CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW) - Page 231

ATH · 2016-10-01, 11:18

The one I'm using that works is CUDA6.5. But they claim now that the bugs in CUDA7 and 7.5 should be fixed in CUDA8. That's why I wanted to try it, but I guess it is not that important.

flashjh · 2016-10-02, 04:55

Quote:

Originally Posted by ATH

The one I'm using that works is CUDA6.5. But they claim now that the bugs in CUDA7 and 7.5 should be fixed in CUDA8. That's why I wanted to try it, but I guess it is not that important.

Hello everyone!

I uploaded a new Windows file to sourceforge with the CUDA 8.0 version of CUDALucas. It's here.

Remember to download the CUDA 8.0 Libs from here.

Please note that I couldn't locate the cufft32_80.dll file, so I didn't compile a 32 bit version for CUDA 8.0. As far as I can tell nVidia didn't include it with this version of CUDA. If someone has it, get it to me and I'll compile and upload the other files.

I ran several tests on a GTX 750ti, but I don't have any of the newer cards to see if this works or not. Can someone test this and let me know if it's producing errors or zero results similar to the other problems? Thanks!

Karl M Johnson · 2016-10-02, 12:09

Hello, Jerry!

Thanks for the new binary, I can confirm that on a GTX 1080 it's a bit faster than the old v8.0 RC binary & the exponents match.

II did a quick test on M1000003, and the runtime was 76s for the old binary and 66s for the new one.

Quote:

Originally Posted by flashjh

Hello everyone!

I uploaded a new Windows file to sourceforge with the CUDA 8.0 version of CUDALucas. It's here.

Remember to download the CUDA 8.0 Libs from here.

Please note that I couldn't locate the cufft32_80.dll file, so I didn't compile a 32 bit version for CUDA 8.0. As far as I can tell nVidia didn't include it with this version of CUDA. If someone has it, get it to me and I'll compile and upload the other files.

I ran several tests on a GTX 750ti, but I don't have any of the newer cards to see if this works or not. Can someone test this and let me know if it's producing errors or zero results similar to the other problems? Thanks!

mognuts · 2016-10-09, 09:58

Just run a benchmark using my new card (GTX 780ti) and the various CUDA binaries.
Driver = 373.06. Test number = M2976221. Card not driving monitor.

CUDA 4.2 32bit = 13min 11sec
CUDA 4.2 64bit = 13min 01sec
CUDA 5.0 32bit = 12min 18sec
CUDA 5.0 64bit = 12min 24sec
CUDA 5.5 32bit = 12min 24sec
CUDA 5.5 64bit = 12min 36sec
CUDA 6.0 32bit = 12min 13sec
CUDA 6.0 64bit = 12min 19sec
CUDA 6.5 32bit = 12min 19sec
CUDA 6.5 64bit = 11min 38sec
CUDA 8.0 64bit = 13 min 56sec

I'm not that surprised at CUDA 8.0, but CUDA 6.5 64bit was a bit out of character compared to how by GTX580 behaves.

ATH · 2016-10-09, 10:47

I did a -cufftbench 2592 8192 20 on the different versions (only the 64 bit versions) so it does 20x 50iterations on each FFT and takes the average.

In most cases 6.5 is fastest but a few of them has 8.0 as the fastest (on a Titan Black, but this is probably GPU dependent).

CUDA 4.2 was quite a bit slower on all of them, so I left it out.

Code:

