Bigger and better GPU sieving drive: Discussion

[frmky]

Originally Posted by Ken_g6

The only two good ways I can think of to provide that are to either recompile the client with the latest CUDA SDK or to vectorize the work like I did for AMD. Neither option is all that easy, but compiling the latest SDK on a clean VM is probably easier.

I recompiled the source at http://github.com/Ken-g6/PSieve-CUDA using CUDA 3.1 for Linux without issue. It runs at the same speed as the precompiled 2.3 binary on the GTX 480 and gives the same results.

[Ken_g6]

Originally Posted by frmky

I recompiled the source at http://github.com/Ken-g6/PSieve-CUDA using CUDA 3.1 for Linux without issue. It runs at the same speed as the precompiled 2.3 binary on the GTX 480 and gives the same results.

It should. The issue is with the 460.

I know; I get those confused sometimes, too.

Edit: My issue with compiling on the latest SDK is that I need to keep the older SDK to compile for BOINC.

[frmky]

replace 10T-20T figures with 1000G-2000G, since people will actually be reserving these in terms of G most likely

I think you mean 10000G-20000G. And I'll kick off the reservations with 10-10.1T.

Edit: Ummm, I mean 10-11T. Lost a power of ten myself there.

[mdettweiler]

Originally Posted by Ken_g6

It should. The issue is with the 460.

I know; I get those confused sometimes, too.

Edit: My issue with compiling on the latest SDK is that I need to keep the older SDK to compile for BOINC.

I have the CUDA 3.1 developer's toolkit installed, so I tried compiling ppsieve-CUDA just now. I got this error:

gary@buttford:~/Desktop/gpu-stuff/ppsieve31$ make
cd pps; make
make[1]: Entering directory `/home/gary/Desktop/gpu-stuff/ppsieve31/pps'
Makefile:77: Extraneous text after `else' directive
gcc -c -m32 -march=i586 -mtune=core2 -o ../main.o -fomit-frame-pointer -s -O3 -DNDEBUG -D_REENTRANT -I. -I.. ../main.c
../main.c:1: error: bad value (core2) for -mtune= switch
make[1]: *** [../main.o] Error 1
make[1]: Leaving directory `/home/gary/Desktop/gpu-stuff/ppsieve31/pps'
make: *** [all] Error 2

BTW @Greg: thanks for helping out!

[gd_barnes]

Is PST aware that we are taking 10T-20T in this forum? I didn't see any posting for it over there.

[mdettweiler]

Originally Posted by gd_barnes

Is PST aware that we are taking 10T-20T in this forum? I didn't see any posting for it over there.

They delete their reservation/completion posts after the tables have been updated accordingly. I did make a post earlier, but it's now been deleted. If you look in the reservation thread you will see this about halfway down the page:

Completed and Reserved Ranges - dat-less sieving

RSP1M 4 < k < 10000 for n < 2M; Please visit the Welcome to RSP Sieving thread for instructions on dat-less sieving.
10000G-20000G NPLB

[Flatlander]

When will k<1001, n<1m benefit from this sieving?

a) Weeks.
b) Months.
c) Never. That range will be finished before anything is snapped of the sieve file.
d) How long is a piece of string?

[Mini-Geek]

Originally Posted by Flatlander

When will k<1001, n<1m benefit from this sieving?

a) Weeks.
b) Months.
c) Never. That range will be finished before anything is snapped of the sieve file.
d) How long is a piece of string?

If the sieve depth listed in the files posted for drives 6 and 7 (which fit your k and n limits) are correct, they are sieved to 26T. You could start benefiting as soon as factors above that are being reported. I don't know if the admins would actually do that due to the extra work and complexity of removing new factors now and then, but it could help. If we estimate that the equivalent of 5 GTX 480s are used on this drive, (no idea how close to the truth that is) and each takes 2.3 days to run 1T, and ignore that some <10T sieving isn't done yet, that could be in as little as a week. This is probably an overly optimistic assuming that number of GPUs, since people reserve in blocks instead of taking minute pieces in order, but that's what it should average to. So maybe figure in the 'weeks' category, assuming about that overall GPU speed.
Some other sieve depths, as gleaned from files:
drive 11 (k=2000-3000, n=425K-650K, outside of your k bounds): 30T
mini-drive (k=3010-3200, n=420259-650K, outside of your k bounds): 30T
individual k (k=300-400, n=1M-2M, outside of your n bounds): 140T
Those are all the depths I could find.

So my best guess is a) Weeks. (if the admins are willing to apply the factors ASAP, and the sieve depths listed aren't terribly off), or b or c (Months-Never) if they're not.
Remember that there's no benefit to snapping out parts of the sieve file, (as in, stopping factoring of those portions) since they keep sieving numbers they don't need to along with the large ones that they're trying to reach the optimal depth for. Of course, at a certain point, you'll say, "we can start testing of this file at the current depth", but factors for them will continue to be found.

It seems to me that calculating the optimal depth for this sieve will be far harder than traditional sieves, since you get a great deal of unusable factors along with usable ones, and must factor that in. Or it's at least much harder than the sort of optimal calculations I do for CRUS bases, maybe it's more comparable to other big sieves...

[mdettweiler]

Originally Posted by Flatlander

When will k<1001, n<1m benefit from this sieving?

a) Weeks.
b) Months.
c) Never. That range will be finished before anything is snapped of the sieve file.
d) How long is a piece of string?

Probably Weeks. I'm thinking that once the new sieve gets to, say, p=100T, that would be a good time to update the k<1001 drives' files. We may only catch the tail end of the 5th Drive with that (in which case it would probably be easier management-wise just to finish that one off on the 26T file), but the 6th and 7th drives should benefit quite a bit.

[Ken_g6]

Originally Posted by mdettweiler

I have the CUDA 3.1 developer's toolkit installed, so I tried compiling ppsieve-CUDA just now. I got this error:

gary@buttford:~/Desktop/gpu-stuff/ppsieve31$ make
cd pps; make
make[1]: Entering directory `/home/gary/Desktop/gpu-stuff/ppsieve31/pps'
Makefile:77: Extraneous text after `else' directive
gcc -c -m32 -march=i586 -mtune=core2 -o ../main.o -fomit-frame-pointer -s -O3 -DNDEBUG -D_REENTRANT -I. -I.. ../main.c
../main.c:1: error: bad value (core2) for -mtune= switch
make[1]: *** [../main.o] Error 1
make[1]: Leaving directory `/home/gary/Desktop/gpu-stuff/ppsieve31/pps'
make: *** [all] Error 2

BTW @Greg: thanks for helping out!

Either (1) get gcc 4.3, or (2) remove the -mtune directive. I sorta thought gcc 4.3 or later would be required to install the CUDA SDK.

What Linux version are you using?

[em99010pepe]

Originally Posted by frmky

The GTX 480 is about 50x faster than a single core:

GTX 480 has a peak power in load of 450 W. GTX 480 sieves at 5103 kp/sec.
A core i5 750 (quad-core) has a peak power in load of 95 W. Core i5 sieves at 120kp/sec/core.

GTX 480 will sieve 1T in 55 hours having a energy consumption of 24.49 kWh. For the same consumption the core i5 would only sieve to 0.45T. This means the quad-core machine would spend 55 kWh of electric energy to sieve 1T.

GTX 480 24.49 kWh/(Tera sieved).
Core i5 750 54.97 kWh/(Tera sieved).

In terms of efficiency the GTX 480 is only 2.24x more efficient than one i5 750.

Another question is to know which one will fail first due to heat, the processor or the GPU, and of course the investment. The GTX 480 cost 2.4x the i5 750. So what's the conclusion? (lol)