GPU much faster than CPU [Yes, and no]

[chengbin]

I use a 8800GT and Q6600 to run F@H. I know that GPU was much faster than CPU but my 8800GT did the same amount of work in 3-4 minutes compared to what it takes about 12 hours to do on a Q6600 (smp 4 cores). Is F@H more optimized for GPU (or vice versa) or is CPU's computational power is miniscule compared to a GPU?

[Beremat]

I'm not 100% sure, but I'm almost certain that the GPU is actually that much faster in raw speed.
The 8800GT has many, many stream processors (somewhere around the 112 mark?) and can do work much faster.

[P5-133XL]

Yes, a GPU's computational capabilities is that much better than a CPU's.

[7im]

All work units are not the same size. One SMP work unit does not equal one GPU work unit. I don't think you can compare 3 minutes of GPU folding to 12 hours of SMP folding without knowing more about which work units are folding, how many molecules each is modeling, implicit or explicit solvent models, etc.

About the only think you can do easily is compare points. And while that is not the best comparison either, it's all that we as contributors can easily judge.

[EvilAlchemist]

I remember viewing a slide from a Presentation a while back about the clients.

Performance
- GPU > PS3 > SMP > Single Core

Flexibility
- Single Core > SMP > PS3 > GPU

GPU are very fast at certain types of projects.
You also have to remember that GPU's can only run a subset of the project.
It is always important to have a good mix of clients. Never a good idea to load up on one client type.

[chengbin]

It takes 3-4 minutes to complete 500 foldings on my 8800GT while it takes 12 hours to do the same amount of folding on my Q6600. How come we can only utilize the great power of the GPU on a few things?

[Xilikon]

Actually, work units is varied in terms of solvent type (implicit or explicit), atoms count and other parameters. The GPU is very fast but they only process implicit solvent units (implicit = computing the solvent as a continous force) while the SMP client is working on explicit solvent units (explicit = solvent atoms is computed individually). Both types is valuable for Stanford and is rewarded appropriately so don't worry

Also, the atoms count can make a big difference in processing time.

[7im]

Actually, work units is varied in terms of solvent type (implicit or explicit), atoms count and other parameters. The GPU is very fast but they only process implicit solvent units (implicit = computing the solvent as a continous force) while the SMP client is working on explicit solvent units (explicit = solvent atoms is computed individually). Both types is valuable for Stanford and is rewarded appropriately so don't worry

Also, the atoms count can make a big difference in processing time.

Sounds like there's an echo in here...

From a few posts back...http://foldingforum.org/viewtopic.php?p=51264#p51264

It takes 3-4 minutes to complete 500 foldings on my 8800GT while it takes 12 hours to do the same amount of folding on my Q6600. How come we can only utilize the great power of the GPU on a few things?

Even those 500 steps are not created equally. As said several times above, 500 steps with 400 atoms is not the same as 500 steps with 10,000 atoms. If the GPU client did 500/400 in 3 minutes, and the SMP client did 500/10,000 in 12 hours, then which is more helpful to the project. It's hard to know without more details.

Steps completed is not a good judge of performance. Especially if you consider some work units step up 100 units at a time, and some step up 5000 units at time. Example...


[10:13:47] Completed 490000 out of 500000 steps (98 percent)
[10:28:48] Timered checkpoint triggered.
[10:36:03] Writing local files
[10:36:03] Completed 495000 out of 500000 steps (99 percent)

versus...

[13:22:13] Completed 1640000 out of 4000000 steps (41)
[13:37:13] Timered checkpoint triggered.
[13:52:13] Timered checkpoint triggered.
[13:59:48] Writing local files
[13:59:48] Completed 1680000 out of 4000000 steps (42)

[chengbin]

How do you get a 4 million steps job? My CPU only gets 500 each time and 25000 each time for a GPU. How do I check how many atoms are there?

[bruce]

At the top of this page there is a link to a page called Project Summary. It lists the atom counts for every active project. I don't know any way to get the step count unless you look in FAHlog.txt where (depending on which FahCore is being used) you will probably find statements like
Completed 1300000 out of 2500000 steps (52%)

[chengbin]

My mistake. The CPU is actually doing 500000 steps instead of 500. My CPU work has 1258 atoms and my GPU work has 576 atoms. My GPU is doing 25000000 steps.

[Xilikon]

My mistake. The CPU is actually doing 500000 steps instead of 500. My CPU work has 1258 atoms and my GPU work has 576 atoms. My GPU is doing 25000000 steps.

You notice there is no correlation between the atoms size and number of steps ? This is why using steps as a measuring stick of value is not the right thing to do.