mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   storm5510 (https://www.mersenneforum.org/forumdisplay.php?f=168)
-   -   A New Toy (https://www.mersenneforum.org/showthread.php?t=25987)

storm5510 2020-09-21 15:38

A New Toy
 
4 Attachment(s)
[B]Zotac RTX 2080 Amp[/B]. I had to remove a front case fan to get this to fit. I definitely need a larger case.

I was concerned about the driver I downloaded from [I]Nvidia[/I]. The screens indicated it was a 3000 series driver. However, it installed properly and displayed the correct series on a little widget I keep on my desktop for monitoring temperatures.

It passed [I]mfaktc-2047's[/I] self-test with no errors. I had a few TF's to run for a benchmark. It never dropped below 2900 GHz-d/day. The temperature leveled off in the upper 60's on the C scale. In a large case with more fans, it might run a few degrees cooler. I downloaded the latest version of GPU-Z. It indicated a board power draw of just under 200W.

It is not my intention to use this for TF work. My preferred is wavefront P-1 using [I]gpuOwl[/I]. I may mix in some LL-DC's as well. My 1080 took around 3.5 hours to run a wavefront P-1. A recent LL-DC in the 57-million area took 39 hours to complete. I don't know how this GPU will handle those work types. I guess I will find out. I have attached some images for inspection.

kriesel 2020-09-21 23:41

[QUOTE=storm5510;557478][B]Zotac RTX 2080 Amp[/B]
...
It is not my intention to use this for TF work. My preferred is wavefront P-1 using [I]gpuOwl[/I]. I may mix in some LL-DC's as well. My 1080 took around 3.5 hours to run a wavefront P-1. A recent LL-DC in the 57-million area took 39 hours to complete. I don't know how this GPU will handle those work types. I guess I will find out. I have attached some images for inspection.[/QUOTE]Congrats. This model is an excellent gpu for TF. Its DP performance is weak by comparison so don't expect much of it or use it much for P-1, LL, or PRP.
Try using "c:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe" -i 0 -pl 0 or similar to reduce power and get even better GhzD/kwhr. Have fun!

storm5510 2020-09-22 13:07

[QUOTE=kriesel;557511]Congrats. This model is an excellent gpu for TF. Its DP performance is weak by comparison so don't expect much of it or use it much for P-1, LL, or PRP.


Try using "c:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe" -i 0 -pl 0 or similar to reduce power and get even better GhzD/kwhr. Have fun![/QUOTE]

Thanks! I use [I]Afterburner[/I] to throttle it down. It's at 90% capacity at the moment. The core temperature stays around 65°C, so that's not too bad. I'm running GPUto72 TF's to 76 bits. It takes about 35 minutes each. Nearly all the P-1's I ran the past several months were factored to 76. So, I'm right on target, for now. There was chatter on their thread a while back about a future need to run to 77 bits. I don't know what their status is. I would have to go look.

Thanks again! :smile:

chalsall 2020-09-22 13:41

[QUOTE=storm5510;557556]There was chatter on their thread a while back about a future need to run to 77 bits. I don't know what their status is. I would have to go look.[/QUOTE]

That was *before* the new Cert methodology, and "optimal" was actually 78.

76 bits is now where we're targeting since the CERT runs take *much* less time than DCs.

storm5510 2020-09-22 14:36

1 Attachment(s)
[QUOTE=chalsall;557559]That was *before* the new Cert methodology, and "optimal" was actually 78.

76 bits is now where we're targeting since the CERT runs take *much* less time than DCs.[/QUOTE]

Thank you for the confirmation. I am where I need to be. I set GPU72Config to pledge 76 and to option 2, Highest TF level. I am doing this in semi-automatic mode. Assign is automatic but submit is manual. I had issues with [I]Misfit[/I] in the past. It is not something I wish to repeat. Keeping it running is the important thing.

