Welcome to the testing thread for what will probably be the third CUDA BOINC project for PrimeGrid, a CUDA Cullen/Woodall sieve (for base 2 only):

CWPSieve-CUDA (Source, on the "cw" branch)

CWPSieve-CUDA is based on TPSieve-CUDA, and by partial coincidence they both get about the same speedup over their CPU counterparts.

Now that I have a compute-capable GPU, I've tested CWPSieve-CUDA on several ranges and it seems to work fine. But I have only a Fermi-based card on a 64-bit Linux box. I need testers using older cards and/or Windows. Despite the speed advantage CUDA provides, most test ranges are slow to provide factors. But here's one that finds three factors in less than a minute on my GTX 460:

Win7 x64
GTX 470 @ 750/1500/1674

Without sievefile:

I was going for ruby badge on CPUs!
This could change sub-project "life" expectancy

Win7 x64/ GTX580 GPU version, very low CPU usage.

Stand alone

NVIDIA GeForce 9800 GTX+ on Linux 64-bit

$ ./cwpsieve-cuda-boinc-x86_64-linux -p2563602e7 -P2563604e7 -n 10000000 -N 25000000
cwpsieve version cuda-0.2.3 (testing)
Compiled Jan 3 2011 with GCC 4.3.3
nstart=10000000, nstep=20
Changed nstep to 19
cwpsieve initialized: 10000000 <= n <= 25000000
25636026136339 | 24184321*2^24184321-1
25636029526061 | 12004589*2^12004589+1
25636030632281 | 14263341*2^14263341+1
Found 3 factors

Can't open init data file - running in standalone mode
Sieve started: 25636020000000 <= p < 25636040000000
Thread 0 starting
Detected GPU 0: GeForce 9800 GTX+
Detected compute capability: 1.1
Detected 16 multiprocessors.

Thread 0 completed
Sieve complete: 25636020000000 <= p < 25636040000000
count=647707,sum=0xe66f848aacfc21bb
Elapsed time: 133.70 sec. (0.01 init + 133.69 sieve) at 150984 p/sec.
Processor time: 1.06 sec. (0.02 init + 1.04 sieve) at 19408738 p/sec.
Average processor utilization: 1.42 (init), 0.01 (sieve)
called boinc_finish

C:\ppsieve_cuda>cwpsievecuda -p2563602e7 -P2563604e7 -n 10000000 -N 25000000
cwpsieve version cuda-0.2.3 (testing)
nstart=10000000, nstep=20
Changed nstep to 19
cwpsieve initialized: 10000000 <= n <= 25000000
Sieve started: 25636020000000 <= p < 25636040000000
Thread 0 starting
Detected GPU 0: GeForce GTX 570
Detected compute capability: 2.0
Detected 15 multiprocessors.
25636026136339 | 24184321*2^24184321-1
25636029526061 | 12004589*2^12004589+1
25636030632281 | 14263341*2^14263341+1

Thread 0 completed
Waiting for threads to exit
Sieve complete: 25636020000000 <= p < 25636040000000
Found 3 factors
count=647707,sum=0xe66f848aacfc21bb
Elapsed time: 30.88 sec. (0.03 init + 30.84 sieve) at 654430 p/sec.
Processor time: 0.66 sec. (0.02 init + 0.64 sieve) at 31508430 p/sec.
Average processor utilization: 0.50 (init), 0.02 (sieve)


Done with GTX570 Windows

Win7 x64
GTS 250 @ 750/2000/1100

Without sievefile:

Win XP Pro 32-bit
9600 GSO

Cannot get application to run. Copied the cudart.dll file from the BOINC directory, but I get the following error message:

The procedure entry point cudaSetDeviceFlags could not be located in the dynamic link library cudart.dll

I may just be an inexperienced idiot trying to do this, but this is what I get:

Michael, are you running 64-bit Linux or 32-bit Linux?
____________

What compute capability and memory is needed? I have an nVidia Quadro NVS 135M (1.1 - 128MB).
____________
Murphy (AtP)

What compute capability and memory is needed? I have an nVidia Quadro NVS 135M (1.1 - 128MB).

What compute capability and memory is needed? I have an nVidia Quadro NVS 135M (1.1 - 128MB).

