[SOLVED] 2x 7950s dropping to 64% load every few seconds

[wazoo42]

I am running these on gentoo with cgminer 2.11.3 (I've tried 2.10.4 as well), and I noticed that the MH/s for each card initially goes up to 505/526 for the first minute or two. Over the next 3-5 minutes it settles into 480/490 MH/s. In the first period the gpu load is 98-100%, but as it declines and once it hits steady state the gpu load alternates between 64% and 98-100%. I have tried altering the intensity (8, 9, 10), the worksize (256, 512, 1024), the # of gpu threads (1, 2, 3), and the queue size (1,3,5), and the kernel (diakgcn, diablo, poclbm). All of these settings give the same behavior (though the starting and steady state hashing speeds differ slightly). I do not see this behavior on my boxes with 2x5830s, 2x5770s, or 1x7950 & 1x5770. Any thoughts as to what might be causing this or some other test I can try?

[1713572812]

1713572812

[1713572812]

1713572812

[1713572812]

1713572812

[DrG]

Two common causes that I have seen:
1) Thermal throttling.  Are the cards getting too hot?  If they're above 85C they might be throttling. When one of my 6870s went bad it did the same thing - throttled back at 106C lol
2) Something running amuck in the back of Windows (like some other program tying up resources too much).  If it's a linux rig I have no idea.  Best solution for windows is just do a clean install, not worth wasting time trying to diagnose the cause.

[-ck]

And not enough power.

[wazoo42]

The temps are 73/57 C and it is linux. They are fed by a 1000 W rosewill bronze psu. I may see if I can get my hands on a bigger one, though this model does have 4 pci connectors, so I thought it could handle it.

[DrG]

Temps probably aren't the problem then.

Rosewill as a general rule makes crappy PSUs, but their Capstone series are pretty nice.  Does the PSU put out at least 60Amps on the 12V lines?

I'm not well versed on power since I'm not an engineer, but if you launch GPU-Z and look at the sensors you can see what the 12v drops down to while both are mining.  It could be the PSU is bad.

[wazoo42]

It has 4x12 V rails with 20, 20, 30, and 30 A on each rail. What they don't say is if they put 1 pci-e per rail. They also don't specify how the regular 12 V molex connectors are distributed across the rails. I will try putting one or two pci-e 6 pin adapters on these. Sadly the it87 sensor on this motherboard does not give me much useful voltage information. I suppose my next test would be to get a 500 W psu and use it to run one of the cards independently.

[wazoo42]

I ran one of the cards off of an antec 550W with 2x12 V rails w/ 19 A per rail. Thus, I really don't know if this tells me anything new b/c I'd run into the same problem if both connectors pulled off the same rail (for the record, both with this psu and the rosewill, I tried all combinations of pci-e and 4 pin molex + pci-e adapters). The kill-a-watt showed about 150 W before mining starts, and then 370 W with one card or 585 W with both cards. It would appear that 20 A probably isn't enough. Needless to say, the end result was the same as before. I will see if I can get my hands on a psu with larger 12 V rails.

[glendall]

Hey fellow new miner.

Think I got the answer for you. Just picked up a 7950 the other day and had same thing. GPU dropped to 66% every so often.

Go into the Catalyst Control Centre software/console.  Go to the AMD Overdrive setting.  There is a power setting option. This sets the power threshold. Increase this to the max, 20% .

Basically by default these cards power down / reduce the clock if running at 100% to save power and GPU life.

After increasing this to 20% I'm maintaining  99% gpu activity at high 70's temperature.

[Vuxil]

Hey fellow new miner.

Think I got the answer for you. Just picked up a 7950 the other day and had same thing. GPU dropped to 66% every so often.

Go into the Catalyst Control Centre software/console.  Go to the AMD Overdrive setting.  There is a power setting option. This sets the power threshold. Increase this to the max, 20% .

Basically by default these cards power down / reduce the clock if running at 100% to save power and GPU life.

After increasing this to 20% I'm maintaining  99% gpu activity at high 70's temperature.

Do you know if there is a config file that can be editted to change this? My *nix box is headless

[wazoo42]

The short answer is that ati control center no longer has overdrive capability and the power settings are controlled by libatiadlxx.so. You can create a program, powercontrol, to alter the power settings using (the export command is necessary for doing this remotely)

Code:

export DISPLAY=:0 ; ./powercontrol

https://bitcointalk.org/index.php?topic=7374.895;wap2

[Sy]

Is this still the solution to go? Got the same problem with my 280x but the fix isn't working -.-

[knapol96]

me too, need to solve it but it still occurs after increasing the power by 20%

[Sy]

It was a cooling problem for me, Sapphire 280x Dual-X

Just downclocked themself if they came close to 80C

[BillGeo]

In my case it WAS poor cooling (i dont know what or how PowerTune +20 fixed).


I got a cooling fan failure (a big external one, not on the card itself) and VRAM temps skyrocketed to 110oC
so the card began to throttle so it won't burn up in smoke!


Go the fans working reliably again, all is good!!!
 

[reactor]

And not enough power.

90% of the time it is this or temperature. ^^^

The rest of the time hardware isn't getting along (system usage, memory speed and availability, rogue processes) or you're pushing the card too hard.  Back in my infancy when I used the GUI-heavy variant for mining I'd just put in settings I found on the wiki for cards and be baffled that I'd get system resets and the like.  Pay careful attention to stats from cgminer - just made a modification to frequency or voltage and have a high HW or R rate?  Tweak down and retry.  

If you have windows, try something like GPU Caps Viewer for monitoring temp/cycles.  If you have MSI, try using something like afterburner (may have name wrong, it's the MSI monitoring app) to see the usage graphs to trace the drops.  For example, one MSI card I had tweaked to the 9's but my cooling wasn't adequate so it'd throttle for temp and eventually crash the drivers.  Removed the plastic cover and did some aftermarket hack cooling and the card screams hashing now.