Simultaneous CPU and GPU Primality Testing on Windows System

[jwnutter]

I run Prime 95 v29.8 Build 6 on a x64 Windows 10 system that has an 8-core Intel CPU running at 4.9GHz, 64gb of DDR4-3200 RAM, and a single NVIDIA RTX 2080 Super (3,072 CUDA Cores and 8 GB of DDR6). I use the CPU for PRP testing, however, I rarely use the GPU to anywhere near it's full capabilities.

Is it possible to support the GIMPS initiative by PRP testing with my CPU while also running a worker on my GPU? The work type isn't really important to me, whatever makes the most sense.

Any links to discussions on this topic would be greatly appreciated.

Thanks in advance for your help.

[Uncwilly]

gpuOwl will run PRP on the GPU. See this area of the forum:
https://mersenneforum.org/forumdisplay.php?f=171

[kriesel]

Originally Posted by jwnutter

NVIDIA RTX 2080 Super (3,072 CUDA Cores and 8 GB of DDR6). I use the CPU for PRP testing, however, I rarely use the GPU to anywhere near it's full capabilities.

Is it possible to support the GIMPS initiative by PRP testing with my CPU while also running a worker on my GPU? The work type isn't really important to me, whatever makes the most sense.

It is possible to do PRP on the RTX2008Super using gpuowl. Or to do P-1 using gpuowl. It is likely, given the relative performance of those applications on that gpu model, to be more help to the project doing TF with mfaktc. RTX20xx are very much faster at TF. (If it was a Radeon VII, PRP would be a better use for that.)
Try a few things, and do what you enjoy.
http://www.mersenneforum.org/showpos...91&postcount=2 contains a PDF attachment with selected specs, and links to download locations and forum threads for numerous programs.
Yes it is possible to run on cpu and gpu(s) in parallel with different programs, and routine.

[jwnutter]

Thank you both for your quick replies. I've reviewed the posts related to mfaktc, however, I may be a bit out of my league. The mfaktc-0.21 readme indicates that "It is assumed that [I've] already set up [my] compiler and CUDA environment." But unfortunately, I wouldn't know where to begin this process.

So, maybe one day it will be possible to develop an *.exe based GUI for the layman in a similar situation.

Thanks again for your help! Much appreciated.

[Uncwilly]

This post has a recent Win64 exe file in the 7zip archive.

https://mersenneforum.org/showpost.p...postcount=2161

[jwnutter]

Thanks!

I've downloaded gpuowl and created a worktodo.txt file in the same directory (which contains a PRP test value), but unfortunately I'm getting the following error when I launch gpuowl-win. The error appears in a new file titled gpuowl.txt. Did I miss a step somewhere along the way?

Note: not found 'config.txt'
device 0, unique id ''
Exception gpu_error: DEVICE_NOT_FOUND clGetDeviceIDs(platforms[i], kind, 64, devices, &n) at clwrap.cpp:77 getDeviceIDs
Bye

[Uncwilly]

I am moving this thread to a location where you will get more help.

[kriesel]

Originally Posted by jwnutter

The work type isn't really important to me, whatever makes the most sense.

Mfaktc will be a much better application of the RTX 2080 Super's capabilities. That gpu is particularly excellent at TF, and relatively weak at DP computations, so much so that I haven't bothered to try gpuowl on mine yet.
See https://www.mersenne.ca/mfaktc.php (3072, ~3 times the speed of a Radeon VII at TF)
and https://www.mersenne.ca/cudalucas.php (no benchmark submitted yet; by extrapolation from similar RTX models, estimated ~1/4 the speed of a Radeon VII at PRP)
Advantage TF (mfaktc), decisively

You don't have to recompile mfaktc if you have a usable CUDA10 compatible version downloaded for the RTX2080Super. Download here

To experiment with gpuowl anyway:
Newer versions of gpuowl for Windows (and a link to a list of older) are at https://mersenneforum.org/showthread.php?t=25624
Run gpuowl-win -h and look for a list of devices between the options and the fft lengths. If it doesn't list your RTX2080Super, check you have OpenCL installed and working with some other utility.
Create a config.txt, including a carriage return. Mine looks like this:

-device 0 -user kriesel -cpu condorella/rx480 -yield -maxAlloc 7500

(I include -cpu systemname/gpuid because I have more than one of each. When I run multiple workers per gpu I distinguish those too by appending -w2 or -w:3 to the gpuid for additional workers.)

