TPU's F@H Team

[newtekie1]

GT1030 forget about that one. I'm not even sure it will run.

I wouldn't say that. It'll run, and should give a few tens of thousands of PPD, so it's worth it.

 

[mstenholm]

Core 22 is out in beta. Config, Expert, Client-type beta. It utilize the GPU a lot better than Core 21(95% vs 90%) but handle with care. I tried with four different WUs on my RTX 2060 starting with my normal +40 MHz OC plus some extra fan speed. Three times it crashed ending in BAD UNIT. Then +20 MHz OC, no,. No OC, no. No OC and 80% power limit, no. But it was sweet when it lasted. My normal PPD is 900K, the 11733 (the only one that is released for Core 22 I think) was 1.06M PPD.

Well I didn't give up so I tried my two 2070s. Lesson from yesterday dictated no OC and lower power limit plus instead of auto fan fixed 60 %. The first one went well with 1.46M PPD (normal 900-1400K PPD). This is like getting a continuous row of 14167s and none of the small WU which only yield 900k.

 

[AmioriK]

I wouldn't say that. It'll run, and should give a few tens of thousands of PPD, so it's worth it.

I think I am going to give Folding At Home a try. The Graphics Cards I currently have available are;

• RX 570 8G
• GT 1030 2G GDDR5
• GT 730 1G DDR3
• GT 710 1G DDR3

I also have an Athlon 200GE with an Integrated Radeon Vega 3 graphics processor. But I believe it is likely not going to contribute a worthwhile amount, with it only having 192 stream processors operating at 1000 MHz. I think the same can be said of the two GT 700 series cards. I would like to set up my GT 1030 for folding. Its small Pascal GP108 core should be able to help I think. It consumes so little power and all it does currently is idle, providing display outputs. The way I see it those CUDA cores might as well be doing some good too.

I apologise if my post is long, but I think I am thinking out loud here. Any input you can provide will be much appreciated though! I am going to highlight my main questions in bold to make it easier to read.

As I currently understand it, Folding at Home works better on Nvidia GPUs using CUDA than it does on AMD ones. Is it worth configuring my RX 570 8G to run Folding when I am not gaming on it? The board has a power limit of 135W, or is the performance per watt of the Polaris 20 GPU simply not feasible for long periods of time?

My plan is to put my GT 1030 into my Ryzen 3 1200 cruncher, and turn the CPU usage allowed by the BOINC client to 50%, so that the Folding Client can use the remaining cycles to feed the GT 1030. The dual 1700 rigs will remain with the 700 series parts and entirely dedicated for WCG.

I would like to invest into new hardware for folding at home. This may ultimately lead me to turn the 1200 PC entirely to F@H, and purchase a motherboard with two mechanically 16X PCIE slots so that I may use multiple GPUs in this machine for Folding. On that note; I assume folding doesn't put a lot of strain on the PCIE interface, is this correct? So with this in mind, would a board like this one be suitable for running two GPUs for F@H? The second slot is connected via the Promontory chipset, using gen2 connectivity.

I think I may investigate the Turing-based GTX 1660 non-ti card that was recently spotted, for this PC.

Thanks for reading. I would love to hear everyone's opinions and thoughts on my plans, thank you!

Ash

 

[mstenholm]

I think I am going to give Folding At Home a try. The Graphics Cards I currently have available are;

• RX 570 8G
• GT 1030 2G GDDR5
• GT 730 1G DDR3
• GT 710 1G DDR3

I also have an Athlon 200GE with an Integrated Radeon Vega 3 graphics processor. But I believe it is likely not going to contribute a worthwhile amount, with it only having 192 stream processors operating at 1000 MHz. I think the same can be said of the two GT 700 series cards. I would like to set up my GT 1030 for folding. Its small Pascal GP108 core should be able to help I think. It consumes so little power and all it does currently is idle, providing display outputs. The way I see it those CUDA cores might as well be doing some good too.

I apologise if my post is long, but I think I am thinking out loud here. Any input you can provide will be much appreciated though! I am going to highlight my main questions in bold to make it easier to read.

As I currently understand it, Folding at Home works better on Nvidia GPUs using CUDA than it does on AMD ones. Is it worth configuring my RX 570 8G to run Folding when I am not gaming on it? The board has a power limit of 135W, or is the performance per watt of the Polaris 20 GPU simply not feasible for long periods of time?

My plan is to put my GT 1030 into my Ryzen 3 1200 cruncher, and turn the CPU usage allowed by the BOINC client to 50%, so that the Folding Client can use the remaining cycles to feed the GT 1030. The dual 1700 rigs will remain with the 700 series parts and entirely dedicated for WCG.

I would like to invest into new hardware for folding at home. This may ultimately lead me to turn the 1200 PC entirely to F@H, and purchase a motherboard with two mechanically 16X PCIE slots so that I may use multiple GPUs in this machine for Folding. On that note; I assume folding doesn't put a lot of strain on the PCIE interface, is this correct? So with this in mind, would a board like this one be suitable for running two GPUs for F@H? The second slot is connected via the Promontory chipset, using gen2 connectivity.

I think I may investigate the Turing-based GTX 1660 non-ti card that was recently spotted, for this PC.

Thanks for reading. I would love to hear everyone's opinions and thoughts on my plans, thank you!

Ash

To answer your last question first - yes it has nice spacing between the slots which the top GPU will love. Edit: Only the top PCI-e will run 3.0 x 16, the other will run 3.0 x 4 and that will reduce the efficiency with some (5-10) %.

The smaller Nvidia sips power so give it a go. You have a kill-a-watt or what you native English speaking people call that ampere measuring device. You decide how much energy you want to spend on it.
I guess the same goes for the 570.

The 1660 Ti is a great GPU point per day/watt wise.

I have been folding on my daily PC for close to 8 years (summers off) with the same PSU and same MB, GPUs have been upgrade but some have around three years of running with a constant OC. The only thing that I killed was some AMD GPUs.

 

[thebluebumblebee]

turn the CPU usage allowed by the BOINC client to 50%,

What we recommend is to run the WCG client at 100%, but reduce the "% of CPUs", commonly called "giving F@H a thread". I have an i7-6700K running with a GTX 980 Ti. I want to maximize F@H on that so I "give" F@H 2 threads. That means I run BOINC at 75 % of CPUs, or to say it another way, I run BOINC on 6 threads and allow the other two threads to process F@H. We don't recommend running F@H on the CPU, only the GPU - you will have to manually remove the CPU slot when you install the F@H client.

 

[AmioriK]

I've set the GT 1030 up and running, on Windows 10 Pro, along with the WCG client:

Spoiler: GT 1030

And I thought, why not try the RX 570 for a while see how it goes.

Spoiler: RX 570

I've got WCG running here too, of course, my 2700X is a big cruncher for that. I've set the WCG client to use 87.5% of the CPU threads; 7 cores, 14 threads. So that gives the FAH client a single core and both its logical threads to work with.

According to this LIST a 560 does 114k PPD and a 580 does 344k PPD so some where in between.

Seems like it's doing around 311K PPD, closer to the 580, of course as it has 32 out of the 36 CU. Power is at 100W, which is quite a lot. Near the same power as 1660 Ti but half the performance. I probably won't fold on the 570 a lot, given the perf/watt... Best wait until I can afford a 1660 Ti for that. I think I will let it complete this work unit and then stop.

 

[hat]

Not sure about the 1660Ti, but I think GTX1070 does roughly 700k PPD on average? The 1660Ti should be more...

 

[AmioriK]

I guess it depends on the workloads; the 1660 Ti has fewer CUDA cores (1536 vs 1920) but dedicated INT32 and FP16 (at double speed) pathways alongside the FP32 ones, which could speed up computing in scenarios where those types of instructions are used together. At least that is my understanding. @mstenholm commented on its performance, being around 600K PPD, but could do more. Either way it seems like a solid investment for my 24/7 F@H rig that I have now set up. Definitely not going to Fold on the 570 though.

 

[mstenholm]

Not sure about the 1660Ti, but I think GTX1070 does roughly 700k PPD on average? The 1660Ti should be more...

The GTX 1070 does 700k PPD stock on Linux, 650k on Windows with an OC. BUT the 1660ti is cheaper to buy and to run.

