Folding Community Board

[TVwazhere]

I just remembered I had a power outage a week ago and uh-

[LAR_Systems]

On 1/28/2023 at 10:16 PM, Mxyzptlk said:

@LAR_Systems I am curious with the PPD Estimates for the 4070ti. They seem.... High. I am seeing my 4070ti in the 11 - 12 range, but almost every WU has a crazy high spike that might be effecting the numbers? 

It's not the spikes as we take multiple samples across work units.

What I commonly see with new GPUs is that F@H will toss work units that can be too easy or more less not optimized for the new GPU and it results is crazy results for the first few weeks.

 

Then things seem to settle (average) down in almost all cases which suggests they benchmark then put the GPUs to work on the kind of work where they can get the most out of the hardware.

 

The chart below showing the first two weeks vs this week so far suggest this is the case and it's normalizing down as they appear to push far more Influenza and pull back on Alzheimer's / remove membrane transport projects.

 

This kind of thing is why I show the chart and I felt the PPD DB was important enough to build because "the ppd a GPU gets" changes a lot over time with F@H, Drivers, OS etc. evolving what they are doing.

 

[Bitter]

Has the Python 2 issue been sorted out with FAH control yet? I'd like to run Linux but don't want to screw with installing Python 2 as I recall it was kind of a pain and caused some other unfortunate problems due to version confusion.

[Mxyzptlk]

12 hours ago, LAR_Systems said:

It's not the spikes as we take multiple samples across work units.

What I commonly see with new GPUs is that F@H will toss work units that can be too easy or more less not optimized for the new GPU and it results is crazy results for the first few weeks.

 

Then things seem to settle (average) down in almost all cases which suggests they benchmark then put the GPUs to work on the kind of work where they can get the most out of the hardware.

 

The chart below showing the first two weeks vs this week so far suggest this is the case and it's normalizing down as they appear to push far more Influenza and pull back on Alzheimer's / remove membrane transport projects.

 

This kind of thing is why I show the chart and I felt the PPD DB was important enough to build because "the ppd a GPU gets" changes a lot over time with F@H, Drivers, OS etc. evolving what they are doing.

 

Thanks for the explanation. I have seen several new GPU's pop onto the list over time, just never saw this happen this dramatically before.  This weeks numbers look a lot more in line with what I am seeing.

[Bitter]

It's really cold, take my extra 250W worth of points until I get sick of the fan noise.

Edit: fan noise due to card I forgot to clean and repaste running at 700RPM higher fan speed and 15C hotter temp. Whoops. I'll fix that tomorrow and let it finish the current work. ~2M PPD at 250W from the wall isn't amazing these days but it is what it is with a pair of 1660S with a mild power limit.

[TVwazhere]

On 1/30/2023 at 8:31 AM, TVwazhere said:

I just remembered I had a power outage a week ago and uh-

I just remembered I forgot to turn on ITNOS and uh-

[RollinLower]

19 hours ago, TVwazhere said:

I just remembered I forgot to turn on ITNOS and uh-

very eco-friendly tho!

[NetoriusNick]

On 1/20/2023 at 6:19 PM, IkeaGnome said:

Oh that's the forums doing the broken attachment thing again. I can't remember what you have to remove from the broken link. I can also send them over Discord if you'd want.

Andrew#9623 PFP matches my PFP on here.

happen to have put off and put off looking at this only to find out that the office stash of compnenets has the exsact cap in it 

[IkeaGnome]

10 hours ago, NetoriusNick said:

happen to have put off and put off looking at this only to find out that the office stash of compnenets has the exsact cap in it 

That makes it a whole lot easier to find! 

[TVwazhere]

On 2/2/2023 at 4:11 AM, RollinLower said:

very eco-friendly tho!

I remembered the thing

If only it kept going in that direction

[GOTSpectrum]

Just now, TVwazhere said:

I remembered the thing

If only it kept going in that direction

[Bitter]

I'm running 2x 1660S with a conservative overclock (naughty I know) in a H97 board with 8GB RAM and an i5 4590T...on Windows... I'm seeing 80-90% CPU use while both cards are running, PPD seems inline with other 1660S. Should I swap in a CPU with a little more processing power or as long as PPD is on target don't worry about it?

[Gorgon]

48 minutes ago, Bitter said:

I'm running 2x 1660S with a conservative overclock (naughty I know) in a H97 board with 8GB RAM and an i5 4590T...on Windows... I'm seeing 80-90% CPU use while both cards are running, PPD seems inline with other 1660S. Should I swap in a CPU with a little more processing power or as long as PPD is on target don't worry about it?

So a low power 4c/4t processor with 2.0 GHz Base and 3.0GHz Boost frequencies. It would likely under perform in Windows with a few of the more tasking WUs but those GPUs would be less likely to get a lot of those.