Good work again Ken!
Is there likely to be an ATI version of this, or are you too fed up of that?

BTW, I found from a colleague that my high CPU with ATI is due to a busy-wait that got introduced between the 2.1 and 2.2 SDK - very nice of them - NOT.
____________

Michael, are you running 64-bit Linux or 32-bit Linux?

Good work again Ken!
Is there likely to be an ATI version of this, or are you too fed up of that?

BTW, I found from a colleague that my high CPU with ATI is due to a busy-wait that got introduced between the 2.1 and 2.2 SDK - very nice of them - NOT.

8400M GS
Vista 32-bit

Win XP

Could not find "cudart.dll"...
Copied several "cudart.dll" from various other folders, but it did not like any of them. Could not find entry ...

???
____________
Murphy (AtP)

This should be the right cudart.dll. If it's not, ask Scott Brown what he did to solve it. ;)
____________

@Ken
Is this the 32 or 64bit version?

For Linux are both versions in the download folder:
http://www.primegrid.com/download/libcudart.so.2.32bit
http://www.primegrid.com/download/libcudart.so.2.64bit is equal to http://www.primegrid.com/download/libcudart.so

[edit]
link
____________
Best wishes. Knowledge is power. by jjwhalen

For Windows there is no 64-bit version, because I can't compile a 64-bit binary with the free MS Express tools. Just use the 32-bit version; it should work as well as TPSieve does.
____________

What is TPSieve-CUDA? I don't see that app listed on the apps page:

http://www.primegrid.com/apps.php
____________
Reno, NV

What is TPSieve-CUDA? I don't see that app listed on the apps page:

http://www.primegrid.com/apps.php

Linux64, GT240(GT215)

Can't open init data file - running in standalone mode
Sieve started: 25636020000000 <= p < 25636040000000
Resuming from checkpoint p=25636022621441 in cwpcheck2563602e7.txt
Thread 0 starting
Detected GPU 0: GeForce GT 240
Detected compute capability: 1.2
Detected 12 multiprocessors.

Thread 0 completed
Sieve complete: 25636020000000 <= p < 25636040000000
count=647707,sum=0xe66f848aacfc21bb
Elapsed time: 201.32 sec. (0.01 init + 201.31 sieve) at 87248 p/sec.
Processor time: 68.40 sec. (0.02 init + 68.39 sieve) at 256823 p/sec.
Average processor utilization: 1.22 (init), 0.34 (sieve)
called boinc_finish

cwpsieve version cuda-0.2.3 (testing)
Compiled Jan 3 2011 with GCC 4.3.3
nstart=10000000, nstep=20
Changed nstep to 19
cwpsieve initialized: 10000000 <= n <= 25000000
Sieve started: 25636020000000 <= p < 25636040000000
Thread 0 starting
Detected GPU 0: GeForce GT 240
Detected compute capability: 1.2
Detected 12 multiprocessors.
p=25636025767169, 96.12K p/sec, 0.35 CPU cores, 28.8% done. ETA 05 Jan 12:56
25636026136339 | 24184321*2^24184321-1
25636029526061 | 12004589*2^12004589+1
p=25636031010049, 87.38K p/sec, 0.33 CPU cores, 55.1% done. ETA 05 Jan 12:57
25636030632281 | 14263341*2^14263341+1
p=25636036252929, 87.38K p/sec, 0.35 CPU cores, 81.3% done. ETA 05 Jan 12:57

Thread 0 completed
Waiting for threads to exit
Sieve complete: 25636020000000 <= p < 25636040000000
Found 3 factors
count=647707,sum=0xe66f848aacfc21bb
Elapsed time: 230.58 sec. (0.01 init + 230.56 sieve) at 87547 p/sec.
Processor time: 78.34 sec. (0.02 init + 78.32 sieve) at 257710 p/sec.
Average processor utilization: 1.24 (init), 0.34 (sieve)

WinXP 32bit:

