Exciting News!!!

Over in the PST forum, Ken and the crew have been working on a GPU application for sieving. The program is called ppsieve and is currently being used on the PPSE sieve. If all goes well, we may be able to merge the PPS and PPSE sieves bringing PPSE into BOINC.

Currently available for 32 & 64 bit Linux AND 32 bit Windows (will run on 64 bit). It should work on cards with any compute capability. You can download it here:

ppsieve-cuda.zip (source)

To test, please use the following command line:

./ppsieve-cuda-boinc-(version) -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -c 60

It should output the following factors:

Range: 42070e9 to 42070030e6
42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070003101727 | 4207*2^1054290+1
42070003511309 | 6057*2^1043547+1
42070006307657 | 1513*2^1771812+1
42070006388603 | 2059*2^1816098+1
42070007177519 | 5437*2^1121592+1
42070007396759 | 7339*2^1803518+1
42070008823897 | 4639*2^952018+1
42070008858187 | 2893*2^317690+1
42070010190569 | 5625*2^1903125+1
42070011430123 | 3821*2^1406279+1
42070012301263 | 1957*2^1185814+1
42070013521999 | 1965*2^404493+1
42070013970587 | 7143*2^1462422+1
42070013989247 | 5037*2^838603+1
42070017332953 | 6237*2^1916994+1
42070018235321 | 1941*2^363948+1
42070019542387 | 8587*2^1703626+1
42070023987581 | 9811*2^318944+1
42070024339237 | 9257*2^1170495+1
42070024532551 | 4311*2^1690093+1
42070024936837 | 5679*2^1726142+1
42070024995961 | 9111*2^1707153+1
42070026021997 | 4039*2^1819590+1
42070027452199 | 1323*2^854008+1
42070029006583 | 5943*2^663870+1
Found 27 factors

Please provide as much details about your system as possible.

Thank you for testing!

p.s. If you wish to test the CUDA time vs. CPU time, you can download the CPU build here: ppsieve-bin.zip (source)

Just run the same test range and then compare the results.

Other sample test cases:

Range: 20070e9 to 20070010e6
20070000475957 | 4995*2^1822738+1
20070001146497 | 4977*2^626298+1
20070001163929 | 3765*2^461308+1
20070001302811 | 7669*2^725426+1
20070001425977 | 5821*2^1775248+1
20070002245151 | 1221*2^646983+1
20070002606341 | 4809*2^497683+1
20070004816819 | 6699*2^1215561+1
20070005914001 | 9847*2^1634140+1
20070006187837 | 9923*2^287853+1
20070006875981 | 1645*2^965954+1
20070007170259 | 3889*2^49730+1
20070008329039 | 9065*2^832569+1
Found 13 factors

Range: 249871e9 to 2498711e8
249871003789289 | 6295*2^266404+1
249871009510013 | 2771*2^1272671+1
249871010360639 | 1743*2^1337710+1
249871027030549 | 8865*2^1534637+1
249871030776329 | 7815*2^1679937+1
249871032591751 | 2335*2^23512+1
249871038523049 | 7527*2^204096+1
249871049497963 | 6497*2^505399+1
249871066947839 | 8497*2^1221770+1
249871068167599 | 7311*2^450531+1
249871089712009 | 9281*2^1650023+1
249871091913587 | 2139*2^1290902+1
249871099624639 | 8381*2^350375+1
Found 13 factors

Range: 42070e9 to 42070100e6
Found 68 factors
____________

System: Debian GNU Linux (Squeeze), Kernel 2.6.32 AMD64
Intel i5-750
GeForce 9500 GT (silent)
Nvidia Driver Ver.: 190.53
Cuda-Toolkit: 2.3 Ubuntu 9.04
Running AP26 on all 4 Cores, no other boinc-based GPU-apps running
GPU-Temperature changes from 38°C to 40°C during the sieve runs

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070003e6 -k 1201 -K 9999 -N 2000000 -z normal
ppsieve version cuda-0.1.1-beta (testing)
Compiled Mar 7 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070003000000
Starting 1 threads.
Detected GPU 0: GeForce 9500 GT
Detected compute capability: 1.1
Detected 4 multiprocessors.
42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070000300049 | 9139*2^461846+1
42070000345343 | 1715*2^635711+1
42070000464001 | 4179*2^1577462+1
42070000949861 | 4707*2^571847+1
42070001011573 | 7113*2^215532+1
42070001040127 | 6471*2^37907+1
42070002482267 | 9951*2^1920408+1
42070002690167 | 2553*2^1888870+1
42070002698543 | 4239*2^368773+1
42070002875941 | 4081*2^1494668+1

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070003000000
Found 12 factors
count=95668,sum=0x37dacb7121ccffe4
Elapsed time: 24.00 sec. (0.01 init + 23.99 sieve) at 131122 p/sec.
Processor time: 3.42 sec. (0.02 init + 3.40 sieve) at 925156 p/sec.
Average processor utilization: 1.14 (init), 0.14 (sieve)
____________

System: OpenSuSE 11.2 Kernel 2.6.31.12 amd64
CPU: Core 2 Quad Q9550 (E0 stepping)
GPU: GeForce GTX 260-192 - NVIDIA driver version: 190.53 (cuda 2.3 support)

Test run with the CPU idling and the GPU at stock clock:

