; why u might ask, the more cards and cpu's; the better for all of us.
I recommend Secunia PSI: http://secunia.com/vulnerability_scanning/personal/
RE: no offense, but i like
)
Howdy ZPM,
Hey, that is perfectly understandable. I do not ‘Hate’ Nvidia cards. The company currently makes some great GPU’s !
I am sure I would be happy with a GTX 260, 275, 285, etc.
That said, I have only owned Two Nvidia cards in the last 12 years.
In 1998, I bought an original TNT 1 Board ! Which I liked. It was a
great card for it’s day.
In 2005 I bought a GS 7200 which, I did not like.
Other than those Two, I have used mostly ATI boards for the last 10 years.
This can become like a ‘Ford’ vs ‘Chevy’ debate. With one preferring one brand over the other.
So__ will I ever personally ‘change sides’ and start buying Nvidia in the future. ( quite possibly )
I have heard that their new GT 300 chip
which has just taped out, will be a powerful distributed computing cruncher !
Best Regards,
Bill
RE: Does Einstein@home use
)
Given that the primary operation is a very long FFT of low-bit noisy data, I would think ABP crunching could easily work with single precision; I think even the folding for the test binary orbits wouldn't suffer since the total duration of the beam pointing is fairly short.
The LIGO signals may be different. Again, one is looking for something weak buried in noise so the quantization noise from single precision math shouldn't hurt much. The trickier math may again be the time transformations for the trial sky location being searched. That might require double precision or other precision-extending gymnastics.
"Better is the enemy of the good." - Voltaire (should be memorized by every requirements lead)
RE: RE: Does
)
Exactly. The ABP1 is already running in beta test on single precision CUDA hardware, so that's not a problem. For the GW app, it's indeed trickier to eliminate double precision stuff but there's hope this can be done as well without loosing to much of the GPU-gain you would expect. Stay tuned.
CU
Bikeman
i've got 2 9400 gt's that
)
i've got 2 9400 gt's that will be here to crunch if they are allowed(unlike milkyway).
I recommend Secunia PSI: http://secunia.com/vulnerability_scanning/personal/
RE: i've got 2 9400 gt's
)
Milkyway needs double precision in the CUDA code so that's probably the reason why you can't use those cards there. The ABP1 CUDA beta apps do not require this and should run fine on your boards .... of course you need a CUDA capable driver installed.
HC
Bikeman
I think crunching in single
)
I think crunching in single precision is only a temporary solution. The accuracy of calculation should be as high as possible and IMO, that could not be reached with single precision.
I believe Einstein@H will come to the same conclusion like Milkyway befor. From this point nVidia-cards lower than a GT200 are a no-go and as long as i read, all actually AMD/ATI-cards supports double precision.
RE: I think crunching in
)
I would not agree with this. Obviously it depends on the math that is required for the individual project, but there is no general rule that double precision HAS to yield better results at all. If you are dealing with input data that itself has low precision because of the wy it was collected and the algorithms that do the processing are insensitive to small perturbations, double precision won't give you anything. If you are lucky, you can also split the computation into parts that can be done in single porecision and those that really require double precision, and if you are really lucky it turns out that the DP part can be run on the CPU and the SP part on the GPU in a way that still gives a good speedup. So I think SP in scientific computations will not go extinct anytime soon.
CU
Bikeman
I read some time ago
)
I read some time ago somewhere about a scientist who had been asked about single- and double precision. He stated, that the double precision requirements are mostly there due to lazyness to make the formula for computation exact enough.
I think I have a very very vague idea what he mean whith that. In nautical school one teacher did write a stability program in excel. We had the task to let excel calculate a formula discribing a graph. Don't remember the type of formula anymore. The max what excel can do on that is to the fifth grade. For some parts that was not enough. Then I increased the decimal places on my part, and then it was enough. So, instead of doing the 6th grade of the formula by hand, we increased the precision by adding more decimal places.
The scientist said, that probably only 10% of all the computation requiering actualy double precision realy need it. The others need a higher grade of the formula and are after that happy with single precision. That's at least, how I understand all that.
Christoph
Greetings, Christoph
RE: I read some time ago
)
He'd be referring to one function approximating another, like a polynomial expansion calculating a sine. To give a closer answer ( to the true sine value ) one could keep the same number of polynomial terms but use increased precision for each OR add more of the lower precision terms. It all depends on how the method has been matched to the problem. Double precision might only achieve more decimal places on a wrong answer, and terms don't necessarily contribute with the same sign.
For the GW WU's at E@H I guess the question would be whether the F-statistic is sufficiently sensitive to precision to be worried about. Thus if a precision change makes a 25.34 become a 25.33, that's quite different to a 25.34 becoming a 12.34.
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal