Folding Forum

Just a quick FYI for folks who might be doing a search on these particular WUs, especially for those with multi-GPU rigs....

We've had reports of these WUs drawing 60W more than we're used to on an Nvidia 580 card so if you're already on the edge for power consumption these might put you over.

I've personally had to bump up my default core voltage on a "stock" MSI 560ti Twin Frozr OC card as well as peg the fan speed to 100% in order to keep folding two sli'd cards, and even with the 100% fan speed my hotter card is hitting around 85c when the norm was around 75c w/ 75%.

That said.... I'm reporting that these WUs will return 18K+ at "stock" speeds vs. the 13K+ I'm used to with the cards OC'd to 950.

I think I should ask since I didn't see the information posted : what is the GPU temperature when the system turns off ?

The OP said it shut down as soon as it started, so I don't know if he was watching the temps.

The PSU model came out some years ago. Of course, he could have bought it more recently.

If he has been folding or other distributed computing for 4 or 5 years, the PSU might just have reached its EOL.

By lowering the frequency of the gpu from 732 to 405 Mhz, it works, but then the machine becomes very slow. If I increase the frequency the noise of the ventilator is too strong. I am going to abandon Folding and to wait for Geforce Kepler, my Geforce 570 is too little power for Wu 8031.

En baissant la fréquence du gpu de 732 à 405 Mhz, cela fonctionne, mais alors la machine devient très lente. Si j'augmente la fréquence le bruit du ventilateur est trop fort. Je vais abandonner Folding et attendre les Geforce Kepler, ma Geforce 570 est trop peu puissance pour les Wu 8031.

I am going to abandon Folding and to wait for Geforce Kepler, my Geforce 570 is too little power for Wu 8031.

I am not sure that is a solution.
Newer cards will be ultimately stressed in a similar fashion. If they are not, it just means that they are not being used efficiently(to their full potential).

I have a AMD 7970. It is running very cool folding now. However, I estimated that it is currently only using about 1/3 to 1/4 of its ultimate compute capability. When the software/drivers/Folding cores are optimized, I am sure it will howl like a leaf blower.

Just the nature of GPU compute.

see this post for a good discussion of the cooling design of cards in comparison to their ultimate compute ability.
viewtopic.php?f=59&t=20878&start=15#p208836

I return to Boinc. Thanks to all for your answers.

I have no work, i have no budget to change material.

If you are stinking rich, you can donate me.

If you are stinking rich, you can donate me.

No just a hardware enthusiast. = the kind of person likely to answer a thread like this.

My GTX 570 is is stock clock and doing the same thing. It's not as bad if I only fold gpu or smp but with these hardcore wu's and both clients it overheats and trashes my system. I am using no flags to get these hot wu's so they must have gone mainstream.

Is there a way to get the tamer wu's that run cooler with this card? I'd prefer to run the gpu at 85% if that were possible but it doesn't look like that is a priority for F@h. Until then, I have to agree with Zarck except I will just remove the gpu client and leave the smp client folding.

Welcome to foldingforum.org, csimon.

Yes, they've gone mainstream.

No, there isn't any way to choose specific projects.

There are plans to fix the GPU cores so that you can set them to 85% or some other value of your choice, but that's still a future consideration. Until that capability is released, there's not much else you can do. It does take a well-designed GPU and a case that's thermally well-designed to handle the maximum heat from both CPU and GPU. Most of us assume that adding a powerful GPU to an existing system is just a matter of having an available slot, a powersupply that's big enough, and some extra drivers/software. The fact that the air circulation wasn't designed for it isn't something we think about, and we should. A reputable computer manufacturer would have designed and tested the heat dissipation for all the options that they expect you to install but GPU manufacturers offer you anything you want and often just dump the extra heat INSIDE the case -- with you having to pay extra for a "reference design" two slot-board that exhausts most of the added heat directly out of the back of the case.

I'm all for doing the fantastic science that only GPUs can do, but is v7 installing SMP+GPU slots by default still a good idea with these perspectives in mind? I've already linked to the v7 thread where I suggested adding notes about lag, but heat is also a good point. Not everyone has my case...

This thread is all about design consistancy : it is useless to have a powerful GPU is the rest of the machine is not capable of handling it. This applies to both CPU and MB which have to be powerful enough and to have enough bandwidth to feed the GPU and to the PSU that have to be powerful enough to power everything in the machine ...

I've never been a fan of multi-rail PSUs, and I think my point has been proved by this thread ...

While this trick is hit and miss at best, lowering the CPU usage setting on the GPU slot only has helped some people in the past. Reducing CPU usage starves the GPU of data, and introduces more wait cycles where other functions, like updating the mouse position, can take place. It can help with both lag and heat. But again, it works on some, but not all, and with limited results. No one has tracked which models and driver versions it works with yet. Caveat folder.

7im wrote:While this trick is hit and miss at best, lowering the CPU usage setting on the GPU slot only has helped some people in the past. Reducing CPU usage starves the GPU of data, and introduces more wait cycles where other functions, like updating the mouse position, can take place. It can help with both lag and heat. But again, it works on some, but not all, and with limited results. No one has tracked which models and driver versions it works with yet. Caveat folder.

I never understood how this worked, this behavior ( introducing wait cycles ) is the way the fah_gpu_idle environment variable works and a client can not influence a core after it was started. So I just checked with V7, and seems cpu-usage is passed through as '-cpu x' where x is the setting you choose. And yes it works with every gpu core I can test ( which is nvidia core 11 and ati core 11 ).

There been a change in the gpu cores which now enables passing through the 'fah_gpu_iddle' flag directly ( I think I even proposed something like this at the time as quick work around until it was possible to directly set usage ).

I expect, since this behavior is in both the ati core and the nvidia core I can test, it's also in the openCL/openMM core's. It's indeed not as fine grained as setting usage directly but it's better then nothing, the long workaround was measuring utilization and adjusting wait cycles so utilization matches the chosen usage percentage, this however requires hardware monitoring something the fahcore is not capable of ( now ).

7im wrote:While this trick is hit and miss at best, lowering the CPU usage setting on the GPU slot only has helped some people in the past. Reducing CPU usage starves the GPU of data, and introduces more wait cycles where other functions, like updating the mouse position, can take place. It can help with both lag and heat. But again, it works on some, but not all, and with limited results. No one has tracked which models and driver versions it works with yet. Caveat folder.

Is this the "Percent CPU usage" in the V7 configuration that you refer to, or is it something else?

MTM, these are two different things.

fah_gpu_idle artificially inserts a consistent 1 in 10 wait cycles (or whatever your setting is), regardless of CPU speeds, or GPU power. But the time slices are much too course. The GPU cycles up and down, and still lags in the 9 out of 10 cycles.

CPU usage slider slows down how often data is fed to the GPU. This is a more constant deficiency, hitting almost all cycles, but the effect is very dependent on CPU speed and GPU power.

Think of the data stream as a fire hose full of water flowing to the GPU. In the first example, the fire hose full of water has some air introduced, like 1 in 10 liters of water is replaced by an equal volume of air. It still runs full speed 9 out of 10 cycles. In the second example, the water valve (cpu usage) that feeds the fire hose is cranked down so that the total volume of water flowing to the GPU is reduced by a steady amount. 1/10th of every liter is withheld, not 1 in 10 liters.

csimon wrote:Is this the "Percent CPU usage" in the V7 configuration that you refer to, or is it something else?

Yes.

Note: the slider will reduce CPU usage for all slots, not just one slot (unless the GPU is your only slot ). But you can set the CPU usage as a slot option for each individual slot. The fah install guide describes how to configure that local option setting.