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070003e6 -k 1201 -K 9999 -N 2000000 -z normal

ppsieve version cuda-0.1.1-beta (testing)
Compiled Mar 7 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070003000000
Starting 1 threads.
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.
42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070000300049 | 9139*2^461846+1
42070000345343 | 1715*2^635711+1
42070000464001 | 4179*2^1577462+1
42070000949861 | 4707*2^571847+1
42070001011573 | 7113*2^215532+1
42070001040127 | 6471*2^37907+1
42070002482267 | 9951*2^1920408+1
42070002690167 | 2553*2^1888870+1
42070002698543 | 4239*2^368773+1
42070002875941 | 4081*2^1494668+1

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070003000000
Found 12 factors
count=95668,sum=0x37dacb7121ccffe4
Elapsed time: 5.81 sec. (0.02 init + 5.80 sieve) at 542504 p/sec.
Processor time: 1.35 sec. (0.02 init + 1.33 sieve) at 2363791 p/sec.
Average processor utilization: 1.10 (init), 0.23 (sieve)

Test run with the CPU idling and the GPU at 667 MHz (shaders linked):

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070003e6 -k 1201 -K 9999 -N 2000000 -z normal
ppsieve version cuda-0.1.1-beta (testing)
Compiled Mar 7 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070003000000
Starting 1 threads.
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.
42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070000300049 | 9139*2^461846+1
42070000345343 | 1715*2^635711+1
42070000464001 | 4179*2^1577462+1
42070000949861 | 4707*2^571847+1
42070001011573 | 7113*2^215532+1
42070001040127 | 6471*2^37907+1
42070002482267 | 9951*2^1920408+1
42070002690167 | 2553*2^1888870+1
42070002698543 | 4239*2^368773+1
42070002875941 | 4081*2^1494668+1

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070003000000
Found 12 factors
count=95668,sum=0x37dacb7121ccffe4
Elapsed time: 4.98 sec. (0.02 init + 4.96 sieve) at 634269 p/sec.
Processor time: 1.22 sec. (0.02 init + 1.20 sieve) at 2617475 p/sec.
Average processor utilization: 0.95 (init), 0.24 (sieve)

To give an impression of the current speeds here are the runtimes of the test range on a Q9550 @ 3.4 GHz (FSB 400 * 8.5) CPU use the linux-x86_64 version of the ppsieve-cpu application.

Running only 1 thread:

ppsieve version 0.3.4a (testing)
Compiled Feb 19 2010 with GCC 4.3.3
Algorithm not specified, starting benchmark...
bsf takes 350000; mul takes 490000; using standard algorithm.
nstart=1999980, nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070003000000
Thread 0 starting

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070003000000
Found 8 factors
count=95668,sum=0x37dacb7121ccffe4
Elapsed time: 16.44 sec. (0.90 init + 15.54 sieve) at 193956 p/sec.
Processor time: 16.41 sec. (0.86 init + 15.55 sieve) at 193898 p/sec.
Average processor utilization: 0.95 (init), 1.00 (sieve)

Running 4 threads:

ppsieve version 0.3.4a (testing)
Compiled Feb 19 2010 with GCC 4.3.3
Algorithm not specified, starting benchmark...
bsf takes 350000; mul takes 540000; using standard algorithm.
nstart=1999980, nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070003000000
Thread 0 starting
Thread 1 starting
Thread 2 starting
Thread 3 starting

Thread 3 completed
Waiting for threads to exit
Thread 0 completed
Thread 1 completed
Thread 2 completed
Sieve complete: 42070000000000 <= p < 42070003000000
Found 8 factors
count=95668,sum=0x37dacb7121ccffe4
Elapsed time: 4.94 sec. (0.91 init + 4.03 sieve) at 748919 p/sec.
Processor time: 16.11 sec. (0.91 init + 15.21 sieve) at 198232 p/sec.
Average processor utilization: 0.99 (init), 3.78 (sieve)

---

The "Elapsed time" on a GTX 260-192 @ 667 MHz is nearly the same as on the Q9550 @ 3.4 GHz. So the output of the GTX 260-192 at stock clock is roughly the same as that of 4 cores of my Q9550 at stock clock. This ratio also applies to the current AP26 apps (1.01 (cuda23) vs 1.04).

Sieve complete: 42070000000000 <= p < 42070003000000
Found 8 factors

64 bit GPU app - Test range 30M

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q

My GTX 260-192 at stock clock with no load on the CPU:

ppsieve version cuda-0.1.1-beta (testing)
Compiled Mar 7 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Starting 1 threads.
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 49.30 sec. (0.02 init + 49.29 sieve) at 611635 p/sec.
Processor time: 3.85 sec. (0.02 init + 3.83 sieve) at 7866200 p/sec.
Average processor utilization: 1.04 (init), 0.08 (sieve)

and again at 667 MHz with no load on the CPU (shaders linked):

ppsieve version cuda-0.1.1-beta (testing)
Compiled Mar 7 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Starting 1 threads.
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 42.05 sec. (0.02 init + 42.02 sieve) at 717374 p/sec.
Processor time: 3.72 sec. (0.02 init + 3.70 sieve) at 8142356 p/sec.
Average processor utilization: 0.82 (init), 0.09 (sieve)