[kriesel]

Set up as previously described in this thread, in separate folders. Stop mfaktc temporarily.
Gpuowl v6.11-325 failed re atomics pragma.
Gpuowl V6.11-318 ran successfully.
M107.8M P-1 factored with default bounds 1M, 30M in both stages 3.4 hours.
Found a 70.5 bit factor that was missed in someone else's earlier TF.
Before the P-1 gcd completed, it had started PRP and indicated about 5 days to do what a RadeonVII could do in about one day.

2020-06-27 09:28:50 dodo/rtx2080super 107803049 OK        0 loaded: blockSize 400, 0000000000000003
2020-06-27 09:28:55 dodo/rtx2080super 107803049 OK      800   0.00%; 4035 us/it; ETA 5d 00:50; 7ad8f79445647366 (check 1.88s)

If put to TF instead, it could probably find factors for multiple first-test wavefront exponents in those same 5 days.
For each factor found, it would save a first test, and either a double check (and possibly third LL), or eventually, a proof, and a verification.
(See the next big development for GIMPS)

Randomly selecting 5 days June 1 through 5 local time, that same gpu had found
[Thu Jun 04 04:50:48 2020]
UID: Kriesel/dodo-rtx2080super, M333039731 has a factor: 231825750049276336445449 [TF:77:78:mfaktc 0.21 barrett87_mul32_gs]
[Thu Jun 04 17:10:13 2020]
UID: Kriesel/dodo-rtx2080super, M333045971 has a factor: 184152303502604990320967 [TF:77:78:mfaktc 0.21 barrett87_mul32_gs]

These 5 days of mfaktc on this RTX2080 saved the project:
Trial factoring 333039731 saved 173 GhzD of P-1 factoring and probably 4954 GhzD of primality testing (or twice that).
Trial factoring 333045971 savings are comparable. Over 10,000 GhzD saved corresponds to ~ 3 weeks of RadeonVII gpuowl time saved, by running mfaktc on the RTX2080.
(173+4954*1.03) * 2 exponents = 10551 GhzD/day. Divide by optimal gpuowl performance seen, 510GhzD/day (5M, linux, rocm) = 20.7 days Radeon VII. (Or ~107. days of RTX2080 gpuowl time)
(173+4954*2.04) * 2 exponents with LL & LLDC = 20558; /510 = 40.3 days RadeonVII. (Or ~7 months of RTX2080 gpuowl time)
These are overstated slightly, by the few percent chance of P-1 finding a factor, avoiding primality tests.

[jwnutter]

Wow, thanks for all the info. This all sounds amazing. But like I said previously I might be a bit out of my league here.

I'm still having a difficult time installing mfaktc-0.21. I've tried mfaktc-0.21.win_cuda80-2047 and mfaktc-0.21.win_cudea100-2047 and keep getting the same system error during install indicating cudart64_80.dll was not found.

I've also been reading about how to install OpenCL on my machine for Gpuowl, and it looks like this process is a bit complex as well.

I'll have some time this evening to do some more reading. That said, if you know where I might find a step-by-step guide for installing either/both of these programs please share. At this point I'm just trying to piece things together using this forum and google searches. However, I'm not a programmer so much of what I find provides more questions then answers. =)

Thanks again!

[kriesel]

Originally Posted by jwnutter

Wow, thanks for all the info. This all sounds amazing. But like I said previously I might be a bit out of my league here.

I'm still having a difficult time installing mfaktc-0.21. I've tried mfaktc-0.21.win_cuda80-2047 and mfaktc-0.21.win_cuda100-2047 and keep getting the same system error during install indicating cudart64_80.dll was not found.
Thanks again!

Hang in there. I started climbing the gpu computing learning curve in March 2017. It takes a while.

Your RTX 2080 Super will require CUDA10-capable driver, CUDA 10 dll, and CUDA10-capable executable for mfaktc. (It varies by gpu model family; GTX10xx CUDA8)

CUDA Dlls are in the CUDA_dlls folder of https://download.mersenne.ca/

You may want to try GPU-Z as a utility on your Windows10 system to see an indication of what the computer thinks is installed for your gpu (CUDA opencl openGl etc), graphically monitor gpu parameters, maybe even log them if you want. One of many utilities listed in https://www.mersenneforum.org/showpo...74&postcount=6