R9 Fury low general performance

[umfaddi]

Hello all, apologies if this doesn't fit within this forum, but I am having trouble pinpointing the exact issue I am having folding with an R9 Fury.

I ran F@h on a brand new install of Win 7 Pro N and with the latest driver for the Fury, the ReLive driver.
It completed 2-3 WU's with an expected PPD upwards of 230k until it downloaded a "broken" WU. This WU manifested itself like this: if folding was paused it showed roughly 30% progress, however, if I resumed the folding it will get stuck at 99.99% and drive PPD numbers through the roof. I've uninstalled F@h and deleted the WU from AppData and subsequent installs of F@h have been riddled with issues.

For example, I had a new WU started: immediately it showed an ETA of 8.00 days (must be a default value), after a couple of minutes it went down to 5 hours and after a while it stabilizes to somewhere between two and three hours. So I left it on for the night (a good 8 hours) and to my surprise in the morning it still wasn't finished (?!). This has been the past trend with any install of F@h I've tried, the performance of the card is very very low and I don't know what the issue is.

Please, if anybody has any suggestions, let me know!

Please let me know if you need any additional info or files, will be uploading them once I get home.

Many thanks for your help!

[foldy]

Your R9 Fury should get around 400k PPD when folding 24/7. It takes some work units and days until the daily PPD stabilizes.
What is your exact R9 Fury vendor/model?
Can you post your FAH logfile?

[umfaddi]

Hi foldy, thanks for taking the time to look into my issue. I am not really bothered by its PPD numbers, but more of the fact that the performance is just not there and I don't know why.

It is an R9 Fury NITRO from Sapphire.

Here is the log, too big to paste in a standard message: http://docdro.id/GNibJlx .

[bruce]

The projected PPD and ETA depend on data gathered from your history with that particular project. When you get a WU from a project that's new to you, it starts with a default value (as you suspected) and as more data is gathered, it stabilizes, but that does take time. If you've processed the project previously, the initial (default) value will be based on your previous history. In the unusual situation that something major has changed on your system since your previous encounter with that project, the ETA and PPD will stabilize to a new value. [This is more likely with a CPU project than with a GPU project.]

[Joe_H]

After taking a look at the log file, there is something very strange about the GPU processing. For some periods of time it will be quite regular, then it slows way down. It is pretty obvious where the log shows multiple reports of a WU being processed on the CPU completing another percent, while there is no progress reported on the GPU slot. At other times the reports alternate, and the GPU TPF's are varying from less than a minute to nearly half an hour.

Can you monitor the GPU to see if it is down clocking at times? That would slow down processing like that.

[foldy]

Code: Select all

04:08:29:WU01:FS01:0x21:Completed 1250000 out of 2500000 steps (50%)
...TPF 3:30
04:12:56:WU01:FS01:0x21:Completed 1275000 out of 2500000 steps (51%)
...TPF 9:30
04:22:23:WU01:FS01:0x21:Completed 1300000 out of 2500000 steps (52%)
... TPF 43:30
05:05:53:WU01:FS01:0x21:Completed 1325000 out of 2500000 steps (53%)

Do you have something else running on that PC?
Is there a heat problem in your PC e.g. fans stop running and GPU must downclock?

[umfaddi]

Thank you bruce for the clarifications. However, I am still left wondering why there was such a big discrepancy when I started folding on the card compared to what I get now after I had to delete the WU because it was stuck.

Joe_H, you might be on to something here. Last night I opened WattMan and had a look at the graphs showing temperature, clocks, activity and such. The card would downclock from its default of 1050MHz to about 300MHz and at the same time the activity curve would plummet to zero and appearing as you say quite irregular. I don't actually remember if the activity percentage went down before the clock went down or the other way around, as the latter would (I assume) mean that there is an issue with the card/driver and not with the application.

No foldy, I even switch to Windows classic theme when I am folding, but the wild variance of TPF persists.

[foldy]

The Crimson Relive driver 16.12.2 is a new feature driver. Maybe try the previous non Relive driver
http://support.amd.com/en-us/download/d ... ev=16.11.5
With AMD drivers I had sometimes the problem that GPU fan was locked at 20% and so GPU thermal throttled.

[umfaddi]

Downloaded 16.11.5, gave it a run and for a moment it looked like the issue was fixed. However, 20 minutes later, the TPF would spike again.

I then noticed a pattern: about 20 minutes after folding was started and TPF value stabilised, the card would once again drop in performance so I started to go through the Power Settings to see what changes after 20 minutes of idle (just folding).

It turns out that HDDs were scheduled to turn off after 20 minutes, I disabled the setting (increased the value absurdly) and what do you know, the Fury powers on like a champ.
Just to make sure, I've retested with the latest driver (16.12.2) and the problem didn't manifest either.

Would anybody have an idea as to why turning off the HDD has a negative impact on the performance of the card, why does it make the card down clock?

Thank you all for your suggestions and your time!

[bruce]

Does your GPU's monitoring software give the change in clock rate when it's clocking down? If not, you're probably reaching the wrong conclusion.

The FAH client writes checkpoint files. When the elapsed time reaches the scheduled time for a checkpoint, it stops long enough to write that information to disk. Normally, that results in brief disk activity and it goes back to folding the next segment. If the disk has been shut down, much more time is spent waiting for the disk to wake up, spin up to speed and ready itself to accept the checkpoint data.

[umfaddi]

That's a perfect explanation and I think that causes the GPU to stagnate.

Sorry, I didn't properly worded it out. I meant that turning of the disk causes the GPU to down clock indirectly:

I am imagining a scenario in which F@h tries to write the information to the disk, but because the disk is turned off, it takes time to write it, time which is probably long enough that the GPU activity is stopped thus the card down clocks.