LTT Official Folding Month 2019!!!

[Stormseeker9]

2 minutes ago, Gorgon said:

+185 seems a might aggressive. You might have a golden sample but you might find it errors out the odd WU.  I used to run one step (13MHz for Pascal & 15MHz for Turing) below stable observed using FAHBench but found that ocassionally I’d get a WU Failure on a GPU so I just bump them all down 2 steps.

 

Most of my cards run +65 but my 2070 runs at 120 and 1070 Ti at 170 for months.

 

The other trade off is the cards at high power limits have greatly reduced efficiency and are harder to cool so I usually reduce power down to 80-90% when not competing. I’ve found that makes it a lot easier to keep cool and easier on the power bill.

I do get the odd error in a WU. That OC been stable since months on gaming - user OC scanner + manual tinkering and 185 seems to work well. 2025-55mhz is what it usually runs at under full load. 

 

So you suggest dropping core to like +160?

[shaz2sxy]

9 minutes ago, Gorgon said:

+185 seems a might aggressive. You might have a golden sample but you might find it errors out the odd WU.  I used to run one step (13MHz for Pascal & 15MHz for Turing) below stable observed using FAHBench but found that ocassionally I’d get a WU Failure on a GPU so I just bump them all down 2 steps.

 

Most of my cards run +65 but my 2070 runs at 120 and 1070 Ti at 170 for months.

 

The other trade off is the cards at high power limits have greatly reduced efficiency and are harder to cool so I usually reduce power down to 80-90% when not competing. I’ve found that makes it a lot easier to keep cool and easier on the power bill.

Oooooooohhhhh

 

Going to set the memoery OC to 0 in a couple of days and see how that plays with the PPD

[Stormseeker9]

23 minutes ago, yaboistar said:

memory OC should be set to 0 but you can disable "force p2 in cuda" to give you the full memory clock.

Why no mem OC ?

Im gonna lower my core by 1 step to 170mhz

[Stormseeker9]

Just now, yaboistar said:

memory instabilities are harder to detect. same with overclocking regular DDR. run the memory at the spec it was designed for for sure, but there's no need to push past that and potentially end up in erroring territory

my ram is running at rated speed.

 

But only for F@H? as when gaming it runs smooth ass

[GOTSpectrum]

6 minutes ago, yaboistar said:

games aren't as sensitive to erroring as FAH and other compute workloads are. for example - my 1080ti will play elite and run benchmarks all day at a certain setting but for FAH i have to turn it down or it will fail work units

Also the errors in folding will compound over tint, as each frame is a child of the one that preceded it. So if you make even just a 0.1% error every frame, you'll be massively out by the end. Rendering the entire simulation scientifically useless 

[LavenderPhantom78139]

My 1060 3gb runs at 2120 Mhz folding generally. Much better than my 1050tis 2012Mhz

[TVwazhere]

On 9/10/2019 at 8:24 AM, TVwazhere said:

Black, 1070ti x2, 280w

Spoiler

 

I decided to let ITNOS run the 1070ti's at 60% power instead of 50% to see what the difference would be.... (also they're now forced to p0 states but I'm not sure if that's doing anything)

 

1070ti's 60% Power, roughly 320w from the wall (same as 1070ti, 970, 750ti)

 

[Gorgon]

14 minutes ago, TVwazhere said:

I decided to let ITNOS run the 1070ti's at 60% power instead of 50% to see what the difference would be.... (also they're now forced to p0 states but I'm not sure if that's doing anything)

 

1070ti's 60% Power, roughly 320w from the wall (same as 1070ti, 970, 750ti)

 

Yes, the lower the power is limited, the more you can push the overclock, but there are Twp other factors, too high of an overclock can lead to an immediate failure when the WU returns from the CPU from a checkpoint or calculation as the chip has cooled down so it tries to boost up to 2200MHz or something equally unstable.

 

The second factor is the Quiescent load from the system is a fairly constant 35 to 70W depending on your non-GPU Hardware and this constant leads to an overall larger decrease in overall system (as opposed to just GPU) efficiency with lower GPU power draws. The efficiency (Work/time or PPD/W) should be read at the wall including the Power supply losses and what I’ve found through testing is that, though the decrease in the QRB due to returning WUs not as fast at lower Power Limits is not as significant as the increase in GPU efficiency, factoring in the System overhead means that the lowest power is usually NOT the most overall efficient but the peak in efficiency is usually somewhere around 70-90% of the GPU’s FE Power Limit.

