CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW) - Page 250

kriesel · 2019-03-09, 21:14

Quote:

Originally Posted by ldesnogu

You can find various threshold for multiplication algorithms for GMP lib implementation on many CPU here: https://gmplib.org/devel/thres

Toom22 is Karatsuba.

The bottom of https://gmplib.org/devel/ has some interesting color diagrams of many multiplication algorithms versus length and asymmetric multiplication for specific cpus. The diagonals would correspond to symmetric operand sizes.

Flammo · 2019-03-17, 09:38

Hi guys,

I'm not really sure where to start, I'm fairly new at GPU computing and cudalucas. I recently posted about errors using 2.05 and then discovered 2.06beta was the way to go.

I'm having a little problem with CUDALucas2.06beta-CUDA8.0-Windows-x64.exe

I get illegal residue every few days, which stops the computing. What does this mean? And what should I do?

GPU: Gefore GTX 750 Ti

Driver: Happens with a fresh install of Windows 10, with both native Windows NVIDIA drivers, and also the latest NVIDIA drivers 419.35.

Thanks in advance,
flammobammo

kriesel · 2019-03-17, 14:21

Quote:

Originally Posted by Flammo

Hi guys,

I'm not really sure where to start, I'm fairly new at GPU computing and cudalucas. I recently posted about errors using 2.05 and then discovered 2.06beta was the way to go.

I'm having a little problem with CUDALucas2.06beta-CUDA8.0-Windows-x64.exe

I get illegal residue every few days, which stops the computing. What does this mean? And what should I do?

GPU: Gefore GTX 750 Ti

Driver: Happens with a fresh install of Windows 10, with both native Windows NVIDIA drivers, and also the latest NVIDIA drivers 419.35.

Thanks in advance,
flammobammo

I suggest reading through https://www.mersenneforum.org/showthread.php?t=23387 and the associated attachments and links. Also, an extensive cudalucas -memtest might be revealing. By extensive, I mean on as much gpu memory as you can, for several hours at least. I've seen a card go from reliable to quite bad, unusable for CUDALucas or CUDAPm1, and eventually TF also unusable, in a year. Check temperatures in your system; check clock rate of your gpu; overclocking can hurt reliability, including factory overclock. Cutting back clock and cooling down may give your card some additional useful life. Or not, if it has significant memory error rate that's stubborn.
Like with a cpu, if it's not reliable enough to complete matching double checks reliably, it probably should not be used for primality first time checks or double checks. It might be reliable enough for P-1 or trial factoring for a while. Computing credit per day or KW-hr or $ on a gpu will be much higher with TF than P-1 or LL (typically 10 fold higher or more). P-1 has the biggest memory footprint; LL is less; TF is least. TF might "dodge" memory errors due to its small requirement. The card I had with increasingly bad memory developed lots of errors scattered in the middle of the gpu address space, judging by block number in the -memtest run. Dialing core clock and memory clock rates down did not help in that case. It was removed from service and replaced.

Good luck.

ATH · 2019-03-17, 15:29

Try for example MSI Afterburner to check the GPU temperature and set a custom fan curve, so fan will go to 100% at around 80 C or so.

I assume you ran the test with -r 1 ? It takes like 30 minutes or so, check all the residues match.

Try another CUDA binary:
https://sourceforge.net/projects/cud...iles/2.06Beta/

Your card has computer capability 3.5 so you can go all the way down to CUDA 4.0. On my old Titan Black with CC3.5 CUDA5.0 or 5.5 are fastest.
Get the required DLL files here:
https://download.mersenne.ca/CUDA-DLLs

kriesel · 2019-03-17, 20:42

Quote:

Originally Posted by ATH

Try for example MSI Afterburner to check the GPU temperature and set a custom fan curve, so fan will go to 100% at around 80 C or so.

I assume you ran the test with -r 1 ? It takes like 30 minutes or so, check all the residues match.

Try another CUDA binary:
https://sourceforge.net/projects/cud...iles/2.06Beta/

Your card has computer capability 3.5 so you can go all the way down to CUDA 4.0. On my old Titan Black with CC3.5 CUDA5.0 or 5.5 are fastest.
Get the required DLL files here:
https://download.mersenne.ca/CUDA-DLLs

Running fftbench on various CUDA levels, renaming the output file between levels run to save them, may be a quick way to compare speeds. (Rename "Geforce GTX 750 Ti fft.txt" "Geforce GTX 1050 Ti fft %Cudalavel%.txt")
Run fftbench and threadbench on the level you select, to tune for performance, and look for anomalies.
I've done that and seen anomalies below 4.2. 7.0 was also known to have issues as I recall. (And 9.0, 9.1, and maybe more.)