32 bit GPU app - Test range 30M

./ppsieve-cuda-x86-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q

My GTX 260-192 at stock clock with no load on the CPU:

ppsieve version cuda-0.1.1-beta (testing)
Compiled Mar 7 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Starting 1 threads.
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 49.59 sec. (0.02 init + 49.56 sieve) at 608232 p/sec.
Processor time: 4.10 sec. (0.02 init + 4.07 sieve) at 7399053 p/sec.
Average processor utilization: 1.11 (init), 0.08 (sieve)

and again at 667 MHz with no load on the CPU (shaders linked):

ppsieve version cuda-0.1.1-beta (testing)
Compiled Mar 7 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Starting 1 threads.
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 42.28 sec. (0.02 init + 42.26 sieve) at 713325 p/sec.
Processor time: 3.96 sec. (0.02 init + 3.94 sieve) at 7648693 p/sec.
Average processor utilization: 1.07 (init), 0.09 (sieve)

64 bit CPU app - Test range 30M

1 thread - Same CPU and clock rate as stated above:

./ppsieve-x86_64-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q
ppsieve version 0.3.4a (testing)
Compiled Feb 19 2010 with GCC 4.3.3
Algorithm not specified, starting benchmark...
bsf takes 350000; mul takes 500000; using standard algorithm.
nstart=1999980, nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Thread 0 starting
p=42070023724033, 196.6K p/sec, 1.00 CPU cores, 79.1% done. ETA 09 Mar 18:11
Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 153.56 sec. (0.87 init + 152.69 sieve) at 196576 p/sec.
Processor time: 153.24 sec. (0.87 init + 152.37 sieve) at 196995 p/sec.
Average processor utilization: 1.00 (init), 1.00 (sieve)

4 threads - Same CPU and clock rate as stated above:

ppsieve version 0.3.4a (testing)
Compiled Feb 19 2010 with GCC 4.3.3
Algorithm not specified, starting benchmark...
bsf takes 350000; mul takes 520000; using standard algorithm.
nstart=1999980, nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Thread 1 starting
Thread 0 starting
Thread 3 starting
Thread 2 starting

Thread 3 completed
Waiting for threads to exit
Thread 1 completed
Thread 0 completed
Thread 2 completed
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 39.77 sec. (0.90 init + 38.88 sieve) at 772082 p/sec.
Processor time: 152.13 sec. (0.89 init + 151.24 sieve) at 198467 p/sec.
Average processor utilization: 0.99 (init), 3.89 (sieve)

does this version work for 1.0 cards?
____________

does this version work for 1.0 cards?

Currently available for 32 & 64 bit Linux. It should work on cards with any compute capability. You can download it here:

oh, sorry, i missed that to. :-(
____________

Ubuntu 9.10, kernel 2.6.31-20-generic
Core2 Quad Q9550
GeForce 9800 GT, NVIDIA driver 190.42

This 30M range was run with the CPU cores busy on TRP sieve. Another run with the cores idle was just as fast (as expected).

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q

ppsieve version cuda-0.1.1-beta (testing)
Compiled Mar 7 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Starting 1 threads.
Detected GPU 0: GeForce 9800 GT
Detected compute capability: 1.1
Detected 14 multiprocessors.
p=42070029884417, 498.0K p/sec, 0.14 CPU cores, 99.6% done. ETA 13 Mar 17:49
Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 60.52 sec. (0.01 init + 60.51 sieve) at 498248 p/sec.
Processor time: 9.42 sec. (0.01 init + 9.41 sieve) at 3203673 p/sec.
Average processor utilization: 0.71 (init), 0.16 (sieve)

____________

Has anyone given any thought to what's going to happen when this goes live via BOINC?

In particular, consider how I've got PrimeGrid set up, which is probably a fairly common arrangement for people with CUDA capable GPUs:

1) I've got a quad-core CPU and a GTX280 GPU.

2) I run AP26 on the GPU

3) I run other PrimeGrid stuff on the CPU

4) I do not want to run AP26 on the CPU because I can run it much, much faster on the GPU (about 5 min/WU).

5) There's no explicit BOINC mechanism to say "run X on the CPU and Y on the GPU", unless PrimeGrid makes two separate sub-projects, "AP26-CPU" and "AP26-GPU".

6) So, I have ONLY the CPU tasks selected on project preferences page to feed the right tasks to the CPU

7) ... and "Send work from any subproject..." to send AP26 to the GPU. This works because nothing else exists for the GPU, so it has to send AP26 tasks.

As soon as there's more than one GPU project, there will be no way of selecting what you want to run on the GPU, unless you want to also allow those to run on the CPU (which I think most people would prefer not to do.)

One possible solution is to make separate sub-projects for the GPU versions, but I realize that's far from ideal.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

I think it would be a mistake to build too much complexity into local bespoke code.

We need the BOINC client to do the right thing with GPU scheduling. We need the BOINC server to allow per subproject CPU/GPU preferences. We should try and feed those requirements into the mainstream BOINC development process and then make use of it when released.

All IMHO of course, and somewhat idealistic, since BOINC development seems to follow whatever direction the Berkeley folks are meandering in at a particular moment in time...

____________

I intentionally avoided the whole concept of "well, the BOINC software really *should* do this..." for the obvious reasons, primarily that we're going to have a problem in the near future whereas the BOINC client might eventually get around to solving a problem like this anywhere from tomorrow to never. They've got much bigger GPU scheduling problems to solve first before they could get around to this one.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

As soon as there's more than one GPU project, there will be no way of selecting what you want to run on the GPU ...