========================

I managed to get the 2nd front panel fan back in. With the GPU being in place, it was a simple matter of sliding it past the end. The 90% capacity core reading dropped to 60°C. Dropping the capacity had a minimal affect on throughput. Still above 2800 GHz-d/day. According to GPU-Z, the board power consumption remains around 200W with a GPU load of 85%. It's tight in this case as the image below demonstrates, so heat is a concern.

kriesel 2020-09-22 16:38

[QUOTE=storm5510;557556]I use [I]Afterburner[/I] to throttle it down. [/QUOTE]Do you shut Afterburner down after making the power setting? Some people prefer nvidia-smi for power setting and occasional sensor monitoring, because as a command line utility, it has low overhead.

storm5510 2020-09-22 23:42

[QUOTE=kriesel;557573][B]Do you shut Afterburner down after making the power setting?[/B] Some people prefer nvidia-smi for power setting and occasional sensor monitoring, because as a command line utility, it has low overhead.[/QUOTE]

No, I don't. I minimize it to the tray. It has never created a problem doing it this way. I also have it set to load at system startup. I looked for it in the Task Manager. It uses 3.5 MB of RAM and intermittent 0.1% of CPU time. I don't have a problem with this.

moebius 2020-10-26 13:20

Can you please do a benchmark/timing for gpu owl - prp 77936867, regarding post [URL="https://www.mersenneforum.org/showpost.php?p=560841&postcount=2533"]2533[/URL]. The GTX 1080 without TI is still missing at the list.gpuowl version 6.11.364. to 6.11.380 should be the fastest. The benchmark values ​​on mersenne.ca are unfortunately only of limited use, as mostly for CudaLucas and sometimes very different from reality.

storm5510 2020-10-26 22:19

[QUOTE=moebius;561175]Can you please do a benchmark/timing for gpu owl - prp 77936867, regarding post [URL="https://www.mersenneforum.org/showpost.php?p=560841&postcount=2533"]2533[/URL]. The GTX 1080 without TI is still missing at the list.gpuowl version 6.11.364. to 6.11.380 should be the fastest. The benchmark values ​​on mersenne.ca are unfortunately only of limited use, as mostly for CudaLucas and sometimes very different from reality.[/QUOTE]

A benchmark. I am not sure how to do this. Can you elaborate?

moebius 2020-10-27 05:50

gpuowl is a command line software that runs with openCL, it does not require a CUDA installation. So just download the Windows version. The linux version should run a few percent faster,but it is more difficult to get running, because e.g. ROCm has to be installed.[URL="https://www.mersenneforum.org/showthread.php?t=25601"]How to[/URL] The Windows version can be used immediately.
[URL="https://www.mersenneforum.org/showpost.php?p=551594&postcount=8"]Gpuowl-win v6.11-364-g36f4e2a[/URL], [URL="https://www.mersenneforum.org/showpost.php?p=556311&postcount=40"]gpuowlv6.11-380-g79ea0cc-ubuntu[/URL]
In principle, you only have to start gpuowl-win -iters 200000 -prp 77936867on the command line.
You can possibly make gpuowl run faster. with e.g. the following opencl options
[CODE]gpuowl.exe -carry short -use CARRY32,ORIG_SLOWTRIG,IN_WG=128,IN_SIZEX=16,IN_SPACING=4,OUT_WG=128,OUT_SIZEX=16,OUT_SPACING=4 -block 100 -maxAlloc 10000 -B1 750000 -rB2 20 -prp 77936867
[/CODE]
examples for timming [URL="https://www.mersenneforum.org/showthread.php?p=561206#post561206"]post561206[/URL]

storm5510 2020-10-27 13:17

1 Attachment(s)
I believe the attached text document is what you need. Actually, I let it run overnight then trimmed down the results.

I have used [I]gpuOwl[/I] (Windows) in the past so I am familiar with it. I used a line capture utility called [I]mtee[/I] to duplicate what appeared on the screen into a text file. It is 300,000 iterations, so you can trim out what you do not need. The version and config lines are at the top also.


All times are UTC. The time now is 05:15.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.