Some numbers: RTX cards

[Theodore]

Running each card at their optimal lowest power setting:

* 4x RTX 2060s at $350 each ($1400). will fold at ~4,1 M PPD,
running at 130W each (that's about 625W at the wall).
* 3x RTX 2080s at $700 each ($2100) will fold at ~4,2 M PPD,
running at 150W each (that's about 525W at the wall).
* 2x RTX 2080 tis, at $1200 each ($2400), will fold at ~4,4M PPD,
running at 230W each (that's about 575W at the wall).

The wall power draw number, roughly equals annual electricity cost, so it seems that:
- the 2080s are most economic for folding in the long run,
- the 2060s are cheapest for folding, with a max of 7 years, 24/7 folding.
- and the 2080 TIs are the fastest (duh) .

Same Intel quad core benchmark system is used, under Linux.
As long as you're using Linux, 4 CPU cores work fine with 4 dedicated GPUs.
No extra CPU core is needed for Linux.
One could use a lower powered quad core (45W TDP) 2Ghz CPU for the 2060, and save ~20-25W on electricity; where as the 2080 TI needs a 3Ghz (~65W TDP) CPU, which brings total system overhead to ~100W.

[ari2asem]

any numbers about gtx-1000 series cards (gtx 1060, 1070 or 1080)? how fast is gtx 1000 serie folding?

is rtx folding faster than gtx? or it depends on cpu type (number cpu cores) or depends of speeds of used ram memory? (ddr3 or ddr4, or amount of ram memory) ?

i am planning to build a rig, only and specially for FAH. now trying to gather information so that i can make my purchase of hardware

[bruce]

The size and speed of VRAM are relatively unimportant. For one number which can be used as a quick comparison, look at either the SinglePrecision GFLOPS or multiply the number of shaders by the clock speed.

[Nathan_P]

TDP is a rough approximation of what a chip will consume at full power, when running F@H on a gpu you will find that your cpu is idling and pulling very little power. Under the load of 2 gpu's my system is using about 75w plus the gpu's and that includes a very power hungry PLX chip.

[ari2asem]

thank you all for answers

[bollix47]

any numbers about gtx-1000 series cards (gtx 1060, 1070 or 1080)? how fast is gtx 1000 serie folding?

is rtx folding faster than gtx?

The GTX models you mention can get ~400K 700K 800k PPD respectively in linux.
The RTX 2060 can get ~1.2M in linux. So, yes, the current RTX models are doing well in terms of PPD.

[bruce]

GTX 1080 Ti GFLOPS = 10609 (11340) TDP = 250W
GTX 1080 GFLOPS= 8228 (8873) TDP = 180W
GTX 1070 Ti GFLOPS= 7816 (8186) TDP = 180W
GTX 1070 GFLOPS = 5783 (6463) TDP = 150W
RTX 2060 GFLOPS = 5242 (6451) TDP = 160W
GTX 1060(5/6) GFLOPS = 3855 (4372) TDP = 120W
GTX 1060(3) GFLOPS = 3470 (3935) TDP = 120W

Compare the prices and as far as FAH is concerned, the 10 series gives you more for your money -- unless you want ray-tracing in your game views.

[MeeLee]

@Bruce,
The RTX 2060 is probably the best card in terms of PPD per dollar.
It folds at the same speed as a 1070 ti to 1080, but at much lower electric cost.

The DDR6 memory is much more energy efficient,plus the RTX cards have more control on power than the GTX cards, making it possible to fold at 96% of the PPD from stock speed, at only 76% of stock power consumption.

1 year of folding on a single RTX 2060, and the difference between a used 1070ti in purchase price is negated.
Not to mention current RTX cards are expected to get a price drop, after the new RTX cards will appear.
A potential RTX 2060 Ti is reported to be arriving soon, with 256 bit memory bus, vs the older 192bit.
These newer cards will be similar in price to the current RTX cards, while current RTX cards will drop in price.

If I were to do it all over again, I probably would have ran 4x RTX 2060 cards, as the 2080 I currently own, will probably never get me the initial purchase difference of $350 on electricity back.