As soon as there's more than one GPU project, there will be no way of selecting what you want to run on the GPU ...

The anonymous platform mechanism lets you control exactly what you want to run. I deployed it today on the Linux side of my quad and it even picked up existing tasks correctly.

Obviously, this isn't a solution for the masses, but it works for me.

As soon as there's more than one GPU project, there will be no way of selecting what you want to run on the GPU ...

The anonymous platform mechanism lets you control exactly what you want to run. I deployed it today on the Linux side of my quad and it even picked up existing tasks correctly.

Obviously, this isn't a solution for the masses, but it works for me.

It's been a loooooong time since I've set up app_info by hand, and I don't really remember how to do it. Anyone know of a good reference for what exactly needs to be done?

<app_info>

<app>
<name>ap26</name>
</app>
<file_info>
<name>primegrid_ap26_1.01_windows_intelx86__cuda23.exe</name>
<executable/>
</file_info>
<file_info>
<name>cudart.dll</name>
<executable/>
</file_info>
<app_version>
<app_name>ap26</app_name>
<version_num>101</version_num>
<plan_class>cuda23</plan_class>
<avg_ncpus>0.050000</avg_ncpus>
<max_ncpus>0.050000</max_ncpus>
<flops>5604000000.000000</flops>
<coproc>
<type>CUDA</type>
<count>1.000000</count>
</coproc>
<file_ref>
<file_name>primegrid_ap26_1.01_windows_intelx86__cuda23.exe</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>cudart.dll</file_name>
</file_ref>
</app_version>

<app>
<name>psp_sr2sieve</name>
</app>
<file_info>
<name>primegrid_sr2sieve_wrapper_1.12_windows_x86_64.exe</name>
<executable/>
</file_info>
<file_info>
<name>primegrid_sr2sieve_1.8.10_windows_x86_64.exe.orig</name>
<executable/>
</file_info>
<app_version>
<app_name>psp_sr2sieve</app_name>
<version_num>112</version_num>
<flops>2876776723.690640</flops>
<file_ref>
<file_name>primegrid_sr2sieve_wrapper_1.12_windows_x86_64.exe</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>primegrid_sr2sieve_1.8.10_windows_x86_64.exe.orig</file_name>
<open_name>primegrid_sr2sieve_1.8.10_windows_x86_64.exe.orig</open_name>
</file_ref>
</app_version>
</app_info>

Thanks, that makes sense and helps a lot.

____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Alright, I've posted a release candidate version to the locations given at the beginning of the thread. It may be slightly faster; it may also lock up your GPU until it's done. Let me know how it goes.

In any case, I've recently been looking at some other projects' GPU speeds, and I'm finding myself disappointed with my speeds. When Milkyway@Home is 17 times faster on high-end NVidia (PDF), and even a simple Collatz app (not the Collatz, but the only source code I could find) is more than twice as fast as a CPU on a mid-range card, but my code is only as fast as a CPU on a high-end card, I wonder if I'm doing something wrong. Would any of the experienced CUDA developers around here care to give my code the once-over, to see if I'm doing something obviously stupid like not giving the card enough threads?

I suppose the other side of the coin could be that my CUDA code isn't bad, but that my and Geoff's CPU code is extraordinarily good.

Edit: P.S. There's a BOINC capable (I think) executable in the zipfile as well. :)
____________

In any case, I've recently been looking at some other projects' GPU speeds, and I'm finding myself disappointed with my speeds. When Milkyway@Home is 17 times faster on high-end NVidia (PDF), and even a simple Collatz app (not the Collatz, but the only source code I could find) is more than twice as fast as a CPU on a mid-range card, but my code is only as fast as a CPU on a high-end card, I wonder if I'm doing something wrong.

Yeah...the application here is computationally-bound, and doesn't require much memory. Probably the slowest part are the 64-bit multiplies. When Fermi comes out, I expect that each stream processor will run my app (once recompiled) twice as fast.

Another part of it could be that others are comparing GPU speed to CPU speed on one core. In that case my app is 4 times as fast as the CPU version. :)
____________

Nice work so far Ken.. I only see one issue, this appears to be a compute-mode only CUDA application, meaning it will not run on the primary adapter under Windows in current form (driver watchdog timer). Correct?
____________

I have no idea what you just said! (I'm new at this CUDA stuff.)

If you mean it's not using the driver API, that's correct. I was hoping to avoid it.

Edit: Did you see cuda_sleep_memcpy.cu?
____________

I have no idea what you just said! (I'm new at this CUDA stuff.)

If you mean it's not using the driver API, that's correct. I was hoping to avoid it.

Have you investigated whether some of the later compute capabilities add features that increase speed? It is nice to see an application that works on all CUDA cards, but given that only a handful of models are compute capable version 1.0 (G80 chips), added features such as atomic functions in compute capable 1.1 cards might help with speed depending on the processes computed in the application.


____________
141941*2^4299438-1 is prime!

In Windows, if a CUDA kernel runs longer than 5 seconds the program will be terminated by the driver. Briefly looking at your posted source, it appears you're running one huge kernel.

