Recommendations on used cards for Folding

[MrFrizzy]

So, I am tempted to add a 3rd card to my arsenal and have been looking at the used market.

I would like to have something that can hit 1m PPD (even if not daily) without exceeding $300. These are the cards that seem to be in budget:

• Vega 56
• Vega 64
• GTX 1070 Ti
• GTX 1080
• GTX 1660 Ti
• RTX 2060 KO

I know the Vega 56 and 64 really flex when given large work units (potentially 1.7-2m PPD peak) but get much slower on smaller projects. Outside of that, I don't have much info on what the other Nvidia cards can do. However, since the new 0.0.13 core_22 version is in beta testing and shows a huge boost in performance due to CUDA, it seems like the Nvidia cards may be the way to go.

Thoughts? Power consumption, heat, and noise is not a concern.

[ipkh]

The 2060KO is your best bet.
Nvidia cards will be quite good once the Cuda core goes out.

[psaam0001]

Don't rule out the GT 10xx or GTX 16xx series, which I believe Newgg offers for between $90.00-$200.00 (USD; depending on the GPU). They are affordable enough to get an older system with a PCIe x16 slot going.

The GT 1030 will likely match an AMD Ryzen 3's Radeon Vega iGPU. The GTX 1650/GTX 1660 will perform about 50-70% better from my experience, but your experience may vary.

Paul

[MeeLee]

GTX 16xx series don't get 1M ppd.
For 1M ppd, you'd need an RTX GPU.
And for you to get a second hand GPU right now, is not a good time.
For the NEW price of $300, you could find a 2060Super or KO. Second hand you MAY find a 2070 or 2070 Super.
For just $100-200 added, you can get a brand new 3080, which leaves the RTX 2000 series in the dust!
It's really recommended to go with one of these. I know the extra $$$ may be hard on your finances, but it really pays off in the end!
A 3080 is still hard to get, but should go for $400-500. Meanwhile a new 2070 goes for the same, and is less than half as fast.

[kiore]

My opinion, it best to go with newer generations if possible so RTX 2060 leads that list, but I bought a bargain GTX 1660ti earlier this year and am very happy with it although it hasn't previously met the 1 M ppd line it has been close and I think improvements in cores utilizing CUDA better will see this easily over 1Mppd. Choose what to can get a good deal on keep your options open there are many cards that meet your criteria.

[marknd59]

Your best bet at the moment is RTX 2060 super.

https://docs.google.com/spreadsheets/d/ ... edit#gid=0

[HaloJones]

GTX1070s can get to 1m - the 1070ti I had didn't outperform the standard 1070s by much so I sold it.
I've also got a TitanX that easily goes over 1m but is a tad power-hungry.

The new CUDA FAH core is taking my 1070s up to 1.5m at times although that's a short term benefit I expect while the kinks are worked out on it.

[gunnarre]

GTX 16xx series don't get 1M ppd.
For 1M ppd, you'd need an RTX GPU.

I have seen the GTX 1660 Super make up to 1.1M ppd in OpenCL due to variability, but a more typical OpenCL performance seems to be 800k to 900k PPD. The 1660 Super has faster VRAM than the regular 1660, and that might help increase speeds even though folding doesn't use much VRAM.

[PantherX]

...The new CUDA FAH core is taking my 1070s up to 1.5m at times although that's a short term benefit I expect while the kinks are worked out on it.

AFAIK, the benchmark is done on OpenCL to ensure that AMD/Nvidia GPUs are on equal grounds. Thus, the boost in CUDA is going to be an additional bonus which will be around for a while. After all, it was many years in between the CUDA enabled FahCore so make the most of it while it lasts

[markdotgooley]

I’m using a 2060 and a 2060KO on the same motherboard. The KO is on a PCIe 3.0 x8 slot and the regular 2060 on a 3.0 x4 because that’s all the cheap motherboard can manage. The two get over 2 million points per day in total, despite a typical 85% GPU utilization on the slow slot, sometimes 2.5 million or more from COVID Moonshot WUs. Mine are running a little over 150 watts each under load. I’d suggest the KO although if you can get a good price on the faster 1000-series cards they might be fine (no personal experience).

[gunnarre]

https://docs.google.com/spreadsheets/d/ ... edit#gid=0

The 1660 Super doesn't do as low a PPD as in that spreadsheet. Maybe that's because that spreadsheet used Core 21 numbers. This spreadsheet is closer to what I get, with GPU Core 22 (OpenCL): https://docs.google.com/spreadsheets/d/ ... Ek/pubhtml

If you can buy two 1660 Super cards, those would perform better at folding than a single 2060 KO, so it depends on what price you can find the cards at. But yes, a 2080 Ti or 3080 will give a higher quick return bonus, so you'll see several millions of PPD on those cards.

