Top GPUs for Folding@Home

[Endgame124]

FYI, the Geforce 520 will never get the fast return bonus, so even if the compute per watt is good, it’s points per watt will be lower than other options. Similarly, the compute pipeline on the 2080ti is so wide that many folding jobs won’t take full advantage of the cards. Since folding points is how F@H values it’s work, we really should optimize for points per day. This likely means somewhere between the 1660 and 2060 super are likely going to be the best value of points / watt / dollar.

There are very few numbers on the 1660 series of cards when it comes to points though, so I’m testing a 1660 super myself.

[NoMoreQuarantine]

In your list for instance, the 2060 Super will definitely be better than the bottom 3 cards if not the bottom 4 in PPD. Also, where does Vega 64 fit it?

Here is a snippet of my full spreadsheet. The 2060 Super got knocked off the list because it was more expensive than the 5700 XT & 5700 which are both capable of higher operations per second.



As far as "agreeing" with my table or the one you linked, I would recommend you compare what each are measuring. PassMark uses a series of benchmarking tests designed to estimate the average ops/sec: https://www.passmark.com/products/perfo ... ompute.php The table you linked is based on user submitted readings of the estimated PPD from either the client or the benchmarking tool. The issues I have is that the FAH core used for the client and benchmarking tool are substantially different (the client has approximately 25% higher performance right now) and the benchmarking tool produces multiple PPD reading which the user should average before submitting, but as there is no guidance given that I have been able to find, it is left up to the user to average, sum, or just pick one they like. Meanwhile the client will produce different PPD estimates based on the current workload, which makes it a bad benchmarking metric (this wouldn't be as bad with a large sample size across varied workloads).

The ideal would be to test all of the processors using only a benchmarking tool optimized for the current FAH core.

Edit: found this spreadsheet which supplies the sample size data I was looking for, it's not great : https://docs.google.com/spreadsheets/d/ ... tput=html#
That said, with a larger sample set, an updated benchmarking tool using the current core, guidance to only use the benchmarking tool (and how), and version control to keep track of future core changes this could be far more accurate than PassMark.

[PantherX]

...The issues I have is that the FAH core used for the client and benchmarking tool are substantially different (the client has approximately 25% higher performance right now) and the benchmarking tool produces multiple PPD reading which the user should average before submitting, but as there is no guidance given that I have been able to find, it is left up to the user to average, sum, or just pick one they like. Meanwhile the client will produce different PPD estimates based on the current workload, which makes it a bad benchmarking metric (this wouldn't be as bad with a large sample size across varied workloads, but I cannot find any way to view the sample size)...

Assuming that you're talking about this Google Sheet: https://docs.google.com/spreadsheets/d/ ... Ek/pubhtml
Scroll to the bottom and where it says source, this is the link: https://www.overclock.net/forum/55-over ... abase.html
On the overclock.net Forum, there's preview screen and if you scroll all the way to the right, there's a column called Sample Size.

[NoMoreQuarantine]

On the overclock.net Forum, there's preview screen and if you scroll all the way to the right, there's a column called Sample Size.

Thanks, I found that a little bit before you posted and added an edit to the end of my post.

[gordonbb]

... As for the 2060 Super, it got knocked off the list by the AMD Radeon RX 5700 XT.

Are you basing your list on theoretical performance rather than measured results?

I have 2 RX5700XTs from 2 different AIB vendors and can say with certainty that they are worse performers than the RTX 2060 Super and I would not hesitate to recommend the RTX 2060 Super over a RX5700XT. Part of this is based on their performance but the other consideration is the maturity of the AMD Drivers. I'm not a Team Green fanboy nor a Team Red Hater and I do feel that Navi has great potential and NVidia certainly does need competition.

Using current drivers under Linux looking at a 3-day average of PPD:

GTX 1660Ti 700kPPD
RX5700XT 963kPPD
RX5700XT 1.01MPPD
RTX 2060 1.1MPPD
RTX 2060S 1.2MPPD
RTX 2070A 1.3MPPD
RTX 2070S 1.85MPPD
RTX 2070S 1.88MPPD

All cards are running stock with no over-clocks but with manual fans to keep GPU core temperatures under 75C

[NoMoreQuarantine]

Are you basing your list on theoretical performance rather than measured results?

PassMark DirectCompute results.

GTX 1660Ti 700kPPD
RX5700XT 963kPPD
RX5700XT 1.01MPPD
RTX 2060 1.1MPPD
RTX 2060S 1.2MPPD
RTX 2070A 1.3MPPD
RTX 2070S 1.85MPPD
RTX 2070S 1.88MPPD

Interesting, you are measuring major differences even between the two 5700 XTs. When you say they are running stock, are they running at the recommended NVIDIA/AMD clock speeds? A lot of cards come factory OCed by the videocard makers and have to be told to operate at standard clock rates. Do all cards have max RAM operating at recommended frequencies?

[PantherX]

...Do all cards have max RAM operating at recommended frequencies?

Officially, F@H recommends stock frequencies which are set by the vendors for CPUs and GPUs. However, you have partners (ASUS, MSI, EVGA, etc.) who either use the reference PCB with custom frequencies or custom PCB with custom frequencies.

[NoMoreQuarantine]

Officially, F@H recommends stock frequencies which are set by the vendors for CPUs and GPUs. However, you have partners (ASUS, MSI, EVGA, etc.) who either use the reference PCB with custom frequencies or custom PCB with custom frequencies.

