This guy has 7 Fermi engineering samples?

[happy medium]

I don't know much about folding. Is this for real? I read it and it seems legit.


Link:http://www.xtremesystems.org/F...?p=4068556#post4068556

 

[ronnn]

Well known to be a poorly done fake. Here is the b3d forum guys take.

 

[Ben90]

Isnt F@H double precision? I remember from the AT article that they improved double precision multiple times over on the gt300 vs gt200. Even though this is fake, im wondering now if we might actually see something like this. Im not an expert but i would imagine if every calculation on F@H required double precision i would expect the gt300 to fold that much faster.

I really dont know how DP codes works though... is it something where absolutely everything gets it, or is it only some calculations are DP and the rest are SP?

 

[wlee15]

F@H GPU does not use double precision nor would it benefit much from double precision.

 

[TurtleCrusher]

Originally posted by: happy medium
I don't know much about folding. Is this for real? I read it and it seems legit.


Link:http://www.xtremesystems.org/F...?p=4068556#post4068556

I almost thought this was a Wreckage thread.

How did you see this before him?

 

[Painman]

*Painman saunters up to hot woman sitting at the bar*

"...I've got 7 Fermis..."

*Painman gets drink thrown in his face and a stiletto heel thrust into his crotch*

 

[Ben90]

Originally posted by: wlee15
F@H GPU does not use double precision nor would it benefit much from double precision.

Why is that? I was under the impression that gpus basically just estimated the answer and its not always 100% correct. For something as complicated as protein folding i would think one small rounding error would butterfly effect. Is it something where it already double checks answers in the code, or does 1 small miscalculation not really matter in the grand scheme of things, or something entirely different?

 

[Idontcare]

Originally posted by: wlee15
F@H GPU does not use double precision nor would it benefit much from double precision.

I would be totally astounded if this were true. Computational chemistry requires rigorous mathematical precision.

The energy changes brought about by changes in conformation (i.e. the process of the protein folding) involve taking the difference between two very large numbers and the rounding errors of single-precision would be disastrous to the error bars of such deltas.

 

[ViRGE]

Originally posted by: Ben90

Originally posted by: wlee15
F@H GPU does not use double precision nor would it benefit much from double precision.

Why is that? I was under the impression that gpus basically just estimated the answer and its not always 100% correct. For something as complicated as protein folding i would think one small rounding error would butterfly effect. Is it something where it already double checks answers in the code, or does 1 small miscalculation not really matter in the grand scheme of things, or something entirely different?

It's more complicated than that. Floating point precision is arbitrary; we could approach infinitely precise numbers, at a cost of infinite computational resources. There are multiple levels of precision, depending on what you need to do. For some things single precision (float) is precise enough, and for other things you're going to want to use a double or more. A higher level of precision isn't better if you don't need it.

In this case the work being fed to GPUs only needs single precision accuracy.

 

[Ben90]

OH! I was thinking something SP/DP were something completely different. My interpretation of it was it basically did the same calculation twice, shame on me for just assuming that based on its name. Lol, everything makes slightly more sense now

 

[GLeeM]

Originally posted by: wlee15
F@H GPU does not use double precision nor would it benefit much from double precision.

I recently read at F@H forum that there are no "double" GPU work units ... yet.

They were limiting GPU work to single (at least for now), as double would take eight times longer.

 

[happy medium]

Originally posted by: Scholzpdx

Originally posted by: happy medium
I don't know much about folding. Is this for real? I read it and it seems legit.


Link:http://www.xtremesystems.org/F...?p=4068556#post4068556

I almost thought this was a Wreckage thread.

How did you see this before him?

I'm no wreckage.

 

[bryanW1995]

Originally posted by: GLeeM

Originally posted by: wlee15
F@H GPU does not use double precision nor would it benefit much from double precision.

I recently read at F@H forum that there are no "double" GPU work units ... yet.

They were limiting GPU work to single (at least for now), as double would take eight times longer.

I would expect that with the vastly increase dp performance of fermi they will be able to run dp on it in the future.

 

[Idontcare]

Originally posted by: bryanW1995

Originally posted by: GLeeM

Originally posted by: wlee15
F@H GPU does not use double precision nor would it benefit much from double precision.

I recently read at F@H forum that there are no "double" GPU work units ... yet.

They were limiting GPU work to single (at least for now), as double would take eight times longer.

I would expect that with the vastly increase dp performance of fermi they will be able to run dp on it in the future.

I had no idea they were running Tinker on single-precision. I can't even begin to imagine the utility of the results they get from that combo. I'm sure there must be some utility, otherwise they wouldn't waste their time collecting the WU results for analyses, but I can't think of it at the moment. How many cancers have they cured so far?

 

[soccerballtux]

Originally posted by: Idontcare

Originally posted by: wlee15
F@H GPU does not use double precision nor would it benefit much from double precision.

I would be totally astounded if this were true. Computational chemistry requires rigorous mathematical precision.

The energy changes brought about by changes in conformation (i.e. the process of the protein folding) involve taking the difference between two very large numbers and the rounding errors of single-precision would be disastrous to the error bars of such deltas.

A friend talked me through it 3 years ago. Basically they only need something like 44 or 48 bits of precision. Curse my long term memory. I lose everything if I don't use it for a year.