Use processors and GPU for higher throughput?

[Runtime Error]

Hi,

I've been running PRPs on a few machines with Intel Core i7-6700 @ 3.40GHz processors. Recently, I've realized that each of these has a NVIDIA GeForce GT 730. I'm new to GPU computing (other than whatever Matlab does automatically), and I realize these aren't particularly good GPUs, but:

Could I get more overall throughput if I run Prime95 on several of the processor cores and one of the GPU Mersenne programs (maybe trial factoring)? Any tips would be appreciated.

Thanks in advance, and happy holidays!

[Uncwilly]

Yes, run a factoring program on the GPU. You should be able to get about 50 GHz-days/day out of each one.

Someone will tell you which is the right program to use.
Set up the program. Use GPU72 to get work and MISFIT to do the managing of the work and submitting the results.

[kriesel]

Run prime95 on all the cpu cores; and gpu applications on each gpu. You may also be able to run LL tests with CUDALucas on the GT 730. Since the gpu model has a 1:12 SP:DP ratio, it is not a bad choice for LL. Start with double-checks to get a sense of the time it takes to finish one and to check your gpu's reliability.
Gpuowl is not an option because of the OpenCL1.1 version, versus 2.0 required.
The 2GB gpu ram is small for P-1 factoring.
To get the most out of your gpus, regardless of application, do the tuning for the application you use.

https://www.techpowerup.com/gpu-spec...e-gt-730.c2590
See also http://www.mersenneforum.org/showpos...91&postcount=2 and the rest of https://www.mersenneforum.org/showthread.php?t=24607

[Gordon]

Originally Posted by Uncwilly

Yes, run a factoring program on the GPU. You should be able to get about 50 GHz-days/day out of each one.

Someone will tell you which is the right program to use.
Set up the program. Use GPU72 to get work and MISFIT to do the managing of the work and submitting the results.

Maybe he can but at 33W to get 53 Ghz days/day is it really worth it?

I have a 1660-Ti and for less than 4x the power (120W vs 33W) I get 42x the throughput (2250 vs 53).

[kriesel]

Originally Posted by Gordon

is it really worth it?

It's a personal decision. It's winter here. People run electric space heaters that don't compute at all. And run mfakto on IGPs with similar GhzD/day/watt ratios as mfaktc will likely yield on the GT 730 he asked about. Or electric train sets and holiday lights, that produce 0 GhzD/day, because it pleases them to do so. He asked about using the gpu hardware he already has. NVIDIA GT 730 implies mfaktc or CUDALucas, or possibly CUDAPm1 on exponents below 200M.

[Runtime Error]

Thanks for the helpful replies!

I do not pay the electric bill for these machines, but I only get them when no one else is logged on and they haven't been taken offline (e.g. winter maintenance). Otherwise "worth it" reignites the "usefulness" conversation; I'll stay out of it. But Gordon, wow from those numbers, it seems like hunting on GPUs is cheaper than on CPUs both in upfront cost and marginal cost. Is that conclusion correct?

I've set CUDALucas up on one machine for double checks. Currently getting 29.60ms/iter on a 56M exponent. Task Manager says it's only using 10% of the GPU capacity, but perhaps that calculation is similar to how Prime95 only uses 50% due to hyperthreading. Does that sound about right?

I'm working my way through kriesel's documentation. There is a lot to unpack! My next goal is to set up MISFIT per Uncwilly's suggestion. Automating assignment requests and result submission is really important so I can avoid juggling manual requests.

Thanks again, and Happy New Year!

[kriesel]

Originally Posted by Runtime Error

Thanks for the helpful replies!

I do not pay the electric bill for these machines, but I only get them when no one else is logged on and they haven't been taken offline (e.g. winter maintenance). Otherwise "worth it" reignites the "usefulness" conversation; I'll stay out of it. But Gordon, wow from those numbers, it seems like hunting on GPUs is cheaper than on CPUs both in upfront cost and marginal cost. Is that conclusion correct?

I've set CUDALucas up on one machine for double checks. Currently getting 29.60ms/iter on a 56M exponent. Task Manager says it's only using 10% of the GPU capacity, but perhaps that calculation is similar to how Prime95 only uses 50% due to hyperthreading. Does that sound about right?

Sounds like Windows 10. The gpu monitoring is a work in progress there, and in my opinion quite unreliable on gpu load. Try GPU-Z for continuous graphical, or nvidia-smi for very low overhead character-cell gpu stats on request. (Both have optional logging.) I usually use a tiny batch loop for nvidia-smi, so it gives me an update whenever I press a key in that command prompt box, otherwise nvidia-smi just sits idle showing the last results.

title nvidia-smi in %cd%
:nv
"c:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe"
pause
goto nv

Gpu utilization displayed is usually 95-100%. There are some exceptions to that; fast gpus running mfaktc and not yet well tuned; CUDAPm1 during gcd on a cpu core; brief periods in gpuowl during the GEC.

C:\Users\Ken\Documents>goto nv

C:\Users\Ken\Documents>"C:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe
Fri Dec 27 07:39:29 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 378.66                 Driver Version: 378.66                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro 2000        WDDM  | 0000:0F:00.0     Off |                  N/A |
|100%   83C    P0    N/A /  N/A |     85MiB /  1024MiB |     99%      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX 106... WDDM  | 0000:28:00.0     Off |                  N/A |
| 53%   71C    P2    83W / 120W |    232MiB /  3072MiB |     99%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0      5860    C   ...n\My Documents\tf-q2000\mfaktc-win-64.exe N/A      |
|    1      5944    C   ...CUDALucas2.06beta-CUDA6.5-Windows-x64.exe N/A      |
+-----------------------------------------------------------------------------+

C:\Users\Ken\Documents>pause
Press any key to continue . . .