That makes sense. I actually meant the recommended frequencies as specified by AMD & NVIDIA for the GPU.

[NoMoreQuarantine]

Since there has been so much discussion on whether the user submitted-PPD data collected by TheBladeMaster01 (https://www.overclock.net/forum/55-over ... abase.html) is more accurate than the PassMark data I used, I decided to repeat my process of generating an order list at each price point using the user-submitted PPD list data. Surprisingly, they match almost perfectly. The only difference is that the TITAN V has been moved below the 2080 Ti (therefore being eliminated by my process as it is more expensive) and there are several GPUs that have not been submitted to the list; Including in the location where RX 5700 & RX 5600 XT sit in my list. I believe it is likely that the RTX 2060 would be eliminated if they were added. That said, gordonbb has some numbers he posted that do not match either lists, just put a fly in my ointment



If I sort the list by the GFLOPS column from the PPD table the TITAN V gets moved back to the top, the Radeon VII moves up, and the Vega 56 moves up and eliminates the RX 5700 & RTX 2060. I am not sure where they got the GFLOPS measurement (I don't see it in their form submission. Maybe from the TPF somehow?) This reminds me, one of the first things I considered doing when making my list was just taking the GFLOPS specified by the GPU maker as the sorting metric, but I still haven't found a single person who has been able to achieve the 110 TFLOPS NVIDIA claims the TITAN V is capable of



I am planning on updating my original post with the user-submitted PPD data as well as adding some sub-$260 GPUs when I have some time.

[foldinghomealone2]

I envy your effort you put in your calculations but frankly said it's not worth a penny.
You could have just looked at the TFLOPs specified in the data sheets, ranked the GPUs by that, and you would have seen similar results.

When it comes down to folding what mainly counts is "shaders x frequency" (aka as TFLOPS), driver implementation (NV wins over AMD) and OS (linux wins over windows).

When comparing PPD from databases you need to have a very high sample size due to the following facts:
- every projects produces different PPD on the same GPU
- even the same project produces different PPD on the same GPU
- different clock speeds on GPUs
- what OS was used
- you don't know how the GPU is connected to PCIe (some miners want to use their mining rigs for folding and wonder about the low output)
- which folding core was used *
- cooling solution of the GPU and its surroundings (see an extreme example of someone folding in Alaska and in Australia at the very same time)
- driver performance
- upload speed of your internet connection

* Most databases lack sample size (see a sample size of 3 for a 2080 and 9 for a 2080 Ti) and are not up to date because they will be filled mainly after a new GPU model is released. That means that most samples used core 21.
All current project use core 22 however. It behaves differently, is 'faster' and produces more PPD

[JimboPalmer]

Let me post some 'alternative' Top GPUs. (as I have found them)

The fastest Nvidia GPU you do not need additional power plugs for is a GTX 1650.
The fastest Nvidia GPU that fits in a Dual Slot Low Profile case is also a GTX 1650
The fastest Nvidia GPU that fits in a Single Slot is a GTX 1070
The fastest Nvidia GPU that fits in a Single Slot Low Profile case is a GT 1030
The fastest Nvidia GPU that is fanless is a GTX 1650

The fastest AMD GPU you do not need additional power plugs for is a RX 560.
The fastest AMD GPU that fits in a Dual Slot Low Profile case is also a RX 560
The fastest AMD GPU that fits in a Single Slot is a RX 550
The fastest AMD GPU that fits in a Single Slot Low Profile case is a R7 240
The fastest AMD GPU that is fanless is a RX 460

Those of you with big beefy cases with big beefy power supplies can ignore these cards!

[HaloJones]

Using a card for folding that you bought for gaming is great but if you're building a rig just for folding do not buy new. Second-hand cards are always better bang for buck for folding. Get a low-power dual-core like a G4400, 4GB RAM, 1070, 80GB SSD, install Linux Mint and you could be cranking out 800Kppd for $400.

[Nathan_P]

Using a card for folding that you bought for gaming is great but if you're building a rig just for folding do not buy new. Second-hand cards are always better bang for buck for folding. Get a low-power dual-core like a G4400, 4GB RAM, 1070, 80GB SSD, install Linux Mint and you could be cranking out 800Kppd for $400.

You could but as I have learnt to my cost that $400 is wasted if the card fails and you have no warranty. At least new you have a warranty to fall back on. I've bought new, refurb and used cards for years , I've always had better longevity with the new cards. YMMV

[foldinghomealone2]

Using a card for folding that you bought for gaming is great but if you're building a rig just for folding do not buy new. Second-hand cards are always better bang for buck for folding. Get a low-power dual-core like a G4400, 4GB RAM, 1070, 80GB SSD, install Linux Mint and you could be cranking out 800Kppd for $400.

And some of us have to consider energy costs as well. I have to pay 0.30€/kWh.
Folding on energy efficient HW is mandatory (for cost reasons but also to consider effects on environment)

(I didn't say that a 1070 is in-efficient in regards to power consumption but can't compete to faster GPUs considering PPD/W)

[HaloJones]

You could but as I have learnt to my cost that $400 is wasted if the card fails and you have no warranty. At least new you have a warranty to fall back on. I've bought new, refurb and used cards for years , I've always had better longevity with the new cards. YMMV

As you rightly say mileage may vary

I've been lucky lately (touch wood). I bought a water-cooled TitanX in 2017 off eBay that has been running 24/7 ever since, overvolted and overclocked. It's produced maybe close to a billion points since then.