@Ken

boinc@vmware2k-3:~/Cuda/gcw$ time ./gcwsieve_1.12_x86_64-pc-linux-gnu -p2563602e7 -P2563604e7 -n 10000000 -N 25000000
gcwsieve 1.3.8 -- A sieve for Generalised Cullen/Woodall numbers n*b^n+/-1.
Read 1108077 terms n*2^n+/-1 from ABC file `sieve.txt'.
gcwsieve 1.3.8 started: 10000005 <= n <= 24999999, 25636020000000 <= p <= 25636040000000
p=25636020524351, 8087 p/sec, 0 factors, 2.6% done, ETA 05 Jan 14:57
p=25636021573147, 8207 p/sec, 0 factors, 7.9% done, ETA 05 Jan 14:56
p=25636022097479, 8179 p/sec, 0 factors, 10.5% done, ETA 05 Jan 14:56
p=25636022621921, 8199 p/sec, 0 factors, 13.1% done, ETA 05 Jan 14:56
p=25636023146329, 8141 p/sec, 0 factors, 15.7% done, ETA 05 Jan 14:56
p=25636023670771, 8121 p/sec, 0 factors, 18.4% done, ETA 05 Jan 14:56
p=25636024195163, 8172 p/sec, 0 factors, 21.0% done, ETA 05 Jan 14:56
p=25636024719653, 8130 p/sec, 0 factors, 23.6% done, ETA 05 Jan 14:56
p=25636025244209, 8146 p/sec, 0 factors, 26.2% done, ETA 05 Jan 14:56
p=25636025768557, 8164 p/sec, 0 factors, 28.8% done, ETA 05 Jan 14:56
25636026136339 | 24184321*2^24184321-1
p=25636026161819, 8113 p/sec, 1 factor, 30.8% done, ETA 05 Jan 14:56
p=25636026686201, 8170 p/sec, 1 factor, 33.4% done, 818 sec/factor
p=25636027210609, 8129 p/sec, 1 factor, 36.1% done, ETA 05 Jan 14:56
p=25636027735061, 8135 p/sec, 1 factor, 38.7% done, 950 sec/factor
p=25636028259529, 7981 p/sec, 1 factor, 41.3% done, ETA 05 Jan 14:56
p=25636028783929, 8181 p/sec, 1 factor, 43.9% done, 1073 sec/factor
p=25636029308333, 8149 p/sec, 1 factor, 46.5% done, ETA 05 Jan 14:56
25636029526061 | 12004589*2^12004589+1
p=25636029570551, 8198 p/sec, 2 factors, 47.9% done, 583 sec/factor
p=25636030094933, 8159 p/sec, 2 factors, 50.5% done, ETA 05 Jan 14:56
25636030632281 | 14263341*2^14263341+1
p=25636030750457, 7987 p/sec, 3 factors, 53.8% done, ETA 05 Jan 14:56
p=25636031274839, 8179 p/sec, 3 factors, 56.4% done, 459 sec/factor
p=25636031799271, 8184 p/sec, 3 factors, 59.0% done, ETA 05 Jan 14:56
p=25636032323789, 8154 p/sec, 3 factors, 61.6% done, 503 sec/factor
p=25636032848201, 8179 p/sec, 3 factors, 64.2% done, ETA 05 Jan 14:56
p=25636033372583, 8057 p/sec, 3 factors, 66.9% done, 553 sec/factor
p=25636033896971, 8120 p/sec, 3 factors, 69.5% done, ETA 05 Jan 14:56
p=25636034421299, 8152 p/sec, 3 factors, 72.1% done, 589 sec/factor
p=25636034945639, 8152 p/sec, 3 factors, 74.7% done, ETA 05 Jan 14:56
p=25636035470027, 8164 p/sec, 3 factors, 77.4% done, 631 sec/factor
p=25636035994523, 8169 p/sec, 3 factors, 80.0% done, ETA 05 Jan 14:56
p=25636036518899, 8131 p/sec, 3 factors, 82.6% done, 677 sec/factor
p=25636037043229, 8141 p/sec, 3 factors, 85.2% done, ETA 05 Jan 14:56
p=25636037567669, 8143 p/sec, 3 factors, 87.8% done, 719 sec/factor
p=25636038092047, 8190 p/sec, 3 factors, 90.5% done, ETA 05 Jan 14:56
p=25636038616417, 8266 p/sec, 3 factors, 93.1% done, 750 sec/factor
p=25636039140739, 8196 p/sec, 3 factors, 95.7% done, ETA 05 Jan 14:56
p=25636039665157, 8119 p/sec, 3 factors, 98.3% done, 807 sec/factor
gcwsieve 1.3.8 stopped: at p=25636040000000 because range is complete.