[bruce]

I know the Vega 56 and 64 really flex when given large work units (potentially 1.7-2m PPD peak) but get much slower on smaller projects. Outside of that, I don't have much info on what the other Nvidia cards can do. However, since the new 0.0.13 core_22 version is in beta testing and shows a huge boost in performance due to CUDA, it seems like the Nvidia cards may be the way to go.

Thoughts? Power consumption, heat, and noise is not a concern.

From all I've seen, the differences between small proteins and large proteins is similar for AMD and NV. I've never seen this carefully analyzed, though. Somebody should compile a spreadsheet based on the number of shaders and the numbers of protein atoms. The theoretial FLOP value is a product of the shader count and the shader speed and a third factor (often equal to 2). I suspect that the ratio of Atoms to shaders is what influences the large vs. small protein performance but until that data is collected, I'm only guessing. Certainly the drivers also matter, so CUDA vs. OpenCL will change things once the new FAHCore_22 revision is available for distribution.

[MrFrizzy]

It seems that the RTX 2060/KO is the way to go for the price. After looking at your responses and checking out the various spreadsheets/databases that were linked, it appears that it is the best performing out of the cards in the budget specified!

Don't rule out the GT 10xx or GTX 16xx series, which I believe Newgg offers for between $90.00-$200.00 (USD; depending on the GPU). They are affordable enough to get an older system with a PCIe x16 slot going.

The GT 1030 will likely match an AMD Ryzen 3's Radeon Vega iGPU. The GTX 1650/GTX 1660 will perform about 50-70% better from my experience, but your experience may vary.

Paul

I already have a GTX 1650 (GDDR5 model) and after a bit of tweaking and some modding, it averages around 300-350k PPD. For the price, it does not perform that well though it does sip power. Anything lower than that card really loses on the value aspect of things. I had my old 980 Ti going for a bit, but one of the FETs popped and I haven't been successful in resurrecting it...yet. That did between 800-950k PPD and while I can get one used for under $200, it just pulls a lot of power and is already 5 years old. My second system has the means for 2 more graphics cards so I figure I might as well add something more powerful than a GTX 1650 in it.

GTX 16xx series don't get 1M ppd.
For 1M ppd, you'd need an RTX GPU.
And for you to get a second hand GPU right now, is not a good time.
For the NEW price of $300, you could find a 2060Super or KO. Second hand you MAY find a 2070 or 2070 Super.
For just $100-200 added, you can get a brand new 3080, which leaves the RTX 2000 series in the dust!
It's really recommended to go with one of these. I know the extra $$$ may be hard on your finances, but it really pays off in the end!
A 3080 is still hard to get, but should go for $400-500. Meanwhile a new 2070 goes for the same, and is less than half as fast.

I don't know where you are getting your prices, but a new 2060 is still well over $300 everywhere I look and the super is well over $400. The KO is much closer to the $300 mark new, but that doesn't make sense when I can get a used one for much less. It is also rare to find a used RTX 2070 under $300 and the supers seem to be almost always around the $400 mark.

While it is a novel idea to have a brand new RTX 3080 for folding, there is just no way for me. Not only is the price far too high (they will likely never sell for the $700 MSRP even when they are readily available), but that caliber of card would be go in my gaming system where I do care about heat, noise, and power consumption. I just want to add a second, used card to my secondary system where the noise, heat, and power does not matter; something that I can add and "write off" the cost to a good cause hence the $300 budget.

Nevertheless, I will likely wait a bit before I make any sort of purchase anyway since the other RTX 30 cards and RDNA2 releases will also likely decrease the used market pricing.

[psaam0001]

I agree on the above.... Though, I am trying to get what I can out of my legacy hardware while I can.

The only other upgrades I plan to do, is ditch the PATA HD in my Win 7.0 system, and pop in 2 spare SATA HD's. One for the OS/Software, and one for the F@H data. No more system hardware changes can be made to the 2 Fedora systems I use (a Dell PowerEdge T105 and an Alienware M17 Laptop); or to the above referenced Win 7.0 system. It's just time for me to win the lottery and build a new machine.

Paul

[marknd59]

The 1660 Super doesn't do as low a PPD as in that spreadsheet. Maybe that's because that spreadsheet used Core 21 numbers.

The numbers come from a Chrome Extension for the F@H web client with real time PPD benchmark database gives feedback on your PPD by GPU model and WU.

For more information about the extention the home page for the extention -

https://folding.lar.systems/


GPU PPD rank real time database page -

https://folding.lar.systems/folding_dat ... pd_overall