 

Your likely fine. I ran 2070s with Pentium Gold 5400 ad 5500s that weren't that much more powerful but did have much higher Base clocks for years with no issues but that was under Linux.

 

I suspect the $50 you'd likely spend on a new CPU might be better put towards a more modern platform than Haswell.

 

A better approach might be to lower the GPU power limits to the minimum to conserve electricity and run them more efficiently unless you need the energy used for heating.

[Bitter]

55 minutes ago, Gorgon said:

So a low power 4c/4t processor with 2.0 GHz Base and 3.0GHz Boost frequencies. It would likely under perform in Windows with a few of the more tasking WUs but those GPUs would be less likely to get a lot of those.

 

Your likely fine. I ran 2070s with Pentium Gold 5400 ad 5500s that weren't that much more powerful but did have much higher Base clocks for years with no issues but that was under Linux.

 

I suspect the $50 you'd likely spend on a new CPU might be better put towards a more modern platform than Haswell.

 

A better approach might be to lower the GPU power limits to the minimum to conserve electricity and run them more efficiently unless you need the energy used for heating.

I should have mentioned, I'm at 90% power limit. Around 2Ghz on both cards with 7000Mhz memory (+125 core, +200 memory, no failed WU's I'm aware of). Seeing about 260W from the wall for the whole rig with a decent Gold rated PSU. I wanted to run Linux but the whole FAHControl Python2 stuff, bleh. I had a Windows install already on a disk with FAH still installed from 2020 so I just picked up where it left off after running Windows updates. CPU is running around 2.6GHhz on all 4 cores. I have a standard i5 4590 I could drop in there...the Haswell rig was formerly the folding rig, then the mining rig, and now back to the folding rig. It's the only Intel board I have at the moment that's somewhat modern and has 2  x16 slots on it. My AMD boards are a couple B450's and a B550. CPU's available for those are an Athlon 200GE (no), 3200G, and a 1400.

I'm not specifically going for heating but the 260W heat in this room has been slightly more pleasant. My aim is to use as little power as possible for the rest of the rig with the most power going toward folding, so the less power I spend on CPU the better.

I've got a spare SATA drive around, might pause the fold and get Linux on it then mess around with that.

I can grab a ~65W 4590S for about $15, might do that. Each GPU is running about 100W (200W total) so the system overhead is 60W from the wall. Not terrible, not great, could be better. I'd love to see about 1/2 of that but I don't think there's anything in desktop socket that's low power with enough threads? I do have an i3 9100 in a SFF Dell incoming, I might upgrade the CPU and then have that available if I can snag a cheap board for that. I'm sure it would be a little bit more efficient than 4th gen Intel?

[Gorgon]

4 hours ago, Bitter said:

I should have mentioned, I'm at 90% power limit. Around 2Ghz on both cards with 7000Mhz memory (+125 core, +200 memory, no failed WU's I'm aware of). Seeing about 260W from the wall for the whole rig with a decent Gold rated PSU. I wanted to run Linux but the whole FAHControl Python2 stuff, bleh. I had a Windows install already on a disk with FAH still installed from 2020 so I just picked up where it left off after running Windows updates. CPU is running around 2.6GHhz on all 4 cores. I have a standard i5 4590 I could drop in there...the Haswell rig was formerly the folding rig, then the mining rig, and now back to the folding rig. It's the only Intel board I have at the moment that's somewhat modern and has 2  x16 slots on it. My AMD boards are a couple B450's and a B550. CPU's available for those are an Athlon 200GE (no), 3200G, and a 1400.

I'm not specifically going for heating but the 260W heat in this room has been slightly more pleasant. My aim is to use as little power as possible for the rest of the rig with the most power going toward folding, so the less power I spend on CPU the better.

I've got a spare SATA drive around, might pause the fold and get Linux on it then mess around with that.

I can grab a ~65W 4590S for about $15, might do that. Each GPU is running about 100W (200W total) so the system overhead is 60W from the wall. Not terrible, not great, could be better. I'd love to see about 1/2 of that but I don't think there's anything in desktop socket that's low power with enough threads? I do have an i3 9100 in a SFF Dell incoming, I might upgrade the CPU and then have that available if I can snag a cheap board for that. I'm sure it would be a little bit more efficient than 4th gen Intel?

IDK if they've fixed the native FaHControl on Linux. I run 8 Linux Rigs with GPUs and I just use FAHControl from my Windows Daily Driver on them. Less monitors that way

[Bitter]

14 minutes ago, Gorgon said:

IDK if they've fixed the native FaHControl on Linux. I run 8 Linux Rigs with GPUs and I just use FAHControl from my Windows Daily Driver on them. Less monitors that way

True. I suppose I can do that easily enough! I'll monkey with that tomorrow and may end up swapping to a B450 board with the 3200G as I can somewhat cheaply get a 6C CPU for AM4 which would let me pop my GTX 1070 into the M.2 slot, right? I've never tried running a graphics card out of a NVME slot before but I wouldn't mind using the 1070 for something besides a door stop.