Found factors for 3 terms in 2454.810 sec. (expected about 0.03)

real 40m56.805s
user 40m53.689s
sys 0m0.764s

I renamed "gcwsieve_1250180098.sieveinput" to "sieve.txt". I found nothing with the smaller "gcwsieve_20110101.sieveinput" file. I this only for testing?

Thanks John. The new App_info file is ready for takeoff...
____________
Best wishes. Knowledge is power. by jjwhalen

Ken,

I know this isn't really your responsibility, but perhaps you know what the plans are.

Once a CW sieve app on CUDA is released, is there going to be a mechanism that allows people to select one CUDA app over the other without using app_info? (While separately selecting CPU tasks, of course.)

P.S. Thanks for all the incredible work. You have essentially revolutionized sieving here at PrimeGrid.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Once a CW sieve app on CUDA is released, is there going to be a mechanism that allows people to select one CUDA app over the other without using app_info? (While separately selecting CPU tasks, of course.)

Once a CW sieve app on CUDA is released, is there going to be a mechanism that allows people to select one CUDA app over the other without using app_info? (While separately selecting CPU tasks, of course.)

This issue has not been resolved...and yes, it's a pretty significant issue. Outside of the app_info approach, there does not appear to be a way to separate the GPU projects.

The "cheat" that has allowed work to flow to the GPU while still being able to select CPU projects will still work. However, there will be no way to select PPS (Sieve) or CW (Sieve) for the GPU. Work will come from both.

At this time, it seems that PG is the only BOINC project that experiences this unique issue with all the sub-projects that are offered. Therefore, there doesn't appear to be much activity at Berkley to resolve it. :(

However, maybe a BOINC developer will see our plight and come to the rescue.

I renamed "gcwsieve_1250180098.sieveinput" to "sieve.txt". I found nothing with the smaller "gcwsieve_20110101.sieveinput" file. I this only for testing?

gcwsieve_20110101 is a new sieve file. Those factors have already been removed from it.

1/5/2011 11:25:48 AM NVIDIA GPU 0: Quadro NVS 135M (driver version 26099, CUDA version 3020, compute capability 1.1, 128MB, 13 GFLOPS peak)

stdout.txt:

cwpsieve version cuda-0.2.3 (testing)
nstart=10000000, nstep=20
Changed nstep to 19
cwpsieve initialized: 10000000 <= n <= 25000000
25636026136339 | 24184321*2^24184321-1
25636029526061 | 12004589*2^12004589+1
25636030632281 | 14263341*2^14263341+1
Found 3 factors

stderr.txt:

11:30:54 (4340): Can't set up shared mem: -1. Will run in standalone mode.
Sieve started: 25636020000000 <= p < 25636040000000
Thread 0 starting
Detected GPU 0: Quadro NVS 135M
Detected compute capability: 1.1
Detected 1 multiprocessors.

Thread 0 completed
Sieve complete: 25636020000000 <= p < 25636040000000
count=647707,sum=0xe66f848aacfc21bb
Elapsed time: 4834.21 sec. (0.08 init + 4834.13 sieve) at 4176 p/sec.
Processor time: 30.42 sec. (0.19 init + 30.23 sieve) at 667620 p/sec.
Average processor utilization: 2.40 (init), 0.01 (sieve)
12:51:28 (4340): called boinc_finish

cwpfactors.txt:

25636026136339 | 24184321*2^24184321-1
25636029526061 | 12004589*2^12004589+1
25636030632281 | 14263341*2^14263341+1

You can not get much more minimal GPU that this.
____________
Murphy (AtP)

Ken,

I know this isn't really your responsibility, but perhaps you know what the plans are.

Once a CW sieve app on CUDA is released, is there going to be a mechanism that allows people to select one CUDA app over the other without using app_info? (While separately selecting CPU tasks, of course.)

P.S. Thanks for all the incredible work. You have essentially revolutionized sieving here at PrimeGrid.

Great stuff Ken, let the gold-rush on the C/W Sieve project begin!