 

[AmioriK]

Is 320K PPD @ 100W board power pretty decent for my 570? I thought it was bad at first, well compared to Turing. But it seems fairly competitive with what I've read for the 1050 / Ti cards, which seem to be around 180-200K PPD @ 75W. I'm debating whether to keep folding on the 570. On that note, I've been eyeing up 4G 570 cards for around £110 on Amazon Warehouse, which is pretty awesome value. I'm debating whether to grab one and set it up in my 1200 PC for folding 24/7... I'm fairly certain it could take a hefty undervolt too; I'm also considering that for my current 570.

Just to clear this up; with regards to errors due to instability in F@H, are there negative repercussions for an error in a Work Unit, or does it just subtract points? This (along with the same in WCG) is one of my biggest worries.

Thanks.

 

[phill]

No issues at all as far as I know/understand for getting an error in some of the work units, I've had a few were that's happened and also when the server has taken down the unit as well.. So no need to worry there at all

My Ti cards run about 1m PPD but I never leave it on long enough to make the points, the solar only lasts a few hours a day at the moment, so I'm just trying to watch the electric bill That rig will munch through 700w or more easy with the 5960X loaded with WCG and 2 1080 Ti's on folding as well.. Nearly a 1Kw an hour is a bit ouchy when it comes to the electric bill...

 

[mstenholm]

Is 320K PPD @ 100W board power pretty decent for my 570? I thought it was bad at first, well compared to Turing. But it seems fairly competitive with what I've read for the 1050 / Ti cards, which seem to be around 180-200K PPD @ 75W. I'm debating whether to keep folding on the 570. On that note, I've been eyeing up 4G 570 cards for around £110 on Amazon Warehouse, which is pretty awesome value. I'm debating whether to grab one and set it up in my 1200 PC for folding 24/7... I'm fairly certain it could take a hefty undervolt too; I'm also considering that for my current 570.

Just to clear this up; with regards to errors due to instability in F@H, are there negative repercussions for an error in a Work Unit, or does it just subtract points? This (along with the same in WCG) is one of my biggest worries.

Thanks.

If a GPU get many failed WUs it will be put on standby for days. If the fail is due to a really bad WU you get some point depending of how far you got, but no bonus points. Normally an non-OC GPU will not fail a WU, exception are beta units. There are no negative points but you slow down the science since the next WU are based on your WU.

Yes 320k PPD is decent for 100 W. Linux@1070 is 135 W and 700 k, a GTX 970 is close to 400k PPD but at 170-180 W. A RTX 2070 is 180-200 for 1-1.5Mk PPD. The bigger the GUP the better PPD/W due the the progressive bonus.

 

[thebluebumblebee]

What are the qualifications for the QRB?

The bonus is applied for users who use a passkey, have successfully returned at least 10 bonus-eligible WUs, have successfully returned 80% or more of assigned WUs, and returned the unit before its Timeout (formerly Preferred Deadline). Bonus points do not apply to partial returns.

Just to clear this up; with regards to errors due to instability in F@H, are there negative repercussions for an error in a Work Unit, or does it just subtract points? This (along with the same in WCG) is one of my biggest worries.

As you can see, you can lose the QRB if you have a failure rate higher than 20%.

Is 320K PPD @ 100W board power pretty decent for my 570

Appears to be https://docs.google.com/spreadsheets/d/1vcVoSVtamcoGj5sFfvKF_XlvuviWWveJIg_iZ8U2bf0/pub?gid=0#

Edit: This is where most of us check on our F@H stats: https://folding.extremeoverclocking.com/user_list.php?s=&srt=1&t=50711 Secondly, the points shown there are without the QRB, so they are much lower than what the client is saying. Get 10 WU's done with a passkey and you'll start getting the bonus. EDIT: Okay, that came out wrong. When you look at your points on EOC, you will see only base credit at this time until you've met your 10 WU's / 80% threshold.

 

[AmioriK]

Many thanks for the advice everyone it is highly appreciated. I think I will run my GPUs entirely at stock to hopefully eliminate any chance of a failed WU being on my end (Same goes for my CPUs for WCG). This makes me feel better as the anxiety of the possibility of me failing a WU/Task is too high for me.