[shaz2sxy]

6 minutes ago, Gorgon said:

Yes, the lower the power is limited, the more you can push the overclock, but there are Twp other factors, too high of an overclock can lead to an immediate failure when the WU returns from the CPU from a checkpoint or calculation as the chip has cooled down so it tries to boost up to 2200MHz or something equally unstable.

 

The second factor is the Quiescent load from the system is a fairly constant 35 to 70W depending on your non-GPU Hardware and this constant leads to an overall larger decrease in overall system (as opposed to just GPU) efficiency with lower GPU power draws. The efficiency (Work/time or PPD/W) should be read at the wall including the Power supply losses and what I’ve found through testing is that, though the decrease in the QRB due to returning WUs not as fast at lower Power Limits is not as significant as the increase in GPU efficiency, factoring in the System overhead means that the lowest power is usually NOT the most overall efficient but the peak in efficiency is usually somewhere around 70-90% of the GPU’s FE Power Limit.

 

[TVwazhere]

3 minutes ago, Gorgon said:

Yes, the lower the power is limited, the more you can push the overclock, but there are Twp other factors, too high of an overclock can lead to an immediate failure when the WU returns from the CPU from a checkpoint or calculation as the chip has cooled down so it tries to boost up to 2200MHz or something equally unstable.

 

The second factor is the Quiescent load from the system is a fairly constant 35 to 70W depending on your non-GPU Hardware and this constant leads to an overall larger decrease in overall system (as opposed to just GPU) efficiency with lower GPU power draws. The efficiency (Work/time or PPD/W) should be read at the wall including the Power supply losses and what I’ve found through testing is that, though the decrease in the QRB due to returning WUs not as fast at lower Power Limits is not as significant as the increase in GPU efficiency, factoring in the System overhead means that the lowest power is usually NOT the most overall efficient but the peak in efficiency is usually somewhere around 70-90% of the GPU’s FE Power Limit.

I'll have to play around with it more in the near future. Currently both 1070ti's are OC's +150 on the GPU and 300 on the memory. So I might have a bit more headroom now for OCing since the cards havent hit temp limit but are bouncing on the 60% power limit. 

[Gorgon]

33 minutes ago, GOTSpectrum said:

Also the errors in folding will compound over tint, as each frame is a child of the one that preceded it. So if you make even just a 0.1% error every frame, you'll be massively out by the end. Rendering the entire simulation scientifically useless 

There’s actual sanity checks built in and you usually can’t deliver a bad WU back. If the folding client detects 3 successive errors in a WU due to Overclocking (Bad States) it will abort the WU and you’ll lose all the computational effort. If we assume this happens at 50% through and the average WU is 3 hours you’ve just lost 90 minutes work for a 0.7% increase in clock (15/2025MHz) in order for that to be acceptable you’d have to make up that 90 minutes before another failure which is 8 1/2 days to recover before you can benefit from the marginal single step increase in clock.

[Gorgon]

7 minutes ago, shaz2sxy said:

 

Hey, I analyze things in IT for a living. Math is your friend. Calculus is better than sex. There ain’t no such thing as a free lunch.

 

One or more of the preceding statements is false.

[Gorgon]

And I’m not a stats junkie

[shaz2sxy]

1 hour ago, Gorgon said:

Hey, I analyze things in IT for a living. Math is your friend. Calculus is better than sex. There ain’t no such thing as a free lunch.

 

One or more of the preceding statements is false.

The false one is .... the IT one lol

56 minutes ago, Gorgon said:

And I’m not a stats junkie

ooo squiggly lines....

[Favebook]

8 hours ago, GOTSpectrum said:

Check out my sig for my main.

 

Secondary system:

 

5960x

ROG rampage V Extreme

4x4gb (quad channel) data ddr3 2666mhz

Gtx 1080 FE (Won last folding month)

Gtx 980 ti palit super Jetstream

Gtx 980 FE

 

Thanks! This will help me tbh!

8 hours ago, Gorgon said:

HEDT (x299 etc) are just not worth it. The boards and CPUs are more expensive and you start getting into expensive power supplies and having to fart around with esoteric eATX cases and PCIe riser cables.

 

The least expensive sli-capable x370, x470, z170, z370 or z390 motherboard. This gives you two PCIe x16 at x8/x8.

