hoom
17-Aug-2008, 08:07
Over here (http://folding.typepad.com/news/2008/08/gpu2-wus-in-production-mode.html) they say that both ATI & Nvidia GPU2 clients are now doing production folding rather than just the test stuff.
ATIs have been on it about a month, NV since last Monday.
There is also another post that they are now allowing both ATI & Nvidia websites to distribute the client to encourage people to start folding.
Now if only they can start making proper use of my 4870...
Edit: also here is an explanation of the difference between what GPU folding is doing & what the CPU clients do:
http://foldingforum.org/viewtopic.php?f=50&t=4843&sid=8125210c5f17f105172526519eaa5d24#p48578
Protein folding occurs in a living cell or in a test tube, not in isolation but surrounded by molstly water, or, more accurately, aqueous solvent. In fact, the prevailing thought is that water is the major force driving proteins to fold, in that oily, hydrophobic bits of the protein (they really do look much like oil molecules) try to "hide" from the water (oil and water don't mix, right?) which gives the protein a stable "inside." Now it's really more complicated than that, which is why we need to do detailed calculations in order to try to understand folding, but that's one of the major ideas.
With water playing such an important role, we have to represent the solvent one way or another. The question is, do we have to represent all three atoms in every water molecule in protein folding simulations? If the answer is yes, this is a huge problem because to accurately model solvent, you generally need many more water atoms than you have protein atoms, so that typically the fun stuff (the protein folding) is only 10 % of the system size; in a simulation like that, 90 %* of the computation is spent on the less fun stuff, the water, which is important but for our purposes most of it is uninteresting. This is the "explicit solvent" case.
*This isn't perfectly accurate, since actually there are lots of interesting optimizations to get water calculations to go really fast on CPUs, but you get my point.
So why not come up with a way to represent the water not as individual hydrogen and oxygen atoms, but as a big blob capable of exerting forces on the protein? This is a much faster calculation than with explicit water since now 100 % of your system is taken up by the very interesting protein itself, and you don't have to calculate the forces and velocities for all the waters far from your system.
Problem is, no one knows how big of an approximation this big-blob "implicit" solvent model is. After all, real water on a molecular scale isn't a blob, but a collection of discrete molecules. One of the problems might be like what you say: the path the protein takes to go from being unfolded to being folded might change when we change from explicit to implicit solvent.
How protein folding might change if you change how the solvent behaves: For instance, say in order for a particular structure to form as the protein folds, a bunch of water molecules have to get out of the way (called "dewetting," which is something like evaporation. Some surface is "wet," ie, covered with water molecules, but in order for the surface to come into contact with another surface, the water has to scram). If you're representing water explicitly, fine, but if there are no discrete water molecules, the structure may end up forming much too quickly. Hence in that case the protein folding would be highly inaccurate.
With the power of GPU2 (and PS3 and SMP) we're finally able to address these kinds of questions ...
He also posted a comparison pic: GPU top, CPU bottom
http://www.stanford.edu/~densign/implicit_v_explicit.jpg
ATIs have been on it about a month, NV since last Monday.
There is also another post that they are now allowing both ATI & Nvidia websites to distribute the client to encourage people to start folding.
Now if only they can start making proper use of my 4870...
Edit: also here is an explanation of the difference between what GPU folding is doing & what the CPU clients do:
http://foldingforum.org/viewtopic.php?f=50&t=4843&sid=8125210c5f17f105172526519eaa5d24#p48578
Protein folding occurs in a living cell or in a test tube, not in isolation but surrounded by molstly water, or, more accurately, aqueous solvent. In fact, the prevailing thought is that water is the major force driving proteins to fold, in that oily, hydrophobic bits of the protein (they really do look much like oil molecules) try to "hide" from the water (oil and water don't mix, right?) which gives the protein a stable "inside." Now it's really more complicated than that, which is why we need to do detailed calculations in order to try to understand folding, but that's one of the major ideas.
With water playing such an important role, we have to represent the solvent one way or another. The question is, do we have to represent all three atoms in every water molecule in protein folding simulations? If the answer is yes, this is a huge problem because to accurately model solvent, you generally need many more water atoms than you have protein atoms, so that typically the fun stuff (the protein folding) is only 10 % of the system size; in a simulation like that, 90 %* of the computation is spent on the less fun stuff, the water, which is important but for our purposes most of it is uninteresting. This is the "explicit solvent" case.
*This isn't perfectly accurate, since actually there are lots of interesting optimizations to get water calculations to go really fast on CPUs, but you get my point.
So why not come up with a way to represent the water not as individual hydrogen and oxygen atoms, but as a big blob capable of exerting forces on the protein? This is a much faster calculation than with explicit water since now 100 % of your system is taken up by the very interesting protein itself, and you don't have to calculate the forces and velocities for all the waters far from your system.
Problem is, no one knows how big of an approximation this big-blob "implicit" solvent model is. After all, real water on a molecular scale isn't a blob, but a collection of discrete molecules. One of the problems might be like what you say: the path the protein takes to go from being unfolded to being folded might change when we change from explicit to implicit solvent.
How protein folding might change if you change how the solvent behaves: For instance, say in order for a particular structure to form as the protein folds, a bunch of water molecules have to get out of the way (called "dewetting," which is something like evaporation. Some surface is "wet," ie, covered with water molecules, but in order for the surface to come into contact with another surface, the water has to scram). If you're representing water explicitly, fine, but if there are no discrete water molecules, the structure may end up forming much too quickly. Hence in that case the protein folding would be highly inaccurate.
With the power of GPU2 (and PS3 and SMP) we're finally able to address these kinds of questions ...
He also posted a comparison pic: GPU top, CPU bottom
http://www.stanford.edu/~densign/implicit_v_explicit.jpg