Appears to be https://docs.google.com/spreadsheets/d/1vcVoSVtamcoGj5sFfvKF_XlvuviWWveJIg_iZ8U2bf0/pub?gid=0#

Oh wow, it seems Vega 20 is quite powerful at Folding. I wonder how well it does with PPD/Watt, especially if voltage tuned, due to the 7nm process. Anyway I'm fairly set on picking up a cheap RX 570 4GB on Friday to add to my 24/7 folding machine, and then I will fold on my 8G 570 in my main PC, of course, when I am not using it to play video games. Our power bill has gone up this month, but not by an enormous amount. I think I worked out we will be paying an extra £12 a month for the set up, as is. Of course the two RX 570s will likely bump that to about 20 quid, but I'm more than happy to foot that.
Here's to a future of happy crunching

Ash

 

[thebluebumblebee]

Well, @mstenholm , since F@H on Linux runs advanced, I have this on my GTX 980. Almost 500K on a 980! (EDIT: normally gets 400K) Have to wonder what my 980 Ti is doing. Of course, the weather's really heating up and we expect to be around 70°F over the next few days.

 

[AmioriK]

Well, @mstenholm , since F@H on Linux runs advanced, I have this on my GTX 980. Almost 500K on a 980! Have to wonder what my 980 Ti is doing. Of course, the weather's really heating up and we expect to be around 70°F over the next few days.

Why does Windows suck so much?

 

[mstenholm]

Well, @mstenholm , since F@H on Linux runs advanced, I have this on my GTX 980. Almost 500K on a 980! Have to wonder what my 980 Ti is doing. Of course, the weather's really heating up and we expect to be around 70°F over the next few days.

Sweet.
I tried with my 1070/Linux and got an error before it even started. Something cuda-index. Forgot about it, maybe I shoud try again. A lot have problems with Core 22 myself included on one 2060 so it was nice that it worked out of the box for you. I didn't expect that big a performance uplift on Linux considering the GPU core is running 99% already.

 

[thebluebumblebee]

Why does Windows suck so much?

It was my understanding that there is not that much of a difference between Linux and Windows for F@H especially since it's easier to OC with Windows.
Second answer: Windows has to be ready to do a wide variety of things, a jack of all trades if you will, while Linux is barebones out of the box and you have to prep it for what you want it to do. Something that's specifically designed to do one thing is always better than something that's multipurpose. Hope it has a screwdriver.

Sweet.
I tried with my 1070/Linux and got an error before it even started. Something cuda-index. Forgot about it, maybe I shoud try again. A lot have problems with Core 22 myself included on one 2060 so it was nice that it worked out of the box for you. I didn't expect that big a performance uplift on Linux considering the GPU core is running 99% already.

That's my oldest Linux F@H box, which most likely has my worst install on it, but it just works, so I'm not touching it.

 

[AmioriK]

I wish some really rich person would come along and break out an NVIDIA DGX-2 (or more ) and run the entire thing for Folding @ Home for Team TPU

Anyone have any estimates on the potential PPD of that thing?

 

[XZero450]

You might get ~30% more than a 2080ti * 16. That's if F@H, and all of the supporting API's, properly supports the cards inside.

 

[mstenholm]

I totally f**ked up my F@H install on my "new" W3670 cruncher so I did the next best thing - added a GPU on my 2700x cruncher/folder

Truth should be told that 14167 is the best yielding WU outside of beta on this GPU. On Windows (2.0x16, X58) I get up-to 1M with a slight OC. This is stock.

Edit: Based on only two WUs I saw that Linux is 6% faster than Win7, even with lower clock. OK that is at the expense of slightly higher consumption since the GPU is 99% busy and 88 % in Window.

 

[phill]

Very awesome mate I've never seen two completely different cards in a system before!!

 

[mstenholm]

Very awesome mate I've never seen two completely different cards in a system before!!

It took a driver update, last time I tried I forgot and got 200k instead. I'm surprised that it ran at all, maybe it didn't. I shut it of when I saw and pulled the card.

 

[phill]

It looked good which is the important bit

 

[thebluebumblebee]

I'm away on an extended weekend, so I had to shut F@H down. Left a couple of WCG boxes running, but I couldn't risk the heat leaving all of the systems running.