Folding Forum

Good evening everyone. I will be writing an article/essay on folding@home for a scientific journal with a particular focus on the role of GPUs. I have read the FAQ and Science pages on the folding@home website, Nvidia's "About CUDA" and Tom's Hardware article on CUDA. I would be grateful if you post links to any resources you know of, e.g: comparisons between CPU/GPU/PS3 calculation speed etc. The sought of information I am looking for is:

Thus, it would take 10,000 CPU days to simulate folding -- i.e. it would take 30 CPU years!

but with more details on what CPU/processing power it is referring to, and what the GPU functions the folding client utilises (e.g: shader processor/video RAM/stream processing etc). It would also be handy to know how points and work units correlate to number of proteins folded. Thanks in advance.

http://fah-web.stanford.edu/cgi-bin/mai ... pe=osstats
From the website above you should be able to calculate the MFLOP/TFLOP output per CPU/GPU/PS3 Client.

What the quote simply does is to explain the basis behind parallel/distributed computing: to split the workload into smaller pieces to be managed by many processors.

The folding client utilizes the stream processors of DX10-capable cards (otherwise known as stream processing, the former architecture involved vertex/pixel *shaders*), while using the video RAM only to store information.

Each work unit corresponds to an instance of a small timeframe of protein folding, with random trajectories and/or other parameters of each atom in a protein. The number of points reflects the amount of time taken to complete a WU, as compared to a benchmark machine, which is in turn pegged against a value which relates to the amount of science processed in that WU.

My two-piece. Not sure if there's more elaborate documentation somewhere.

Don't fall in to the trap of directly comparing TFLOPS from one client to the next. Speed of processing is not the whole picture. GPUs are REALLY fast, but only do implicit solvent models, while the SMP client does explicit models. Explicit are more helpful to the project, but then SMP clients are slower than GPU clients. That's one reason the FLOPS count is not a direct link to science completed by each project.

Read the History section of the GPU FAQs for some insight in to this. Unfortunately, the project doesn't publish hard numbers to campare the clients. And even if they did, it wouldn't mean that much because of that implicent vs. explicit thing I mentioned.

I suggest you email Dr. Pande with your questions. Best to get the answers directly from the source.

Please check out our papers on our web site for more info or PM me with specific questions.