I've done Mac builds (x86/x86_64 boinc/non-boinc), get them here http://www.pyramid-productions.net/downloads/cwpsieve-cuda.tar.gz. As usual, CUDA 3.2 is required since that's my build platform.

I've tested the range you suggested and it looks fine.

Cheers

- Iain

After the successful standalone tests I tried to run a few Cullen/Woodall Sieve WUs via BOINC.

Good news: WUs are completed, results are uploaded, run time ~130s (on GTX460), CPU time ~4s.

Bad news: They are declared invalid.

Is it just the validator ignoring them because of the short run time (I remember this issue with the very first ppsieve-CUDA WUs) or are there some more modifications to be done?
____________

Michael,

As one of the proponents of running VM's to solve BOINC's shortcomings, couldn't a solution be done with them?

i.e., - setup a single VM to run on all but one core. In the non-VM, run the GPU by selecting either CW or PPS sieves and deselecting the run on CPU option. On the VM, deselect the run on ATI and run on NVidia options and pick the CPU projects you want. You miss one core, but you should be able to crunch only the things you want to.

After the successful standalone tests I tried to run a few Cullen/Woodall Sieve WUs via BOINC.

Good news: WUs are completed, results are uploaded, run time ~130s (on GTX460), CPU time ~4s.

Bad news: They are declared invalid.

Is it just the validator ignoring them because of the short run time (I remember this issue with the very first ppsieve-CUDA WUs) or are there some more modifications to be done?

Bad news: They are declared invalid.

Is it just the validator ignoring them because of the short run time (I remember this issue with the very first ppsieve-CUDA WUs) or are there some more modifications to be done?

WIN7 x64 GTX275:

c:\Test>cwpsieve-cuda-x86-windows.exe -p2563602e7 -P2563604e7 -n 10000000 -N 25000000
cwpsieve version cuda-0.2.3a (testing)
nstart=10000000, nstep=20
Changed nstep to 19
cwpsieve initialized: 10000000 <= n <= 25000000
Sieve started: 25636020000000 <= p < 25636040000000
Thread 0 starting
Detected GPU 0: GeForce GTX 275
Detected compute capability: 1.3
Detected 30 multiprocessors.
25636026136339 | 24184321*2^24184321-1
25636029526061 | 12004589*2^12004589+1
25636030632281 | 14263341*2^14263341+1
p=25636033631489, 227.2K p/sec, 0.01 CPU cores, 68.2% done. ETA 06 Jan 10:09

Thread 0 completed
Waiting for threads to exit
Sieve complete: 25636020000000 <= p < 25636040000000
Found 3 factors
count=647707,sum=0xe66f848aacfc21bb
Elapsed time: 98.39 sec. (0.03 init + 98.36 sieve) at 205214 p/sec.
Processor time: 0.67 sec. (0.03 init + 0.64 sieve) at 31508430 p/sec.
Average processor utilization: 1.14 (init), 0.01 (sieve)


Linux x64 GTX480:

./cwpsieve-cuda-x86_64-linux -p2563602e7 -P2563604e7 -n 10000000 -N 25000000
cwpsieve version cuda-0.2.3a (testing)
Compiled Jan 5 2011 with GCC 4.3.3
nstart=10000000, nstep=20
Changed nstep to 19
cwpsieve initialized: 10000000 <= n <= 25000000
Sieve started: 25636020000000 <= p < 25636040000000
Thread 0 starting
Detected GPU 0: GeForce GTX 480
Detected compute capability: 2.0
Detected 15 multiprocessors.
25636026136339 | 24184321*2^24184321-1
25636029526061 | 12004589*2^12004589+1
25636030632281 | 14263341*2^14263341+1

Thread 0 completed
Waiting for threads to exit
Sieve complete: 25636020000000 <= p < 25636040000000
Found 3 factors
count=647707,sum=0xe66f848aacfc21bb
Elapsed time: 41.97 sec. (0.02 init + 41.95 sieve) at 481141 p/sec.
Processor time: 11.26 sec. (0.03 init + 11.23 sieve) at 1797425 p/sec.
Average processor utilization: 1.48 (init), 0.27 (sieve)

--
Lumiukko

MAC OS X 10.6.5, iMac, with a GT120 cuda 3.2.17, BOINC suspended, using Iain's build posted several hours ago:

Ken,

I know this isn't really your responsibility, but perhaps you know what the plans are.

