GPU Folding Facts and Fiction - Configuration and Settings

[Whaler_99]

With the new Cores and the release of new GPUs, there is a lot of information floating around from various sources. Not, if any, is official unfortunately. So, in this thread I am going to post some info I have found to be pretty trustworthy as well as results I know from my own tests.

 

I invite other users to post their findings. NOTE!! Do not simply post PPD results and the like. This thread is for posting config examples that you know to work based on your testing. Yes, testing, meaning you have multiple systems, GPUs, whatever, to back up your claims. Or you have read, post links if possible, trusted sources.

 

This will hopefully allow you, the average user, who might be considering some GPU upgrades, to get some information that may pertain to your environment. Also, NOTE, all this info is very subjective to the testers environment and settings. Mileage will always vary. Most of this directly effects multi GPU configs. 

 

General Information

Unfortunately most of the info you see on the Stanford site is out of date. Much of the information is from 2013 when the new website was launched and has not been updated much. So, when reading things there, take it with a grain of salt. There have been a lot of upgrades and changes over the past two years.

 

Supported GPUs

There is a file, called the GPU White-list, that contains a listing of all supported GPUs. This does NOT mean that because your GPU is in there, it is supported. This list is updated with new cards but never has old cards removed.

 

As per the Folding Forum, the somewhat official information source, here are the currently supported cards:

 

Officially Supported GPUs
AMD -> HD 5000 Series or higher
Nvidia -> GTX 400 Series (Fermi) or higher

 

That was from 2013. I think the minimum's supported have now gone up, but I am not positive. 

 

Also the equivalent Quadro, FireGL and such cards are also supported.

 

The dedicated GPU compute cards, NVidia Tesla, Intel Phy and others, are not supported.

 

PCIe Slots

You can use PCIe x1 size through x16 sized slots for folding with GPUs. There are various types of expanders that allow your x16 card to plug into a x1, for instance, slot. Also, v1 through v3 of the PCI Express specifications is also supported. Now, here is where is gets complicated. Different Cores and Projects need differing amounts of bandwidth. 

 

In my testing, and some others, about two years ago, we found no PPD decrease when using x1 slots versus x16 slots (PCIe v2). Back then though Core17 was just coming out and there was no Core18. This has changed somewhat. If you get Core15 and some Core17 assignments, this still won't make a difference. But, on Core18 it can. Some have seen up to a 20% drop on PDD when going from x16 to x4 slots. At this point though, v2 or v3 slots don't seem to make much of an impact. So, where does this leave you?

 

If planning to make a multi GPU folding system, try and get/use as many x16 slots as possible. You can use x4 and x1 with the use of expanders, but depending on the work you get assigned, you might notice some PPD decline now on those cards.

 

CPU Impact

This is primarily a NVidia issue, mostly attributed to NVidia drivers. So, the rule of thumb when folding with a NVidia GPU and CPU folding, is to leave one core free per GPU to maximize your GPU Folding performance. In the case of AMD, it is simply recommended to leave a single core free for one or more cards as the impact isn't as great.

 

So, what is going on? Well, it turns out when Folding on NVidia there seems to be a lot of polling. If you watch the CPU, yes, you can see it pegged pretty high, but if you look at something like the heat output, it actually isn't generating much. The CPU isn't actually working that hard, just being consumed with polling queries. Apparently not having the core free can greatly increase processing time, which in turn affects the QRB bonus. But, the CPU isn't actually doing a whole heck of a lot. The GPU is still doing all the work, just there seems to be this polling mechanism for some reason. 

 

SLI and Crossfire

If you have two cards with either of these enabled, no fear. There is minimal impact on Folding from having these technologies set.

 

---

 

Hopefully some of you find this useful and as mentioned, feel free to add your findings. 

 

Just remember, I will remove posts not on topic and specifically containing config info.

 

---

 

Update October 2015

 

Drivers

As many have seen, sometimes updating to the latest version, especially when folding, is not always the best thing. One of the more current versions for NVidia is simply crushing folding. If you are on a version and everything is running tip top, pay close attention to what those new updated version are doing. 98% of the time, you don't need them as they have zero performance gain for most things. What they are doing is adding specific support for games, fixes, etc. If you don't play those games, then you don't need the update.

Just pay attention before updating drivers and be prepared to roll back if things don't go something and wait for the next round.

 

Minimum Support Cards

Officially Supported GPUs
AMD -> HD 5000 Series or higher
Nvidia -> GTX 400 Series (Fermi) or higher

 

Looks like they has been tweaked a bit, nothing official, but looks like these are now the minimum needed to get most WU's completed:

AMD -> 6000 series or higher

NVidia -> GTX 620 or higher

Edited October 24, 2015 by Whaler_99
some updated information

[MrLomax]

 But, the CPU isn't actually doing a whole heck of a lot. The GPU is still doing all the work, just there seems to be this polling mechanism for some reason. 

 

U sure?

