GPU CORE22 0.0.2 coming to FAH - p11737-9 feedback thread

[toTOW]

See rafwiewiora's announcement thread here : viewtopic.php?f=24&t=32070

Please use this thread to report you experiences with the new core and the test project.

[MeeLee]

The new core immediately dumps WUs as a cause of too high of an overclock, where as 21 resumes from prior savestate.

I've seen as high as +3M PPD on a 2080Ti, but need to lower overclock by 1-2Mhz compared to core 21.

Other than that, I have had a WU hang, meaning after the results were uploaded, no new Wu was downloaded. I'm not sure if this is related to core 22, as other users had reported slow downloads as well (about 2 to 5 days ago).

[toTOW]

Yes, the test project (p11737) is more demanding for the GPU than we're used to. Some overclocks (or power optimizations) might become unstable and must be reviewed.

[MeeLee]

My main issue is the immediate dump of the WU, rather than the continuation from the last savestate.

[MeeLee]

BTW, does core 22 also work on CPU? Or only GPU?

[toTOW]

My main issue is the immediate dump of the WU, rather than the continuation from the last savestate.

We need logs showing the errors ...

BTW, does core 22 also work on CPU? Or only GPU?

It's the continuation of core 21, so it's a GPU core.

[rafwiewiora]

6% failure rate in the first batch -- not great, not too bad -- (cf. core21 failure rate on full F@h is ~5%) -- I call this good enough to start a second batch of about 500 WUs - distributing now. Very interestingly, 8/9 failures seen in the first batch were due to the same error - `clWaitForEvents`, and I'm investigating further.

MeeLee -- any chance you were seeing the clWaitForEvents errors? I'm looking for someone who's seeing this and willing to help us and the OpenMM team debug.

[rafwiewiora]

Currently seeing only a 3% failure rate in the second batch, so I'm leaving the ADVANCED flag on constantly now.

[bruce]

The new core immediately dumps WUs as a cause of too high of an overclock, where as 21 resumes from prior savestate.

I've seen as high as +3M PPD on a 2080Ti, but need to lower overclock by 1-2Mhz compared to core 21.

That's useful information for other overclockers reading this topic, but it's not something FAH is going to do anything about. Officially, FAH does not support overclocking. If you do overclock you're operating totally on your own. If the margin/tolerance for overclocking changes, you're responsible for dealing with it. (but you already knew that).

There will be an unknown percentage of FAH donors who will not be critically overclocked, and their global performance increase will be good for science. To time required to compensate for those of you who have to dial things back will be a temporary issue.

[road-runner]

Looks like I had a few of them in the benchmark viewer of HFM

Code: Select all

Project ID: 11737
 Core: OPENMM_22
 Credit: 7498
 Frames: 100


 Name: 1070s little room Slot 02
 Path: 192.168.1.11-36330
 Number of Frames Observed: 197

 Min. Time / Frame : 00:01:09 - 823,910.4 PPD
 Avg. Time / Frame : 00:01:10 - 806,318.1 PPD


 Name: 1080 living room Slot 01
 Path: 192.168.1.39-36330
 Number of Frames Observed: 300

 Min. Time / Frame : 00:01:01 - 991,194.7 PPD
 Avg. Time / Frame : 00:01:01 - 991,194.7 PPD


 Name: 1080 living room Slot 02
 Path: 192.168.1.39-36330
 Number of Frames Observed: 300

 Min. Time / Frame : 00:01:00 - 1,016,077.0 PPD
 Avg. Time / Frame : 00:01:01 - 991,194.7 PPD


 Name: 1080 TI Upstairs Slot 01
 Path: 192.168.1.225-36330
 Number of Frames Observed: 100

 Min. Time / Frame : 00:00:43 - 1,674,753.8 PPD
 Avg. Time / Frame : 00:00:43 - 1,674,753.8 PPD


 Name: 1080 TI Upstairs Slot 02
 Path: 192.168.1.225-36330
 Number of Frames Observed: 300

 Min. Time / Frame : 00:00:44 - 1,617,985.2 PPD
 Avg. Time / Frame : 00:00:44 - 1,617,985.2 PPD


 Name: home Slot 01
 Path: 127.0.0.1-36330
 Number of Frames Observed: 200

 Min. Time / Frame : 00:00:44 - 1,617,985.2 PPD
 Avg. Time / Frame : 00:00:45 - 1,564,353.0 PPD

[bruce]

Looks like I had a few of them in the benchmark viewer of HFM

Right. Small variations are to be expected. Live with it.

[road-runner]

Looks like I had a few of them in the benchmark viewer of HFM

Right. Small variations are to be expected. Live with it.

There different cards 1070, 1080 and 1080ti lot more points than core 21

[bruce]

We're not comparing PPD with a 1070 with a 1080 or with a 1080Ti.

Where do you see a variation of > 10% for a specific GPU and project? I don't see anything in your report that suggests otherwise. If I missed it, please point it out.

[road-runner]

We're not comparing PPD with a 1070 with a 1080 or with a 1080Ti.

Where do you see a variation of > 10% for a specific GPU and project? I don't see anything in your report that suggests otherwise. If I missed it, please point it out.

I didnt say anything about point variation you did, I was just pointing out it was different cards is why there different. I was just posting what the viewer showed I was happy to see more points

[Zangetsu]

FAHCore22:
project 11737
3 X Vega64 on PCIe 1X gen 2 with Risers v007
celeron G4400.
no over/underclock.
1 GPU got the project.
Gives 568k PPD with TPF of 1min 29.

core21 on overall projects (disregarding outliners) gave 600k PPD, so on risers(or on low end CPU) it gives a slight bottleneck.
It runs stable thou

*EDIT the PPD were wrong, now corrected