Once a CW sieve app on CUDA is released, is there going to be a mechanism that allows people to select one CUDA app over the other without using app_info? (While separately selecting CPU tasks, of course.)

P.S. Thanks for all the incredible work. You have essentially revolutionized sieving here at PrimeGrid.

Michael,

As one of the proponents of running VM's to solve BOINC's shortcomings, couldn't a solution be done with them?

i.e., - setup a single VM to run on all but one core. In the non-VM, run the GPU by selecting either CW or PPS sieves and deselecting the run on CPU option. On the VM, deselect the run on ATI and run on NVidia options and pick the CPU projects you want. You miss one core, but you should be able to crunch only the things you want to.

After the successful standalone tests I tried to run a few Cullen/Woodall Sieve WUs via BOINC.

Good news: WUs are completed, results are uploaded, run time ~130s (on GTX460), CPU time ~4s.

Bad news: They are declared invalid.

Is it just the validator ignoring them because of the short run time (I remember this issue with the very first ppsieve-CUDA WUs) or are there some more modifications to be done?

Stderr output
<core_client_version>6.10.56</core_client_version>
<![CDATA[
<message>
Maximum elapsed time exceeded
</message>
<stderr_txt>
Unrecognized XML in parse_init_data_file: hostid
Skipping: 115189
Skipping: /hostid
Unrecognized XML in parse_init_data_file: starting_elapsed_time
Skipping: 0.000000
Skipping: /starting_elapsed_time
Sieve started: 102554440000000 <= p < 102554500000000
Thread 0 starting
Detected GPU 0: GeForce GT 240
Detected compute capability: 1.2
Detected 12 multiprocessors.

</stderr_txt>
]]>

After the successful standalone tests I tried to run a few Cullen/Woodall Sieve WUs via BOINC.

Good news: WUs are completed, results are uploaded, run time ~130s (on GTX460), CPU time ~4s.

Bad news: They are declared invalid.

Is it just the validator ignoring them because of the short run time (I remember this issue with the very first ppsieve-CUDA WUs) or are there some more modifications to be done?

It's running.
First unit with resultid=214457695 waits for validation....

Wonder how we can tweak the Linux version a bit to run on Mac...or is that too far in the future?

I think the invalidation problem is fixed in v0.2.3b. The factors file needed to be non-empty even when no factors were found.
____________

[quote][quote]After the successful standalone tests I tried to run a few Cullen/Woodall Sieve WUs via BOINC.

Good news: WUs are completed, results are uploaded, run time ~130s (on GTX460), CPU time ~4s.

Bad news: They are declared invalid.
...
/quote]
How do I get the C/W CUDA to run under BOINC?
____________
Murphy (AtP)

With an app_info.xml file. Very carefully. ;)
____________

And app_info.xml should look like... ? :)

Card/OS/Drivers: GTX 460 / 64 Bit Linux 2.6.32 / Driver Version 256.53
Clocks: Factory overclocked @725 MHz

6 Jan 2011 22:36:07 UTC 6 Jan 2011 22:43:22 UTC Completed and validated 145.32 25.13 182.43 Cullen/Woodall Prime Search (Sieve)
Anonymous platform (NVIDIA GPU)
6 Jan 2011 22:56:51 UTC 6 Jan 2011 23:01:49 UTC Completed and validated 143.06 25.31 182.43 Cullen/Woodall Prime Search (Sieve)
Anonymous platform (NVIDIA GPU)
6 Jan 2011 22:50:45 UTC 6 Jan 2011 22:58:32 UTC Completed and validated 142.79 24.49 182.43 Cullen/Woodall Prime Search (Sieve)
Anonymous platform (NVIDIA GPU)

Thanks to rroonnaalldd for the app_info templates.
____________

I think the invalidation problem is fixed in v0.2.3b. The factors file needed to be non-empty even when no factors were found.

Another question:

Will there be a CPU version of cwpsieve?
Or will it not offer any benefit over the existing gcwsieve?
____________

Another question:

Will there be a CPU version of cwpsieve?
Or will it not offer any benefit over the existing gcwsieve?