A 650W Gold power supply as they’re not much more than a Bronze at this price point and 650W lets you run up to 2 RTX 2070s with a bit to spare.

a 2c/4t processor with 16 PCIe3 lanes off the CPU (i.e not AMD with APU) Intel Pentium Gold 5400 or 5500 or AMD 1200, 1600, 1700, whatever is cheapest and at least 1.9GHz

a single stick of DDR4-2400 you don’t need lots of memory or memory bandwidth.

2 NVidia 1660Ti or RTX 2060 (Supers) (Best bang for the buck in Canada currently)

a M.2 SATA Disk so no cable clutter

Ubuntu 18.04 LTS Desktop

A box to put it in or a piece of plywood

110 or 140mm fans

 

Recently I’ve been getting 2700 and 2700x CPUs and running BOINC on the unused threads as these are found deeply discounted.

 

If money were no object I’d get a x299 with 4 PCIe slots and 3 M.2, a mining chassis, a 2920x, couple of 1000W power supplies and 3 x Powered M.2 to PCIe3 x4 adapters and 4 PCIe extenders and 7 cards, 4 x 2070 or 2080 Supers and 3 x 2060 Supers. PCIe3 x4 under Linux is sufficient to drive a 2080Ti with little to no loss in performance.

Yeah, thank you for explaining everything so thoroughly, but I think I will go for some HEDT which will I probably use as my main rig and this one that I currently have will become secondary.

Also, I've always used Windows, I've never had experience with Ubuntu, so that might be interesting to thing to do.

6 hours ago, yaboistar said:

https://www.ebay.co.uk/itm/i7-4820k-motherboard-16-GB-hyperx-ram/174023864701

even this for example - 40 pcie lanes, 16GB of ram, 3 16x slots (split it out into 5 8x slots??)
£210 with the option to haggle it down

Yeah, but sadly... "May not post to Serbia"

[Stormseeker9]

1 hour ago, Gorgon said:

And I’m not a stats junkie

where can i find these squigly lines...

[shaz2sxy]

4 hours ago, yaboistar said:

all i gotta do is keep my folding output up for another cycle and you'll be back under me where you belong

I hate it when your right....

 

[LavenderPhantom78139]

Sept is my 2md most productive month so far and im only about 2 or 3 days from it being my most productive month. Next month imma slay though

[shaz2sxy]

@DeeGee I'm slowly catching up...

 

[GrockleTD]

What is the PPD difference between 2070S and 1080Ti? The latter is a bit cheaper and better available so I was wondering what the gap was (power efficiency be dammed)

[shaz2sxy]

2 minutes ago, GrockleTD said:

What is the PPD difference between 2070S and 1080Ti? The latter is a bit cheaper and better available so I was wondering what the gap was (power efficiency be dammed)

From what I can tell they are very similar . 2070s might just pip it.

[GrockleTD]

8 minutes ago, shaz2sxy said:

From what I can tell they are very similar . 2070s might just pip it.

But the ability to obtain one warrants the difference?

[Sequence211]

6 hours ago, Spotty said:

I'm so glad that 14180's have finally buggered off.
But 3.3mill PPD? RIP my 1080Ti's usefulness

I am running a 14180 WU now, and it shows as Core 21.

Seems like they fixed the bonus also, since i remember that 14180 did not show bonus credit in the advanced control. but it does now.. 

[Spotty]

Just now, Struck said:

Seems like they fixed the bonus also, since i remember that 14180 did not show bonus credit in the advanced control. but it does now.. 

Ahh, I was actually curious whether or not they fixed the servers to calculate the QRB or if we all just moved on to WU's on a different server that wasn't having the problem. Good to know it's all fixed.

[Gorgon]

1 hour ago, Favebook said:

Thanks! This will help me tbh!

Yeah, thank you for explaining everything so thoroughly, but I think I will go for some HEDT which will I probably use as my main rig and this one that I currently have will become secondary.

Also, I've always used Windows, I've never had experience with Ubuntu, so that might be interesting to thing to do.

Yeah, but sadly... "May not post to Serbia"

That would be my preference for a new daily-driver too. In my case I'm still using a Xeon E3-1231-v3 with 32GB ECC on a SuperMicro Motherboard. With it's rocking 3.57GHz all-core boost on it's 8 threads it's starting to get a little slow so I've been eyeing Threadripper but they're coming out with new models soon so I'll likely wait until next year.

 

I'm using Windows 10 for my daily driver and other than that along with a Win 7 Pro VM for iTunes the rest of my systems are Ubuntu or other Linuxes.

 

My next system will likely have Linux as the Host Operating System and Windows as a VM on it.