[Whaler_99]

U sure?

 

Yes, both myself and some other sources have seen that this is what seems to be going on. I fold with two 780's and no CPU folding, and my CPU is pegged pretty high  most of the time, yet hovers at like 40C (I am also doing other things). If it was actually doing any Folding work, the temps would be a lot higher, considering the GPU's are dumping a massive amount of heat into my loop. Same has been reported by some senior guys at some other sites.

[AMICLG]

I know for sure that as far back as the HD 5770 works in folding, I had two running a month ago on an old rig in crossfire. Without CPU folding enabled the CPU usage on a Q6600 was around 40%. I can get pics and info if needed, they just won't be in Xfire but rather in separate computers now.

[MrLomax]

Yes, both myself and some other sources have seen that this is what seems to be going on. I fold with two 780's and no CPU folding, and my CPU is pegged pretty high  most of the time, yet hovers at like 40C (I am also doing other things). If it was actually doing any Folding work, the temps would be a lot higher, considering the GPU's are dumping a massive amount of heat into my loop. Same has been reported by some senior guys at some other sites.

i fold with a 980 a 580 and an i7 5930k wAtercoled and "ocED" to 4,5GHz 37C on iddle and 65 when folding, as well as the core and cpu usage shows perf that cpu is working beastly.

That´s why I asked. Thanks for your reply

[iHardware Shelden]

I tested a GT 210, 8600 GTS and 9800 GTX to see if they worked, no luck, looks like they did update the support list oh well...

[Imakuni]

Yes, both myself and some other sources have seen that this is what seems to be going on. I fold with two 780's and no CPU folding, and my CPU is pegged pretty high  most of the time, yet hovers at like 40C (I am also doing other things). If it was actually doing any Folding work, the temps would be a lot higher, considering the GPU's are dumping a massive amount of heat into my loop. Same has been reported by some senior guys at some other sites.

So.... in theory, we could just get a HT CPU and bypass this limitation? We could dedicate the HT cores for the GPUs, while using the non-HT cores for CPU folding, with little performance drop. I would love to test that, though I lack hardware to do so.

 

Also, as far as old GPUs go, I have a Radeon Hd 6670. It can get and complete tasks, but just barely within the deadlines on a 24/7 scheme, so I don't recommend using it or anything slower.

[Whaler_99]

So.... in theory, we could just get a HT CPU and bypass this limitation? We could dedicate the HT cores for the GPUs, while using the non-HT cores for CPU folding, with little performance drop. I would love to test that, though I lack hardware to do so.

 

Also, as far as old GPUs go, I have a Radeon Hd 6670. It can get and complete tasks, but just barely within the deadlines on a 24/7 scheme, so I don't recommend using it or anything slower.

 

Yes, I believe 6000 series is the minimum now.

 

The trick with Folding, is when you tell it to use say 6 of 8 or 4 of 8, or whatever, it still spreads the load across all cores in many cases, it just has the load that you would normally find if those were dedicated it seems. So there has been talk previously about setting affinity and such, but there has never really seemed to work either. In the end, your idea has merit, but might be night on impossible to do.

[Imakuni]

Yes, I believe 6000 series is the minimum now.

 

The trick with Folding, is when you tell it to use say 6 of 8 or 4 of 8, or whatever, it still spreads the load across all cores in many cases, it just has the load that you would normally find if those were dedicated it seems. So there has been talk previously about setting affinity and such, but there has never really seemed to work either. In the end, your idea has merit, but might be night on impossible to do.

Huh. So I guess Process Lasso won't be of any help.....

 

Oh well, that's life, I guess.

[AlexAlcala]

CPU Impact

This is primarily a NVidia issue, mostly attributed to NVidia drivers. So, the rule of thumb when folding with a NVidia GPU and CPU folding, is to leave one core free per GPU to maximize your GPU Folding performance. In the case of AMD, it is simply recommended to leave a single core free for one or more cards as the impact isn't as great.

So, what is going on? Well, it turns out when Folding on NVidia there seems to be a lot of polling. If you watch the CPU, yes, you can see it pegged pretty high, but if you look at something like the heat output, it actually isn't generating much. The CPU isn't actually working that hard, just being consumed with polling queries. Apparently not having the core free can greatly increase processing time, which in turn affects the QRB bonus. But, the CPU isn't actually doing a whole heck of a lot. The GPU is still doing all the work, just there seems to be this polling mechanism for some reason.

wait!, i have a i7 4790k and an asus 970, i have the CPU at 7 threads, you are saying that if a lower the threads the GPU is going to work faster? Should i leave a full core or just one thread? noob here ^^

[Whaler_99]

wait!, i have a i7 4790k and an asus 970, i have the CPU at 7 threads, you are saying that if a lower the threads the GPU is going to work faster? Should i leave a full core or just one thread? noob here ^^

 

You just need the one core free for one 970, so you are fine.

[AlexAlcala]

You just need the one core free for one 970, so you are fine.

