Folding Forum

foldinghomealone2 wrote: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.

Just because there is another method to arrive at the same results doesn't make the results worthless. The benefits of doing it my way is that it is real world measurements as opposed to theoretical. Plus, I was able to start from existing tables, if I had to look up the datasheet of every single GPU, then I would have been at this for days.

I completely agree that sample size is an issue and mentioned it in my previous post. Personally, I trust the PassMark table more than the user submitted PPDs table as it uses standardized tests designed to measure the actual ops/sec, but it still has low sample sizes for some of the cards and does not restrict overclocking.

JimboPalmer wrote: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!

Thanks JimboPalmer! It seems like a lot of people would like to see a list targeted at HW for dedicated folding PCs. For something like that, performance/watt is critical and running multiple GPUs would be the way to go (single slot would be required). I am not interested in making such a list right now, but your post along with what others have commented in this thread will hopefully help those seeking such a thing.

NoMoreQuarantine wrote:...Personally, I trust the PassMark table more than the user submitted PPDs table as it uses standardized tests designed to measure the actual ops/sec, but it still has low sample sizes for some of the cards and does not restrict overclocking.

But you have to understand that PassMark is not designed to benchmark folding hardware. It tests hardware with several different tests to build an average score. Like it makes some tests with double precision but FAH uses mostly single precision. It's just more or less acurate by accident.
Like you can see with the 5500XT benchmark. It's rated higher than in a real world folding scenario.

The only benchmark useful for folding is FahBench but it doesn't work anymore as it doesn't support core 22.

foldinghomealone2 wrote:But you have to understand that PassMark is not designed to benchmark folding hardware. It tests hardware with several different tests to build an average score. Like it makes some tests with double precision but FAH uses mostly single precision. It's just more or less acurate by accident.
Like you can see with the 5500XT benchmark. It's rated higher than in a real world folding scenario.

The only benchmark useful for folding is FahBench but it doesn't work anymore as it doesn't support core 22.

I agree, having an up to date FAHBench and a large sample size of GPUs would be the ideal, but we don't currently have that.

I assume that you meant the 5700 XT vs 2060 Super discussion. Even if I take the single precision performance listed by AMD and NVIDIA for those GPUs, the 5700 XT is still higher. 5700 XT @ 9.75 TFLOPs > 2060 Super @ 7.2 TFLOPS.

Vega is so hit and miss, you can get 1M PPD out of one if the project has enough atoms, if the projects are on the smaller side performance tanks.

I'd be happy to partake in a drag race to 10 million points, my 2060 Super vs a 5700XT so we can test the theory.

My Super is factory OC'd and hits core clocks of 1920mhz under full load, 73 degrees celcius and pulling 181 watts and running on Windows 10 1909

I can provide FahBench build with core_22 and mixed precision for Windows.
viewtopic.php?f=38&t=24225&p=325202#p325202

Juggy wrote:I'd be happy to partake in a drag race to 10 million points, my 2060 Super vs a 5700XT so we can test the theory.

My Super is factory OC'd and hits core clocks of 1920mhz under full load, 73 degrees celcius and pulling 181 watts and running on Windows 10 1909

I believe you, but I can only work with the numbers I have available and everything I can find says that the 5700 XT should perform better for FP32 compute workloads. That said, your suggestion wouldn't be very scientific. An alternative would be to run FAHBench at AMD/NVIDIA spec frequencies on cards with max VRAM, ideally on the same system or as close as possible. I know that FAHBench uses an old core, but it should give us an indication of their relative performance. Run a couple rounds and average the results. Maybe our friend gordonbb could help us put this discussion to rest (assuming he can take a break from crushing us all in the FAH rankings )

After getting the baseline performance, we could look at comparing OCed performance. This will vary depending on the binning and would require a significant number of samples to creative predictive results.

foldy wrote:I can provide FahBench build with core_22 and mixed precision for Windows.

Awesome! Where did you get that? Did you compile it yourself?

NoMoreQuarantine wrote:

Juggy wrote:I'd be happy to partake in a drag race to 10 million points, my 2060 Super vs a 5700XT so we can test the theory.

My Super is factory OC'd and hits core clocks of 1920mhz under full load, 73 degrees celcius and pulling 181 watts and running on Windows 10 1909

I believe you, but I can only work with the numbers I have available and everything I can find says that the 5700 XT should perform better for FP32 compute workloads. That said, your suggestion wouldn't be very scientific. An alternative would be to run FAHBench at AMD/NVIDIA spec frequencies on cards with max VRAM, ideally on the same system or as close as possible. I know that FAHBench uses an old core, but it should give us an indication of their relative performance. Run a couple rounds and average the results. Maybe our friend gordonbb could help us put this discussion to rest (assuming he can take a break from crushing us all in the FAH rankings )

After getting the baseline performance, we could look at comparing OCed performance. This will vary depending on the binning and would require a significant number of samples to creative predictive results.

I agree with you, also don't think using numbers posted on passmark are very scientific either as it is not application/project specific. It's almost like saying this car is faster because it does 0-60 quicker but what we're actually doing is driving around a track and the slower car handles better so in practice it is quicker.

foldy wrote:I can provide FahBench build with core_22 and mixed precision for Windows.

That would be awesome, thank you

NoMoreQuarantine wrote:

foldy wrote:I can provide FahBench build with core_22 and mixed precision for Windows.

Awesome! Where did you get that? Did you compile it yourself?

Yes compiled on Windows as it is open source at github.
viewtopic.php?f=38&t=24225&p=325202#p325202

foldy wrote:Yes compiled on Windows as it is open source at github.
viewtopic.php?f=38&t=24225&p=325202#p325202

You are the best! Now all we need is a database to store the data, a frontend, and guidance on how to submit. We'll need to make sure people are not submitting OCed results and that we keep track of VRAM quantity. I've never done anything like that before, any suggestions would be appreciated. TheBlademaster01 used a google docs for his database and frontend; maybe I could do something similar. The ideal would be to have it all handled by FAHBench, but I'm not one to beg for features.

What settings do we use or simply leave it as-is and run for 5 minutes?

With all the charts, a new way to benchmark, and talks of a spreadsheet.... you should all tremble in fear.

When I unleash the on board Vega 11 graphics on this 2400G things are going to get heated!




(yes that was completely tongue in cheek)


Being serious now, I'm surprised with so many people folding that more databases haven't been made over the years. Years ago it seemed that almost everyone folding was in the geek/power user/overclocker/gamer club. At least from the contact I had with people (and yes I considered myself in that club as well). But now systems have so much more tech that people can go from just about any level as a base, up to insane folding systems.

But even now, having more data is good. I'm not a power user any more, and unless I took it up as a hobby of sorts, chances are I'd just put a decent power efficient card in our every day use machine. Power efficiency and heat would be much more a factor than cost really, since we don't want to upgrade cases, fans, etc.... or have a loud system running on our desktop.

So "best" would remain highly subjective to many of the people using the data. Though I could spring for a higher end card, I'd probably opt for something along the lines of the 1660 Super that a couple of people have mentioned.



Just food for thought. Even back in my overclocker power user days I realized that chasing max performance is a money pit that is ever changing, and many people decide not to play that game.