Folding Forum

Adam A. Wanderer wrote:http://www.nytimes.com/2010/06/01/scien ... &emc=aua21
Time marches on. It's tragic the F@H project doesn't have a truly strong, or custom designed, super computer. It might make a nice complement to the network home PCs.

Actually PG has said before many times it prefers the networked home pc's. Not having to pay the power bill but being able to funnel that money into actual research could be one of them ( and pretty important ) at it.

Edit:

Consideration changed my mind, especially since I didn't emphasize the 'complement' part as much ( or frankly didn't seem to register it at all ). I agree it would probably make a very nice complement when it's also custom designed

This begs the question, how would the F@H distributed network compare to the #1 supercomputer in terms of FLOPS?

Jaguar is rated at 1.8 PFLOPS.
http://en.wikipedia.org/wiki/TOP500

Fah smokes it by double, conservatively speaking. They don't give out Guiness World Records for no reason.
http://fah-web.stanford.edu/cgi-bin/mai ... pe=osstats

7im that's still a theoretical 'free' 50% increase isn't it

MtM wrote:7im that's still a theoretical 'free' 50% increase isn't it

No that was answered some time ago. The FLOPS that Stanford is reporting for the GPU's are actual not theoretical.

Come on this isn't a technical thread which you're trying to turn it into.

It would be a nice theoretical free increase, forget the numbers they are not important.

So, it would be nice, as all free additions would be nice, there is no more logic to this argument then that.

MtM wrote:7im that's still a theoretical 'free' 50% increase isn't it

I'm not sure what you are asking. If it's a question of how FAH calculates FLOPS, that is answered on that stats page I linked. And I only used the conservative FLOPS number, not the much higher x86 FLOPS equivalent. I'm rather sure the super computer pages don't use anything conservative.

As I tried to convey to p5-133XL I think the thread isn't aimed at being a hard numbers topic but more of an 'what if' philosophical discussion. In which I noted the important aspect would be 'free addition' leaving out any numbers, and already pointed that 'even' supercomputers have their downsides.

If it was a thread aimed at a technical discussion, there is enough info to be found about the subject using the search function I believe

MtM wrote:Come on this isn't a technical thread which you're trying to turn it into.

Right.

My first response to Slash_2CPU's "...how would ... F@H compare to the #1 supercomputer in terms of FLOPS?" was to ask why FLOPS since that's a poor way to compare FAH performance (repeatedly discussed elsewhere -- use 'search') without ever getting into a discussion of which measurement of FLOPS is better (also repeatedly discussed elsewhere) but I didn't bother to post it.

Fact: FAH is bigger than that supercomputer.
Fact: FAH's analysis methodology can be run on supercomputers and it would complement the work being done by donors.
Fact: If you've got a supercomputer in your garage and want to donate it to FAH, I'm sure arrangements can be made. All donations are accepted.

Chose FLOPS since it's an easily comparable standard, albeit an admittedly poor one.

Second question: Could the servers handle an immediate increase in load of that magnitude, or would we be out of SMP, bigadv, A3, etc in a short time?

Slash_2CPU wrote:Chose FLOPS since it's an easily comparable standard, albeit an admittedly poor one.

Second question: Could the servers handle an immediate increase in load of that magnitude, or would we be out of SMP, bigadv, A3, etc in a short time?

No, but that's hardly the point. A supercomputer can't run the FAH client in it's present form and it wouldn't make sense to connect it in the same way. If you have one you want to donate, special arrangements will be made. It would need special WUs and would not connect using the current servers.

Once upon a time there was just one type of client ("classic") and one type of server. As new types were developed (GPU, SMP, PS3) new types of servers were added. I'd expect that they'd add a server or two to distribute supercomputer projects.

Vijay said in the past that they have contingency plans to support up to 1 Million clients. Unknown how long it would take to ramp up those plans. So while they may fall short of WUs in the short term during the ramp up, they would fix it long term.

Sorry to burst everyone's bubble. To be ranked on the Top500 list you need to perform 64-bit floating point arithmetic. FAH is almost exclusively 32-bit FP. Therefore the FLOPS on the client page are not comparable to those oft reported Linpack benchmark scores.

Computers with very high double precision performance are also subject to trade restrictions due to their potential use in nuclear simulation. The export rules don't apply to single-precision behemoths, as far as I know.

Wrish wrote:The export rules don't apply to single-precision behemoths, as far as I know.

Drifting off-topic, I suppose . . . but what classes as a behemoth? The X6 or the I7? Or the upcoming 48-core chip from Intel? Or perhaps that chip only if installed in a 4-socket server board? Or perhaps a GPU with some high number of shaders? or ....???

The dividing line is becoming fuzzy.

There are no single-precision behemoths. Even the 8086 ran double precision although you had to add a 8087 Math Coprocessor to do any floating point math -- whether FAH ever uses it or not.

Wrish wrote:Sorry to burst everyone's bubble. To be ranked on the Top500 list you need to perform 64-bit floating point arithmetic. FAH is almost exclusively 32-bit FP. Therefore the FLOPS on the client page are not comparable to those oft reported Linpack benchmark scores.

Fah is almost exclusively 32 bit, yes, but the hardware it runs on is not. Wouldn't it be possible to extrapolate?

For example, intel rates their xeon 5300 at 1.84 GFLOPS using that Top500 Linpack test.