Ok, i'll have to pay more attention to the code when reading. That should work fine in Windows, too.
____________

64 bit GPU app - Test range 30M

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q

My GTX 260-192 at stock clock with no load on the CPU:

ppsieve version cuda-0.1.1-rc1 (testing)
Compiled Mar 17 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Thread 0 starting
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 49.12 sec. (0.02 init + 49.10 sieve) at 613957 p/sec.
Processor time: 3.81 sec. (0.02 init + 3.79 sieve) at 7951250 p/sec.
Average processor utilization: 1.04 (init), 0.08 (sieve)

and again at 667 MHz with no load on the CPU (shaders linked):

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q
ppsieve version cuda-0.1.1-rc1 (testing)
Compiled Mar 17 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Thread 0 starting
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 41.87 sec. (0.02 init + 41.86 sieve) at 720229 p/sec.
Processor time: 3.69 sec. (0.02 init + 3.67 sieve) at 8215573 p/sec.
Average processor utilization: 1.04 (init), 0.09 (sieve)

____________

64 bit GPU app - Test range 30M - speed comparison

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q

My GTX 260-192 at stock clock with no load on the CPU:

ppsieve version cuda-0.1.1-beta (testing) : Elapsed time: 49.30 sec. (0.02 init + 49.29 sieve) at 611635 p/sec.
ppsieve version cuda-0.1.1-rc1 (testing) : Elapsed time: 49.12 sec. (0.02 init + 49.10 sieve) at 613957 p/sec.

and again at 667 MHz with no load on the CPU (shaders linked):

ppsieve version cuda-0.1.1-beta (testing) : Elapsed time: 42.05 sec. (0.02 init + 42.02 sieve) at 717374 p/sec.
ppsieve version cuda-0.1.1-rc1 (testing) : Elapsed time: 41.87 sec. (0.02 init + 41.86 sieve) at 720229 p/sec.
____________

Thanks for all your testing help, guys! I've got one more thing to test, and I hope you won't mind because I expect it to be slower. (But I've been wrong before!)

Linux 64-bit users *only*, with pre-Fermi cards (Fermi isn't in stores yet if you didn't know), please try the two binaries in this zipfile. This is an experiment in 24-bit multiplies instead of 64-bit ones. Both binaries do 24-bit multiplies, despite their names, but they do other stuff differently. Even if it doesn't work here, this is a plausible algorithm for ATI if I can ever figure out how to develop for OpenCL without their GPU.

If anyone reading this *does* have a Fermi (GTX4xx), I'd love to see a benchmark from the original code (linked in the first post by John). If Fermi doesn't run 50-100% faster per shader, I may have to recompile or something for maximum speed.
____________

SYSTEM Ubuntu 9.10
Intel Core2Duo T9550 @ 2.66GHZ
G105M
195.36.15 drivers

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070003e6 -k 1201 -K 9999 -N 2000000 -z normal

ppsieve version cuda-0.1.1-rc1 (testing)
Compiled Mar 17 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070003000000
Thread 0 starting
Detected GPU 0: GeForce G 105M
Detected compute capability: 1.1
Detected 1 multiprocessors.
42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070000300049 | 9139*2^461846+1
42070000345343 | 1715*2^635711+1
42070000464001 | 4179*2^1577462+1
42070000949861 | 4707*2^571847+1
42070001011573 | 7113*2^215532+1
42070001040127 | 6471*2^37907+1
42070002482267 | 9951*2^1920408+1
42070002690167 | 2553*2^1888870+1
42070002698543 | 4239*2^368773+1
42070002875941 | 4081*2^1494668+1

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070003000000
Found 12 factors
count=95668,sum=0x37dacb7121ccffe4
Elapsed time: 80.06 sec. (0.05 init + 80.02 sieve) at 39313 p/sec.
Processor time: 12.80 sec. (0.03 init + 12.77 sieve) at 246337 p/sec.
Average processor utilization: 0.67 (init), 0.16 (sieve)


pps-cuda-a1
./ppsieve-cuda-64bit-x86_64-linux -p42070e9 -P42070003e6 -k 1201 -K 9999 -N 2000000 -z normal

ppsieve version cuda-0.1.0-beta (testing)
Compiled Mar 29 2010 with GCC 4.3.3
nstart=76, nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070003000000
Starting 1 threads.
Detected GPU 0: GeForce G 105M
Detected compute capability: 1.1
Detected 1 multiprocessors.
42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070000300049 | 9139*2^461846+1
42070000345343 | 1715*2^635711+1
42070000464001 | 4179*2^1577462+1
42070000949861 | 4707*2^571847+1
42070001011573 | 7113*2^215532+1
42070001040127 | 6471*2^37907+1
42070002482267 | 9951*2^1920408+1
42070002690167 | 2553*2^1888870+1
42070002698543 | 4239*2^368773+1
42070002875941 | 4081*2^1494668+1

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070003000000
Found 12 factors
count=95668,sum=0x37dacb7121ccffe4
Elapsed time: 166.18 sec. (0.04 init + 166.14 sieve) at 18934 p/sec.
Processor time: 13.11 sec. (0.04 init + 13.07 sieve) at 240683 p/sec.
Average processor utilization: 1.01 (init), 0.08 (sieve)