[Mstenholm]

For some real world numbers on 1070/1070ti/2060 and 2070 have a look here - https://docs.google.com/spreadsheets/d/ ... edit#gid=0

[ari2asem]

For some real world numbers on 1070/1070ti/2060 and 2070 have a look here - https://docs.google.com/spreadsheets/d/ ... edit#gid=0

thank you for the chart....but some questions, maybe stupid ones. then sorry for that

question 1.) i see some PPD values like 636,001 or 1140,658? is it PPD more than 600 thousand and more than 1 million??

question 2.) i see a tab called Run, Clone, Gen ?? what does this mean?

question 3.) i see same card with different projects having different PPD values. is this because every project is different and no-any project has the same difficulty? if yes, then i was wrong thinking that PPD values are more or less stable and reproducible values, meaning constant values?

question 4.) if PPD is not constant value, are FP32-values constant? or are FP32-numbers also changing so much as PPD (depending on project)?

thanks in advance,
erik

[MeeLee]

1, RTX cards all run at 1Million points or more.
2, Run, Clone, Gen have to do with whatever project you're running. The numbers don't really matter for those who fold, but perhaps some of the forum seniors can explain this to you.
3, Yes, every project, and even some WUs of the same project, have a different amount of atoms. This will result in your graphics card being either fully or only partially used. Most RTX 2060 or up cards are often used partially. This results in lower GPU usage, but quicker return bonus.
4, the above is the reason why PPD values can change. FP32 values are constant on a GPU unless you over/underclock, but PPD values also vary, depending on how much percentage of the card is used.
You can't change that value, and you can't pick and choose.
Whatever WU is assigned to the card, is assigned, and needs to be finished.
Some WUs will give lower PPDs, while others give higher than normal PPDs.

[Nathan_P]

RTX do not all run at 1m PPD or more all the time. My RTX 2060 does anything from 829k to 1.1m

[bruce]

Run, Clone, Gen have to do with whatever project you're running. The numbers don't really matter for those who fold, but perhaps some of the forum seniors can explain this to you.

Every WU is unique and is generally assigned to only one machine. If, for any reason, it is not returned by the deadline, it is reassigned to someone else because every WU must be completed if possible. Once a WU has failed, it may be assigned to more than one person to avoid secondary delays after the initial deadline expiration.

The four values ... Project, Run, Clone, Gen numbers ... identify that particular WU. As was stated above, Project is the one that you need to know about. Run and Clone simply index the WU by a couple of simulation-related values. The Gen number is worth understanding.

A Trajectory is a molecular motion starting at some time and continuing to some later time. Generally speaking, it's difficult to simulate trajectories that are "long enough" to keep the scientists happy. In any case, a single trajectory is divide up into a series of shorter steps which are eventually joined together into a time-ordered sequence -- called Gens. The trajectory XYZ (Project X, run Y, Clone Z) starts with Gen 0 and continues to Gen 1, 2,3,4... to some ending point. If you're assigned Gen K, someone MUST finish it before Gen K+1 can be generated.

Some WUs will give lower PPDs, while others give higher than normal PPDs.

(First, I would add the words "... on my GPU")

FAH makes a concerted effort to align the PPDs of all projects to similar values. Nevertheless, it's impossible. PPD performance is not linear and some GPUs prefer small proteins while others prefer large proteins.

Consider 2 WUs that are two different proteins which give the same baseline PPD on GPU X. Now run them both on GPU Y. (Assume Y is twice as fast as X, but it really doesn't matter.) One WU may get 2.1x the PPD and the other WU may get 1.9 the PPD. This is further exaggerated when you add bonus points to baseline points.

We're now OFF TOPIC. Let's get back to numbers for the RTX cards and not discuss this any further in this topic.

[toTOW]

RTX do not all run at 1m PPD or more all the time. My RTX 2060 does anything from 829k to 1.1m

It can go down to 1060 performance on small WUs on Windows ...

[Nathan_P]

RTX do not all run at 1m PPD or more all the time. My RTX 2060 does anything from 829k to 1.1m

It can go down to 1060 performance on small WUs on Windows ...

My numbers are from linux, no windows gpu folding here anymore but i can fully believe that level of performance drop