Ken is it possible to release one Cuda-app for both Sieve projects?

as slow as it can get?

NVS 3100M:


7 Jan 2011 17:02:50 UTC Completed and validated 3,599.50 15.94 182.43 Cullen/Woodall Prime Search (Sieve)
Anonymous platform (NVIDIA GPU)

The "cheat" that has allowed work to flow to the GPU while still being able to select CPU projects will still work. However, there will be no way to select PPS (Sieve) or CW (Sieve) for the GPU. Work will come from both.

Another way around the problem:

Create 2 new sub-projects for GPU only --
CW (Sieve) for the GPU.
PPS (Sieve) for the GPU.

And make the current sub-projects CPU only --
CW (Sieve) for the CPU.
PPS (Sieve) for the CPU.

Credit could be combined or not as Project Admins determine.
____________
Murphy (AtP)

A third way around the problem: These GPU-capable projects should be advancing very quickly. If they're not past the point of optimal efficiency on CPUs they should be soon. Once they are, why not disable the CPU clients and make them GPU-only projects?
____________

A third way around the problem: These GPU-capable projects should be advancing very quickly. If they're not past the point of optimal efficiency on CPUs they should be soon. Once they are, why not disable the CPU clients and make them GPU-only projects?

A third way around the problem: These GPU-capable projects should be advancing very quickly. If they're not past the point of optimal efficiency on CPUs they should be soon. Once they are, why not disable the CPU clients and make them GPU-only projects?

Even with the latest CWPSieve CUDA some WUs are not validated: 151348618

Seems to be a validator issue, all WUs with less than 3 seconds of CPU time are marked as invalid. ;)
____________

Even with the latest CWPSieve CUDA some WUs are not validated: 151348618

Seems to be a validator issue, all WUs with less than 3 seconds of CPU time are marked as invalid. ;)

<app>
<name>gcwsieve</name>
<user_friendly_name>Cullen/Woodall Prime Search (Sieve)</user_friendly_name>
</app>
<file_info>
<name>gcwsieve_20110101.sieveinput_in</name>
<status>1</status>
<sticky/>
</file_info>
<file_info>
<name>gcwsieve_1250180098.sieveinput_in</name>
<status>1</status>
<sticky/>
</file_info>
<file_info>
<name>cwpsieve-cuda-boinc-x86_64-linux</name>
<status>1</status>
<executable/>
</file_info>
<file_info>
<name>libcudart.so.2</name>
<status>1</status>
</file_info>
<file_info>
<name>stat_primegrid.png</name>
<status>1</status>
</file_info>
<file_info>
<name>primegrid_slideshow_00.png</name>
<status>1</status>
</file_info>
<app_version>
<app_name>gcwsieve</app_name>
<version_num>112</version_num>
<platform>x86_64-pc-linux-gnu</platform>
<avg_ncpus>0.020000</avg_ncpus>
<max_ncpus>0.020000</max_ncpus>
<flops>1000000000.000000</flops>
<plan_class>cuda23</plan_class>
<api_version>6.2.18</api_version>
<file_ref>
<file_name>cwpsieve-cuda-boinc-x86_64-linux</file_name>
<open_name>cwpsieve-cuda-boinc-x86_64-linux</open_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>libcudart.so.2</file_name>
<open_name>libcudart.so.2</open_name>
</file_ref>
<coproc>
<type>CUDA</type>
<count>0.500000</count>
</coproc>
<cmdline></cmdline>
<gpu_ram>268435456.000000</gpu_ram>
</app_version>

I think the invalidation problem is fixed in v0.2.3b. The factors file needed to be non-empty even when no factors were found.

Oops. I meant to add cudart.dll to GeneferCUDA.zip. I added it to CWPSieve.zip by mistake. As it's the wrong version for CWPSieve, I've removed it.

Thanks for catching that!
____________

Despite of running apps on GPU for quite a long time so not new to GPU computing, I'm pretty new to nVidia (preferred ATI/AMD for years). And since Fermi is an interesting card, I got one.

My question would be - any difference in speed vs driver version on nVidia?
Or is it just fine when you make it runnning (see ATI issues with OpenCL).
____________
My stats

On Linux, 256.53 is the best nVIDIA driver. But you're on Windows, and I don't know what's best there.
____________