./ppsieve-cuda-24bit-x86_64-linux -p42070e9 -P42070003e6 -k 1201 -K 9999 -N 2000000 -z normal

ppsieve version cuda-0.1.0-beta (testing)
Compiled Mar 29 2010 with GCC 4.3.3
nstart=76, nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070003000000
Starting 1 threads.
Detected GPU 0: GeForce G 105M
Detected compute capability: 1.1
Detected 1 multiprocessors.
42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070000300049 | 9139*2^461846+1
42070000345343 | 1715*2^635711+1
42070000464001 | 4179*2^1577462+1
42070000949861 | 4707*2^571847+1
42070001011573 | 7113*2^215532+1
42070001040127 | 6471*2^37907+1
42070002482267 | 9951*2^1920408+1
42070002690167 | 2553*2^1888870+1
42070002698543 | 4239*2^368773+1
42070002875941 | 4081*2^1494668+1

Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070003000000
Found 12 factors
count=95668,sum=0x37dacb7121ccffe4
Elapsed time: 163.43 sec. (0.04 init + 163.39 sieve) at 19252 p/sec.
Processor time: 12.32 sec. (0.04 init + 12.28 sieve) at 256167 p/sec.
Average processor utilization: 1.01 (init), 0.08 (sieve)
____________

64 bit GPU app - Test range 30M

./ppsieve-cuda-x86_64-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q

My GTX 260-192 at stock clock with no load on the CPU:

ppsieve version cuda-0.1.0-beta (testing)
Compiled Mar 29 2010 with GCC 4.3.3
nstart=76, nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Starting 1 threads.
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.
p=42070020971521, 349.5K p/sec, 0.06 CPU cores, 69.9% done. ETA 01 Apr 05:42
Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 87.11 sec. (0.02 init + 87.09 sieve) at 346140 p/sec.
Processor time: 4.61 sec. (0.02 init + 4.59 sieve) at 6566015 p/sec.
Average processor utilization: 1.11 (init), 0.05 (sieve)

./ppsieve-cuda-24bit-x86_64-linux -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q

My GTX 260-192 at stock clock with no load on the CPU:

ppsieve version cuda-0.1.0-beta (testing)
Compiled Mar 29 2010 with GCC 4.3.3
nstart=76, nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Starting 1 threads.
Detected GPU 0: GeForce GTX 260
Detected compute capability: 1.3
Detected 24 multiprocessors.
p=42070019922945, 332.0K p/sec, 0.06 CPU cores, 66.4% done. ETA 01 Apr 05:47
Thread 0 completed
Waiting for threads to exit
Sieve complete: 42070000000000 <= p < 42070030000000
Found 97 factors
count=955289,sum=0x2dbc17167afb6a8d
Elapsed time: 94.13 sec. (0.02 init + 94.12 sieve) at 320306 p/sec.
Processor time: 4.69 sec. (0.02 init + 4.67 sieve) at 6449443 p/sec.
Average processor utilization: 1.10 (init), 0.05 (sieve)

____________

Yeah, apparently my test version is slower as expected. There's probably no need to test it any more. But thanks for the testing you did!
____________

Ken, whenever you get the code finalized, I can build a Win32 version if you need.
____________

Okay, took another range: 174550 to 174551

Intel Xeon W3520 => 43 factors found; 855.903 k p/s
Nvidia GTX 260 => 43 factors found; 894.724 k p/s
Nvidia FX 580 => manually aborted due to long runtime; ~115 k p/s

Like in AP26 with mfl0ps newest app the GTX260 is a bit faster than the Xeon W3520.

OK, new version uploaded, probably the finalized version, at the links in the first post. I found a somewhat major bug in previous versions: around 30 of the highest N's in the ranges were being skipped! But that's fixed now.

Bryan, see if you can get this code to build in VS, perhaps without BOINC first. If you need to make changes, perhaps I should set up a GitHub account?
____________

Ok, will try building Win32 version soon. Thanks Ken
____________

Ken, if you're ready for me to do a Mac port, I'd be very happy to start on that if I can get the source code.

Cheers

- Iain

@Iain the source code is linked in the first post of this thread
____________

Well, I thought I was done, but I've made a few more changes for the release version of PPSieve-CUDA. The biggest change is compiling with CUDA 3.0. I hope it works for everyone!

The biggest code change is that I gave up using boinc_init_parallel() in favor of boinc_init(), because it's more compatible. The rest of the code changes are to header files and paths to BOINC header files. So nothing major there.

By the way, apparently CUDA 3.0 introduces an easier way to lower CPU usage. It might go from 5% down to 1 or 2%. But I'm going to leave lowering CPU usage for V0.1.2, if it's needed.
____________

Thanks Ken - I just got the 0.1.1-rc2 version ported to Mac OS X (only minor tweaks required as __thread attribute is not supported by GCC on Mac OS X), I'll pick up the new code and rebuild ASAP and post here when it's done (hopefully in a couple of days)...

Compiling with CUDA 3.0 will probably mean that many will be forced to upgrade drivers to at least the 195.xx series. Just FYI...the 196.xx and 197.xx dirvers have been noted to slow down many cards computational speeds compared to the 190.xx and 191.xx drivers (especially 8xxx and 9xxx series cards under Win7 and Vista), so the gain in freer CPU may actually be lost (and maybe exceeded) by the loss in speed on some cards.


____________
141941*2^4299438-1 is prime!

OK, it doesn't have to be compiled with 3.0 (yet). I wasn't sure if 2.3 would support Fermi. Since it looks like it does (PDF), I'll see about going back to 2.3.