		     8.0     6.5       6.0      5.5      5.0
 2592   48471289   1.6135   1.6683   1.6897   1.6145   1.6166
 2744   51250889   2.0056   1.8606   1.8682   1.9980   1.8714
 3136   58404433   2.0937   2.0480   2.0710   2.0201   2.0337
 3200   59570449   2.4195   2.3907   2.4056   2.4150   2.4175
 3240   60298969            2.4266   2.4388   2.4404   2.4435
 3375   62756279            2.5147   2.5301
 3888   72075517            2.5348   2.5584            2.4558
 4000   74106457   2.4631   2.5498   2.5821   2.4590   2.4639
 4096   75846319   2.5115   2.5800   2.5976   2.5200   2.5375
 4320   79902611   3.2614   3.2573   3.2760   3.2685
 4374   80879779   3.3753   3.2784   3.2924   3.2946   3.3003
 4500   83158811            3.3535   3.3845   
 4536   83809729            3.3810   3.4006
 5184   95507747   3.4279   3.3836   3.4208   3.2994   3.3273
 5292   97454309                              3.9568
 5488  100984691   3.8843   3.8345   3.8336   3.9979   3.7484
 5600  103000823   4.1630   4.3082   4.3430   4.1882   4.1943
 5832  107174381   4.5283   4.3644   4.3842   4.3847   4.3903
 6000  110194363            4.5328
 6048  111056879   4.5338                     4.5138   4.5275
 6075  111541967                                       4.5586
 6125  112440191            4.5530   4.5645   4.5780   4.5867
 6144  112781477   4.6858   
 6250  114685037   4.8079   4.6418   4.6620   4.6714   4.6787
 6272  115080019            4.6678   4.6820   4.6835   4.6908
 6400  117377567   4.9088   4.7531   4.7721   4.7719   4.7809
 6480  118813021   4.9438   4.8401   4.8670   4.8669
 6561  120266023   5.1164   4.8818   4.9012   4.8968
 6750  123654943   5.1792            5.0356   5.0432
 6912  126558077   5.1957         
 7776  142017539   5.2343   5.0001   5.0441            4.8219
 8000  146019329   5.2537   5.1762   5.2350   5.0527   5.0997
 8192  149447533   5.3838   5.2219   5.2593   5.1617   5.2106

kladner · 2016-10-09, 11:47

Many thanks for the info above. I had forgotten what version of CuLu I was running, and I'm still not sure. However, I switched in the CUDA 6.5, 64 bit version, and the GTX460 went from 7.4304 to 7.3026 ms/it.

airsquirrels · 2016-10-09, 12:33

I upgraded a couple systems from 6.5 to 8.0 last night. My Titan/Black based systems saw a 4.6% gain after letting everything burn up to normal temps. I'm not sure how effective any one off benchmark will be given the amount of active thermal/power management on most GPUs. Titan X from 7.5->8.0 saw no improvement.

storm5510 · 2016-11-06, 04:06

Quote:

Originally Posted by flashjh

I uploaded a new Windows file to sourceforge with the CUDA 8.0 version of CUDALucas. It's here.

Remember to download the CUDA 8.0 Libs from here.

Thank you!

I had been looking for the libraries for a day or so. I'm only lacking an INI file. It replies "using defaults for non-specified options." then goes on. Where can I find the INI file?

Thanks,
Dwayne.

storm5510 · 2016-11-06, 05:30

Disregard the request above. I found it with a bit more searching.

I changed the screen output options so I could see what was going on. I reserved one doublecheck from PrimNet. CUDALucas reports it can complete it in a little under six days.

Now for my quandary: Prime95 indicates it can do the test in the same amount of time. Six days for a LL test is nothing to sneeze at. CUDALucas did not seem to be utilizing my GPU as much as I thought it might. It's a GTX-750Ti. I could tell by observing the core temperature. mfaktc runs it in the upper 50's one the C scale. CUDALucas only made it into the low 40's.

Just in case anyone wonders about my setup, it all runs with CUDA 8.

kladner · 2016-11-06, 06:54

Quote:

Originally Posted by storm5510

Disregard the request above. I found it with a bit more searching.

I changed the screen output options so I could see what was going on. I reserved one doublecheck from PrimNet. CUDALucas reports it can complete it in a little under six days.

Now for my quandary: Prime95 indicates it can do the test in the same amount of time. Six days for a LL test is nothing to sneeze at. CUDALucas did not seem to be utilizing my GPU as much as I thought it might. It's a GTX-750Ti. I could tell by observing the core temperature. mfaktc runs it in the upper 50's one the C scale. CUDALucas only made it into the low 40's.

Just in case anyone wonders about my setup, it all runs with CUDA 8.

That card should do better. I run CuLu on a (very overclocked) GTX 460 with the 6.5 libraries. You would probably do better with 6.5, as well. CUDA 8.0 mainly seems to benefit GTX 10-series architecture. The 750ti has 640 CUDA cores, at a base clock of 1020 MHz. The 460 has 336 CUDA cores, and mine is running at 848 MHz. I did have to slow the memory from 1900 to 1700 MHz to get reliable results with CuLu.

Temperature is not a good guide comparing mfaktc and CUDALucas. In my experience, mfaktc runs a card hotter. This may have to do with greater (throttled) floating point usage under CuLu. (I could be wrong on this point.)