I've still got mining risers and a splitter card but from what I can recall, even under Linux the 1x interface is too slow for FAH to fully use the GPU, correct?

[Gorgon]

3 hours ago, Bitter said:

True. I suppose I can do that easily enough! I'll monkey with that tomorrow and may end up swapping to a B450 board with the 3200G as I can somewhat cheaply get a 6C CPU for AM4 which would let me pop my GTX 1070 into the M.2 slot, right? I've never tried running a graphics card out of a NVME slot before but I wouldn't mind using the 1070 for something besides a door stop.

I've still got mining risers and a splitter card but from what I can recall, even under Linux the 1x interface is too slow for FAH to fully use the GPU, correct?

Try the 1x riser first. A 1070 is a comparatively small card by today's standards. A m.2 riser will work but you might be surprised by how little of difference you will see.

 

Under Linux the 3200G might not be enough for 3 GPUs but the 1400 with 8 threads should be able to handle it.

[Bitter]

24 minutes ago, Gorgon said:

Try the 1x riser first. A 1070 is a comparatively small card by today's standards. A m.2 riser will work but you might be surprised by how little of difference you will see.

 

Under Linux the 3200G might not be enough for 3 GPUs but the 1400 with 8 threads should be able to handle it.

I'll try the 1070 on 1X riser tomorrow with FAH bench and then into a full 16X slot on a spare test bench, thanks! I'll also see how my P106-100 does on the riser as well. I may end up with a 4 card system again which for sure will need more than a 4 core CPU!

I was under the impression that Nvidia GPU's need whole cores something something Nvidia drivers kind of suck lol, and that SMT didn't count? I did score a H370 board fairly cheap (2x full length and several 1X...and a PCI slot for some weird reason) so when I can snag an 8500T or 9500T i5 cheaply enough I'll have a low watt 6C6T system to swap over to from Haswell which is limited to 4C8T maximum. I'll get it running on Linux then. I took a closer look at the LAR database and my 1660's on Windows seem to be averaging around what they do in Linux. Even with the weak CPU somehow I think I managed to get everything just barely right and scraping by, likely dumb luck and the overclock are able to overcome the deficient OS I'm currently using. The 1660S's I have are the same model but not identical, one has Samsung and the other Hynix memory. For mining they needed different settings but for folding running them both the same seems to work out just fine. I'm not going to bother folding with my RX 580 this go around, it's not efficient enough.

[Gorgon]

@Mxyzptlk Congratulations on breaking 20 Billion! That's 3 of us in Team LTT now in the roaring 20s with @Justaphf set to break 25 Billion soon.

[Bitter]

@Gorgon

Ran the 1070 in my B550/3200G test bench on FAH Bench, 1st on the 1X  and then on the 16X, same slot, same settings. 1X gave a score of 104 over a 5 minute test and 16X gave a score of 109 over a 5 minute test. For laughs the Vega8 did 22.73. I think the M.2 to 16X adapter would be ideal, they're not very expensive. I did score a H370 board cheaply, have RAM, just need a CPU with 6 cores for it then I can rebuild the folding rig with 3 or 4 GPU's. I might try testing the P106-100 but I feel it's just not efficient enough to be worth the power use.

Fired up the 1070 and 1650S to fold, they're sort of jammed together in the mATX board so I set the 1070 on a about 110W power limit and the 1650S on a 90% target both with a mild overclock. They'll do one WU each and I'll decide if I'm going to keep using them or not. I'm pretty sure the 1650S can run on a 1X riser with no detriment given it's rather low compute power.

[rkv_2401]

Interesting research paper on the increasing compute used by machine learning models over the years.

https://arxiv.org/pdf/2202.05924.pdf

[Mxyzptlk]

On 2/5/2023 at 11:26 AM, Gorgon said:

@Mxyzptlk Congratulations on breaking 20 Billion! That's 3 of us in Team LTT now in the roaring 20s with @Justaphf set to break 25 Billion soon.

Thanks!!

 

Soon to be 21 Billion

[Bitter]

Got a H370 and i5 8400T inbound to build the new folding rig. Same low power use, 2 more cores to work with. Being that it's the bottom barrel i5 8400T it might not even use the whole 35W running FAH.

[Bianks84]

Well guys, basement temps are still great, but had to deal with minor flooding this morning

[Bitter]

20 minutes ago, Bianks84 said:

Well guys, basement temps are still great, but had to deal with minor flooding this morning

Wrong way to do water cooling bud.