Edit: To be clear, only the binaries will change, not the source code.
____________

Compiling with CUDA 3.0 will probably mean that many will be forced to upgrade drivers to at least the 195.xx series. Just FYI...the 196.xx and 197.xx dirvers have been noted to slow down many cards computational speeds compared to the 190.xx and 191.xx drivers (especially 8xxx and 9xxx series cards under Win7 and Vista), so the gain in freer CPU may actually be lost (and maybe exceeded) by the loss in speed on some cards.

OK, it doesn't have to be compiled with 3.0 (yet). I wasn't sure if 2.3 would support Fermi. Since it looks like it does (PDF), I'll see about going back to 2.3.

Edit: To be clear, only the binaries will change, not the source code.

...And we're back to binaries compiled with CUDA 2.3. :)

Edit: Scott, I think that's right. I doubt there are any drivers older than that for Fermi.
____________

The Mac OS X / CUDA version of ppsieve is now available for testing:

Mac 32 bit (OS 10.5+ required) - http://www.pyramid-productions.net/downloads/ppsieve-cuda-boinc-i686-apple-darwin.tar.gz

Note that only 32 bit CUDA executables are supported on the Mac, but as most runtime is spent on the GPU, this is not a problem. Since upgrading to Mac OS 10.6.3, Apple now only support CUDA 3.0, so this app is build and linked with the CUDA 3.0 libraries. However, it should work fine with machines where CUDA 2.3 is installed. If you have a Mac running OS 10.5 and/or CUDA 2.3 I'd be very grateful for your testing.

To test the app, please use the same inputs as in the original post, and obviously the output should be the same!

On my machine (MacBookPro, 2.66 GHz Core 2 Duo, GeForce 9400M / 9600M GT) with the CPU idling, the 9400M takes

Elapsed time: 67.96 sec. (0.02 init + 67.94 sieve) at 46303 p/sec.
Processor time: 6.93 sec. (0.03 init + 6.90 sieve) at 455624 p/sec.
Average processor utilization: 1.54 (init), 0.10 (sieve)

and the 9600M GT takes:

Elapsed time: 41.52 sec. (0.02 init + 41.50 sieve) at 75805 p/sec.
Processor time: 3.90 sec. (0.03 init + 3.87 sieve) at 812352 p/sec.
Average processor utilization: 1.35 (init), 0.09 (sieve)

Any problems or performance results please post to this thread.

Thanks

- Iain

Does anyone have plans for a Proth Prime Sieve ATI GPU app in the near future?
____________
May the Force be with you always.

An ATI app should be possible. This is the kind of highly-parallel, low-memory work that should work very well on ATI.

However, I haven't even been able to get their OpenCL compiler to run. Right now I'm focusing on getting the CPU and CUDA apps into BOINC, so ATI is off my radar for now.
____________

I ran the Mac version. Here's how the BOINC client sees my computer (24" iMac, OS X 10.6.3):

Processor: 2 GenuineIntel Intel(R) Core(TM)2 Duo CPU E8335 @ 2.93GHz [x86 Family 6 Model 23 Stepping 10]
Processor features: FPU VME DE PSE TSC MSR PAE MCE CX8 APIC SEP MTRR PGE MCA CMOV PAT PSE36 CLFSH DS ACPI MMX FXSR SSE SSE2 SS HTT TM SSE3 MON DSCPL VMX EST TM2 SSSE3 CX16 TPR PDCM SSE4.1
OS: Darwin: 10.3.0
Memory: 4.00 GB physical, 559.57 GB virtual
Disk: 595.85 GB total, 559.32 GB free
Local time is UTC -7 hours
NVIDIA GPU 0: GeForce GT 120 (driver version unknown, CUDA version 3000, compute capability 1.1, 256MB, 80 GFLOPS peak)

I first suspended all BOINC tasks. Here's the output:

% /usr/bin/time ./ppsieve-cuda-boinc-i686-apple-darwin -p42070e9 -P42070003e6 -k 1201 -K 9999 -N 2000000 -z normal
ppsieve version cuda-0.1.1 (testing)
Compiled Apr 16 2010 with GCC 4.2.1 (Apple Inc. build 5646) (dot 1)
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070000300049 | 9139*2^461846+1
42070000345343 | 1715*2^635711+1
42070000464001 | 4179*2^1577462+1
42070000949861 | 4707*2^571847+1
42070001011573 | 7113*2^215532+1
42070001040127 | 6471*2^37907+1
42070002482267 | 9951*2^1920408+1
42070002690167 | 2553*2^1888870+1
42070002698543 | 4239*2^368773+1
42070002875941 | 4081*2^1494668+1
Found 12 factors
28.37 real 3.61 user 0.13 sys

Screen repainting was pretty herky-jerky when the test was running, not that that's unexpected... it would have been pretty annoying if I was trying to do anything else.

-- Gary

That's great - thanks for testing Gary!

I think we have a problem here...:

I downloaded the ppsieve-cuda from the link, the ppsieve i use is from http://primesearchteam.com/showthread.php?t=25 and of version 0.3.4

Here come my results:

ppsieve-0.3.4 on Intel Core 2 Quad Q9550

Hm, not good is right. The CUDA app performed fine - all those factors are valid. But the CPU app didn't! :Q

I'm in another race, but in about 3 hours I'll have enough free memory to look into this.
____________

Okay. Meanwhile i am cross checking some other ranges i have on file and keep the results posted.