The 460 mentioned above, with all conditions (voltage, clock, no competition from P95, same ambient) stabilizes at 64 C with CuLu, at 7.2521 ms/it.. It does a 40.8M LLDC in 3-4 days. This is about twice as fast as an FX-8350 worker, when running 2 workers with 4 threads each. Just about any i5 or i7 chip from Sandy Bridge on would beat the snot out of the AMD CPU.

The 460 holds at 67 C running mfaktc, where it delivers ~206 GHz-d/d.. In both cases, usage was 100%, according to MSI Afterburner. This is a secondary card. It is not driving the display.

If you are running Windows, I really recommend Afterburner. Even if you don't use it to OC, it has nice, configurable monitoring functions. Finding out what the actual usage is would be a good start to analyzing your performance.

Here is a question: have you run CUFFTbench and threadbench on this card?

henryzz · 2016-11-06, 13:48

Doesn't the 460 have a much better single precision/double precision ratio than the 750ti?

2016-10-09, 09:58	#2534
mognuts Sep 2008 Bromley, England 43 Posts	CUDA binary benchmarks Just run a benchmark using my new card (GTX 780ti) and the various CUDA binaries. Driver = 373.06. Test number = M2976221. Card not driving monitor. CUDA 4.2 32bit = 13min 11sec CUDA 4.2 64bit = 13min 01sec CUDA 5.0 32bit = 12min 18sec CUDA 5.0 64bit = 12min 24sec CUDA 5.5 32bit = 12min 24sec CUDA 5.5 64bit = 12min 36sec CUDA 6.0 32bit = 12min 13sec CUDA 6.0 64bit = 12min 19sec CUDA 6.5 32bit = 12min 19sec CUDA 6.5 64bit = 11min 38sec CUDA 8.0 64bit = 13 min 56sec I'm not that surprised at CUDA 8.0, but CUDA 6.5 64bit was a bit out of character compared to how by GTX580 behaves. Last fiddled with by mognuts on 2016-10-09 at 10:32 Reason: Rounded off values for clarity

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Don't DC/LL them with CudaLucas	LaurV	Data	131	2017-05-02 18:41
CUDALucas / cuFFT Performance on CUDA 7 / 7.5 / 8	Brain	GPU Computing	13	2016-02-19 15:53
CUDALucas: which binary to use?	Karl M Johnson	GPU Computing	15	2015-10-13 04:44
settings for cudaLucas	fairsky	GPU Computing	11	2013-11-03 02:08
Trying to run CUDALucas on Windows 8 CP	Rodrigo	GPU Computing	12	2012-03-07 23:20

2016-10-01, 11:18	#2531
ATH Einyen Dec 2003 Denmark C56₁₆ Posts	The one I'm using that works is CUDA6.5. But they claim now that the bugs in CUDA7 and 7.5 should be fixed in CUDA8. That's why I wanted to try it, but I guess it is not that important.

2016-10-09, 11:47	#2536
kladner "Kieren" Jul 2011 In My Own Galaxy! 2×3×1,693 Posts	Many thanks for the info above. I had forgotten what version of CuLu I was running, and I'm still not sure. However, I switched in the CUDA 6.5, 64 bit version, and the GTX460 went from 7.4304 to 7.3026 ms/it.

2016-10-09, 12:33	#2537
airsquirrels "David" Jul 2015 Ohio 11·47 Posts	I upgraded a couple systems from 6.5 to 8.0 last night. My Titan/Black based systems saw a 4.6% gain after letting everything burn up to normal temps. I'm not sure how effective any one off benchmark will be given the amount of active thermal/power management on most GPUs. Titan X from 7.5->8.0 saw no improvement.

2016-11-06, 05:30	#2539
storm5510 Random Account Aug 2009 11110100100₂ Posts	Disregard the request above. I found it with a bit more searching. I changed the screen output options so I could see what was going on. I reserved one doublecheck from PrimNet. CUDALucas reports it can complete it in a little under six days. Now for my quandary: Prime95 indicates it can do the test in the same amount of time. Six days for a LL test is nothing to sneeze at. CUDALucas did not seem to be utilizing my GPU as much as I thought it might. It's a GTX-750Ti. I could tell by observing the core temperature. mfaktc runs it in the upper 50's one the C scale. CUDALucas only made it into the low 40's. Just in case anyone wonders about my setup, it all runs with CUDA 8.

2016-11-06, 13:48	#2541
henryzz Just call me Henry "David" Sep 2007 Cambridge (GMT/BST) 2×3³×109 Posts	Doesn't the 460 have a much better single precision/double precision ratio than the 750ti?