It was a year ago today when I received that thrashing in thread 847

So the good book's right there is nothing new under the sun.



Dave

The rosetta at home grid has 39,729 computers running at 12.5 teraflops.

This is an interesting number because there is no redunancy on workunits.

Everyone does there workunit then that becomes the result and the workunit is closed out.

From what I can see of WCG there is I belive a redunancy of 3 and maybe as much as six. I have seen both figures tossed around the forums.

An educated guess??? I belive we have an effective teraflops (taking redunancy into account) of probably the same as rosetta at home 12.5 teraflops.

have a nice day everyone...........maybe someone can give some reasons for the high redunancy rates on this Grid?..............is the average member so keen on points that they would overclock to the point of screwing up there calculations?

with six projects eventually coming into the fold maybe the redunancy will have to go down in order to process data in a reasonable amount of time?

Just some thoughts to stir theings up a little...got to keep things lively....

Cheers

It may be a little off topic, but:
I've read somewhere, that ussually WSG sends one Work Unit to 5 devices, and waits for calculations. If users statistics are known, WCG knows which user send back calculations regularly. So it can send one Work Unit to two certauin users (for comparision of results), and wait for example, 48 hours, for callculations (if user don't send it back, WCG can send this work unit to another users then). If WCG works that way, I think productivity of the grid would be better than 20%. May I be wright? May the grid works that way?

Hi all,
I'll weigh in with some more information. The basic FPU unit for handling 64-bit floating point (actually, 80-bit internally with some underflow protection) has sped up tremendously the last 4 or 5 years. In point of fact, both AMD and Intel can deliver short bursts with better than 0.9 flops per cycle. But this is very unrealistic. The LINPACK benchmarks are what people rely on, and they show that the true sustained speed is much slower. Even so, in some ways a current CPU is the equivalent of a supercomputer in the early 1970s.

The streaming instructions add much more complexity. In particular, it is possible to stream 32-bit floating point instructions, with 2 32-bit floats per 64-bit word. Finally, there are some very fast 12-bit streaming instructions that use a table lookup rather than any sort of ALU. These are not true floating point instructions, but are intended for screen graphics calculations. 4096 * 4096 allows calculating the pixel placement for a 2048 * 2048 display without introducing moire patterns (by Nyquist's theorem).

If 32-bit floating point is good enough, it can make sense to use these instructions, as long as you are aware that:
1) Not all CPUs will support all these streaming instructions.
2) Therefore, some results will return with slightly different values computed using 64-bit double floats.

Note that the standard LINPACK benchmarks are using the 64-bit FPU.

Just tossing in my 2 cents.
mycrofth

Here's a good read

http://courses.ece.uiuc.edu/ece512/Papers/Athlon.pdf