Flammo · 2019-03-18, 08:36

ATH and kriesel, thankyou both kindly for your time. That gives me a lot of material to check through and work it out.

James Heinrich · 2019-03-20, 01:06

Quote:

Originally Posted by ATH

Get the required DLL files here: https://download.mersenne.ca/CUDA-DLLs

It has been brought to my attention that I'm missing some CUDA versions from my DLL archive, notably:
7.5
8.0
9.2
10.1 (I have cudart101 but not cudafft101)

I've forgotten where to get these DLLs from (something about NVIDIA toolkit I think) but if somebody has these DLLs on hand and can PM/email me a link I can add them to the archive.

ATH · 2019-03-20, 16:26

CUDA 7.5 and 8 dlls:
https://sourceforge.net/projects/cud...s/CUDA%20Libs/

except 1 file from CUDA8 is missing: cufftw64_80.zip

CUDA 9.2: CUDA9_2dlls.zip

CUDA 10.1: CUDA10_1dlls.zip

Use 7-zip to unpack them.

James Heinrich · 2019-03-20, 17:00

I got most of what I needed from the source at https://developer.nvidia.com/cuda-toolkit-archive

I didn't bother filling in the missing bits of old versions e.g. 3.1 -- if someone thinks it's important I have a complete set let me know and I can go digging for those too.

I have also reorganized the archive into CUDA-version subfolders to make things easier to find.
https://download.mersenne.ca/CUDA-DLLs

kriesel · 2019-03-20, 17:05

Quote:

Originally Posted by James Heinrich

I got most of what I needed from the source at https://developer.nvidia.com/cuda-toolkit-archive

I didn't bother filling in the missing bits of old versions e.g. 3.1 -- if someone thinks it's important I have a complete set let me know and I can go digging for those too.

I have also reorganized the archive into CUDA-version subfolders to make things easier to find.
https://download.mersenne.ca/CUDA-DLLs

Excellent. I think you can ignore 4.1 and below at this point.

James Heinrich · 2019-03-20, 17:12

What about in-between versions, like 9.0 and 9.1 ?

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

2019-03-17, 09:38	#2741
Flammo Feb 2019 2²×3 Posts	Hi guys, I'm not really sure where to start, I'm fairly new at GPU computing and cudalucas. I recently posted about errors using 2.05 and then discovered 2.06beta was the way to go. I'm having a little problem with CUDALucas2.06beta-CUDA8.0-Windows-x64.exe I get illegal residue every few days, which stops the computing. What does this mean? And what should I do? GPU: Gefore GTX 750 Ti Driver: Happens with a fresh install of Windows 10, with both native Windows NVIDIA drivers, and also the latest NVIDIA drivers 419.35. Thanks in advance, flammobammo

2019-03-17, 15:29	#2743
ATH Einyen Dec 2003 Denmark 2·1,579 Posts	Try for example MSI Afterburner to check the GPU temperature and set a custom fan curve, so fan will go to 100% at around 80 C or so. I assume you ran the test with -r 1 ? It takes like 30 minutes or so, check all the residues match. Try another CUDA binary: https://sourceforge.net/projects/cud...iles/2.06Beta/ Your card has computer capability 3.5 so you can go all the way down to CUDA 4.0. On my old Titan Black with CC3.5 CUDA5.0 or 5.5 are fastest. Get the required DLL files here: https://download.mersenne.ca/CUDA-DLLs

2019-03-18, 08:36	#2745
Flammo Feb 2019 12₁₀ Posts	ATH and kriesel, thankyou both kindly for your time. That gives me a lot of material to check through and work it out.

2019-03-20, 16:26	#2747
ATH Einyen Dec 2003 Denmark C56₁₆ Posts	CUDA 7.5 and 8 dlls: https://sourceforge.net/projects/cud...s/CUDA%20Libs/ except 1 file from CUDA8 is missing: cufftw64_80.zip CUDA 9.2: CUDA9_2dlls.zip CUDA 10.1: CUDA10_1dlls.zip Use 7-zip to unpack them.

2019-03-20, 17:00	#2748
James Heinrich "James Heinrich" May 2004 ex-Northern Ontario 11·311 Posts	I got most of what I needed from the source at https://developer.nvidia.com/cuda-toolkit-archive I didn't bother filling in the missing bits of old versions e.g. 3.1 -- if someone thinks it's important I have a complete set let me know and I can go digging for those too. I have also reorganized the archive into CUDA-version subfolders to make things easier to find. https://download.mersenne.ca/CUDA-DLLs

2019-03-20, 17:12	#2750
James Heinrich "James Heinrich" May 2004 ex-Northern Ontario 6535₈ Posts	What about in-between versions, like 9.0 and 9.1 ?