Thanks for the reply!

But im still confused, the 7 threads are for the 8 threads on the core(since the counting go 0,1,2...7) so i should put the cpu on folding at 5 threads? Or at 6? I heard even numbers on the cpu dont work well on folding.

Sorry to bother, i just want to do the best job i can

[Judahnator]

Thanks for the reply!

But im still confused, the 7 threads are for the 8 threads on the core(since the counting go 0,1,2...7) so i should put the cpu on folding at 5 threads? Or at 6? I heard even numbers on the cpu dont work well on folding.

Sorry to bother, i just want to do the best job i can

 

Not exactly. The cores are numbered 0-7, however in the folding client you set 7 cores because you want cores 0-6 folding. This will leave core 7 free to assist the GPU.

 

Does that help?

[AlexAlcala]

Not exactly. The cores are numbered 0-7, however in the folding client you set 7 cores because you want cores 0-6 folding. This will leave core 7 free to assist the GPU.

 

Does that help?

 

i think, but its the same if i let it on -1. So i guess as long it show 7 cores on the F@H client im ok

[Whaler_99]

Numbering on cores is weird from program to program and even within the same one, 0-7, 1-8... Within the advanced View when configuring the number of cores to use, you are not picking "what" cores but just the number to use.... 1, 3, 4, 6, whatever... FAH actually spreads the load amongts all of them regardless, simulating the load for say 5 if you choose 5 as the number of cores to use.

 

In some rare cases, CPU folding would have issues if using a ODD number of cores for some specific WU's, but that issue doesn't come up much any more.

[astranger200]

The gt 610 does fold but it is pitifully slow.

[Silverer]

48 cores, 810MHz, i feel sick looking at those specs

Hey, the laptop version manages to play CSGO at medium settings 60fps and Black Ops 2 at 40fps minimum (though I feel the cpu let it down a bit in BO2).

Pretty good for an entry/casual level gaming pc...

 

Edit: Whatddya know, the 610m and the 610 are exactly the same chip http://gpuboss.com/gpus/GeForce-610M-vs-GeForce-610

[Ryanh19]

nice

[astranger200]

my gainward 610 has horrific temps, horrific performance, just not a good card for me. Averaged 10FPS in FSX

well i find its temps good even when overclocked (68 degrees at the max) and its performance reasoneble. war thunder on medium 720p I get 60 fps

[SirSquid]

I tested a GT 210, 8600 GTS and 9800 GTX to see if they worked, no luck, looks like they did update the support list oh well...

iv also tried a gts 8800 512 and a nvidia grid k2 and they did NOT work.

[Christopher E.R.]

Hey guys, is it possible for a mining style rig, 5 GPUs or more to crunch folding while utilizing all cards at 100%? just curious.

[Extingquish]

Hey guys, is it possible for a mining style rig, 5 GPUs or more to crunch folding while utilizing all cards at 100%? just curious.

You can have multiple cards in a system, although with the newer work units, the GPUs work together with the CPU, so you'd need to make sure you had enough bandwidth for the communications and a powerful enough CPU

 

Mining would happen on just the GPUs, and there wouldn't be much communication between the CPU and GPUs, hence why you would have seen mining rigs with 6 GPUs along with something like a G3220 

 

Hopefully that made sense

[Christopher E.R.]

You can have multiple cards in a system, although with the newer work units, the GPUs work together with the CPU, so you'd need to make sure you had enough bandwidth for the communications and a powerful enough CPU

 

Mining would happen on just the GPUs, and there wouldn't be much communication between the CPU and GPUs, hence why you would have seen mining rigs with 6 GPUs along with something like a G3220 

 

Hopefully that made sense

It did clarify a whole lot   Thank you for putting things into perspective. I guess the only limiting factor would be bandwidth as adapting a mining rig using 1x to 16x risers would have seriously nerfed bandwidth with a close to or completely saturated 1x connection <_< . I would be stuck with 4 GPUs on a board that supports 4 way crossfire/ sli 

[Whaler_99]

It did clarify a whole lot   Thank you for putting things into perspective. I guess the only limiting factor would be bandwidth as adapting a mining rig using 1x to 16x risers would have seriously nerfed bandwidth with a close to or completely saturated 1x connection <_< . I would be stuck with 4 GPUs on a board that supports 4 way crossfire/ sli 

 

Check my sig for build logs and info on this exact type of build. Taken into perspective, if the option is to use a 1x to 16x expander to get extra GPU's folding, due it. Yes, it may effect the PPD somewhat, but not on a massive scale.

 

What brand of GPU's?

[CubiX]

Hello there just started folding at milkyway@home it officially supports a gtx260 and above and i am using a gtx 285  

I plan in contributing for about 16 hours per day until my gpu dies and the weekend will be full 24 hrs  

The gpu runs stable at 75 degrees buut there is a small problem some memory chips arent working correctly and so i have underclocked them anyways i hope i help enough the team even with this little computational power