Well, I can't reproduce your CPU results. I get exactly the same results as you got with CUDA. But there could be several reasons for that.

For one thing, the version I downloaded from the PPSE Sieve Reservations page is 0.3.3. Not that 0.3.4 should produce bad results like that, but I can't compare directly.

PPSieve runs my CPU hot, at least as hot as fast LLR tests. Is it possible your machine isn't entirely stable?

Otherwise, PM me and I'll see about getting your version of files to test with.
____________

Okay, i took 0.3.3 and all is fine.

Intel C2Q Q9550 vs Nvidia GeForce GTX260

All okay, the only thing that was different is:

What is that while computing the 198900G to 199000G range while using cuda-0.1.1-1rc?

Two possibilities for computation errors. Most likely this one is because you're using 0.1.1-rc1. I believe I fixed a bug between rc2 and the final release that could rarely cause this error. It could definitely cause factors to be missed near NMax.

So please download the latest version from the link in the top post.

A computation error means that the GPU says it found some factor (it doesn't return what factor), but the CPU failed to find a factor in that range. So it could also be caused by an unstable GPU or rarely an unstable CPU.
____________

FYI, the source for PPSieve CUDA is now on GitHub!

http://github.com/Ken-g6/PSieve-CUDA

So now you can see the various directions I've considered. The redc branch is the current one, but I have an idea for the other branch that might pull it ahead, if I can find a large enough, fast-enough area of memory; maybe texture memory.

But first, since I've heard nothing from mfl0p, I think I'd better try to set up a WinXP VM and build a version for Windows.
____________

Thank you very much for setting up the repository. This makes it easier to follow the developments. I think it is time for me to reinstall the NVIDIA drivers and their CUDA toolkit under Lucid Lynx and get my GTX 260-192 out of hibernation mode again (in the last few weeks I've crunched with a HD 4770 under Windows and Linux).

By the way: The repo contains a file named pps/ppse_37TE1.txt that is a link to a file in a /downloads/... directory that is not in the repo. Is this file too large to include in the repository or are there other reasons not to include the file?
____________

Heh, didn't know that file was in there. It's a 1.2 GB file, so there's no way to include it. Plus it's not going to be used with BOINC, so its only purpose here would be for testing with many_n_test.sh and maybe some of the other testing scripts. It's not useful for the testing we're doing in this thread.

Edit: By the way, the code hasn't changed in about a month. I just made the code and its previous changes easier to access.
____________

Alright, people, I need HELP with MSVC++.

I tried compiling the source with VC++ 2008 Express. The files compile fine, but when linking, it's like no file sees any other file's header. I included the header files - even some new ones to replace missing Linux versions, so I'm not sure what's going on.

If you know anything about MSVC++ (since I don't), could you please take a look at my source code?

Thanks!

P.S. What all has to be included in the source code to save the proper build instructions? Does the .sln file need to be there? I really want to avoid including the gigantic .ncb file.
____________

Tanya, angle brackets (<>) are used for including non-user-written libraries that are (or should be) in your compiler path. You use the quotes ("") when what you're including is in the same directory or in the directory of another file that includes it. If it's still not found, I think it then checks the compiler path.
____________

I also noticed that a couple of the headers you are trying to load don't appear to exist, "util.h" and "gfn_app.h" There is a "putil.h" so maybe the name was just mistyped.

Getting better. It seems most of my problem was self-inflicted. I read something about how to get rid of

LINK : warning LNK4098: defaultlib "LIBCMT" conflicts with use of other libs; use /NODEFAULTLIB:library

Alright, I think I have a working binary! So if you have Win32 or Win64 and want to test it, please download my source and binary zipfile, and run the usual test on the binary in the Release folder.

Next steps include making it work for BOINC and fixing a checkpointing bug in *all other versions*. Don't let me forget to do that!
____________

I downloaded the zipfile and tryed to run the exe in the release folder. I got a message that the program can't start because cudart.dll is missing from my computer. I think I may have found a place to get the cudart.dll. Do I need to get it and put it in the directory with the exe, or is there something else I need to do?

i7-920 (stock clocks)
6GB RAM
Vista Home Premium 64-bit [Version 6.0.6002]
BOINC suspended for tests
9500GT (512mb, factory OC card, 191.07 driver)


C:\Users\Scott\Downloads\ppsieve-cuda-vc\Release>ppsieve-cuda.exe -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -z normal -q

ppsieve version cuda-0.1.1 (testing)
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000
Sieve started: 42070000000000 <= p < 42070030000000
Thread 0 starting
Detected GPU 0: GeForce 9500 GT
Detected compute capability: 1.1
Detected 4 multiprocessors.
p=42070007340033, 1.490K p/sec, 0.01 CPU cores, 24.5% done. ETA 04 Jun 00:09

all factors match.



____________
141941*2^4299438-1 is prime!

p=42070007340033, 1.490K p/sec, 0.01 CPU cores, 24.5% done. ETA 04 Jun 00:09

all factors match.

p=42070007340033, 1.490K p/sec, 0.01 CPU cores, 24.5% done. ETA 04 Jun 00:09

all factors match.

I did a spit take! But then I realized that's the wrong line. What did the line that starts with "Elapsed time" say?