Folding Forum

HaloJones wrote:wasted capacity effectively. they become no more efficient than a card with half the cores and end up with only maybe 50% utilisation. give them a really big protein with >100000 atoms and the card can get up to >90% utilisation that then allows them to get comparatively low TPF, to return far more quickly than a lower spec card and get an exponentially large quick return bonus

Ah, because some things have to be done in order before continuing through the simulation it winds up spending clock cycles with unused compute units. Thanks!

It would be cool if we could give cards like that additional WUs when they have compute units available (like extra "logical" GPUs in FahClient). For now though, it makes me think that an "optimal" system might be one which has several GPUs that can efficiently handle the average protein size plus standard deviation in parallel rather than one big one with an equal number of compute units.

i think some people have tried running two units on a single card but over time it's better (science and points) to just run one unit at a time.

one day, we will have a system that optimises the science by giving the biggest units to the cards with the most cores, the smallest units to the cards with the lowest cores, etc. but we are a long way from that.

FAH is at best a very amateur effort (and I mean that with no disrespect). some serious money and programming expertise could be achieving far more throughput

HaloJones wrote:...FAH is at best a very amateur effort (and I mean that with no disrespect). some serious money and programming expertise could be achieving far more throughput

Their approach of divide and conquer where the FAHControl, FAHViewer, etc are open source for all the OS which will allow the community to be more involved. Thus, there will be a reduction in tasks allowing the F@H Team to focus on what they do best, more science and more research

I gave you all fair warning, time to update your drivers, tweak your cooling and memory timings, and brace for what was to come.





Device - Vega 11 onboard Ryzen 5 2400G
Compute - OpenCL
Precision - Mixed
WU - Real
Accuracy check enabled
NaN Check - Disabled
Run length - 1 minute

Score - 10.1878
Scaled Score - 30.7433

Atoms 64614


Hey, but it's only pulling about 12w, so some of the cards you guys have pull that at idle!

I did it mostly to give everyone a laugh and claim the bottom spot in hopes others with lesser cards might run the bench. But TBH, with the caliber of cards quite a few of you have, I was really expecting most if not all the upper end cards to best this thing 10-12x over. It pulls 16-19K a day (raw points without QRB) but the fans don't even come off idle on the system, and you wouldn't know it was running without looking at temps of the FOH client.

Maybe I'll bump the clocks up and crack 20K PPD. But I'll give you all some time to work on your rigs first.




As for the info on current GPUs, in this case all the data in the world that isn't FOH numbers doesn't seem to be worth much. Actual PPD numbers from users seems to be the way. The link from Overclockers has a lot of good data, and a continuation of a similar idea would be good. But it does get complex with the various OS, hardware, user tweaks, driver revisions, etc, etc. Not to mention the changes in WU size and efficiency over time. It seems that for it to work well people would have to reach out to various other forums and such, as this forum isn't very active for the number of people folding.

And thanks to doing some folding again after a long break, I've now been looking at GPU's. But this is strange to me, because it's the first time it's not for gaming or productivity goals.

BobWilliams757 wrote: Score - 10.1878
12w
It pulls 16-19K

It's beautiful

ptchernegovski wrote:

JimboPalmer wrote: I have a low profile no extra power GTX 1650, I get 320k PPD. A full height with better fans will be faster.

Previous to that, Low Profile GTX 1050 ti was 150k PPD, so the 1650 was a real step up.

Is that a normal 1650, or a super? I am deciding between 5500 XT and a 1650 currently

Unless someone has one out there that I haven't seen, the only 1650 cards that require no extra power are the regular 1650. The 1650 super requires a 6 pin PCIE at a minimum, I think some have 8 pin. But I'm looking hard at both myself, leaning towards the 1650 Super.

The EVGA GeForce RTX 2060 KO GAMING should be interesting to see test results from as it has a fused down 2080 core. Apparently it has 2080 like compute performance in blender.

Tohya wrote:The EVGA GeForce RTX 2060 KO GAMING should be interesting to see test results from as it has a fused down 2080 core. Apparently it has 2080 like compute performance in blender.

viewtopic.php?f=38&t=32143

overall the thread concluded that the EVGA RTX2060KO despite the claims on GamersNexus about Blender did not show any advantage at Folding

Juggy wrote: There are quite a few motherboards that support more than 2x PCIe 8x. X399 and X299 supports 16x/8x/16x/8x in 4 slots for instance (differenmt combo's of this as well) simultaneously on PCIe 3.0.

On paper, yes.
But you have no idea how many motherboards I had to return, because they had issues running more than 2 GPUs. Some ran 3 (on paper 4), but once I added another GPU, they would either not boot, or the GPU wouldn't be recognized (more the former).

HaloJones wrote:
wasted capacity effectively. they become no more efficient than a card with half the cores and end up with only maybe 50% utilisation. give them a really big protein with >100000 atoms and the card can get up to >90% utilisation that then allows them to get comparatively low TPF, to return far more quickly than a lower spec card and get an exponentially large quick return bonus

Not really.
If you can afford it, the best card to get is a 2080 Ti, because it's just the most efficient card on the market.
While the Titan RTX is even better, few people agree that their nearly triple price is worth the ~10% performance gain.

But sometimes on Amazon they sell some Gigabyte Windforce 2080 Ti refurbished units for $900-999, which is a pretty good deal!
One of these units can pull 2,4M PPD in Linux on core 21, all the way to 4,4M PPD on Core 22!
They run best in an open bench power capped to 220W.

If you run core 21, with lower atom count, you can easily move the power slider to the left, and run your GPU at 180W.
Less atoms just means lower power consumption is possible.
Many newer and high end GPUs run small projects at much higher boost frequencies, and sometimes hit 2100Mhz or higher, without errors.

If you're looking for a better deal, try looking at the 2060KO (or 2070KO, or 2080KO).
It's a step up from the 2060/2070/2080 Super GPUs, which are a step up from the regular 2060/2070/2080 GPUs.
Next up the 1660 Ti, then 1660 Super.
Only after the 1660 Super, would I look at AMD for folding, simply because they're not as energy efficient as Nvidia, and their Linux and OpenCL support is sometimes lacking.

MeeLee wrote:Many newer and high end GPUs run small projects at much higher boost frequencies, and sometimes hit 2100Mhz or higher, without errors.

That's pretty cool. I imagine there is still a limit to the amount of switching the reduced number of transistors can handle even with the higher temperature overhead; it's probably still not as efficient for small projects as two smaller cards running at similar frequencies processing two WUs at the same time. Of course the best case is just to have multiple 2080 Ti's running in parallel

MeeLee wrote:If you're looking for a better deal, try looking at the 2060KO (or 2070KO, or 2080KO)

Neat! From the testing of those cards, it doesn't look like they have much benefit for direct FP32 computing unfortunately.

I hadn't seen results for the cards I use, so here you go:

Windows 10 Pro - MSI MPG X570 Gaming Pro Carbon Wifi - Ryzen 3800x

Device - GeForce GTX 970
Compute - OpenCL
Precision - Mixed
WU - Real
Accuracy check enabled
NaN Check - Disabled
Run length - 3 minute

Score - 40.8069
Scaled Score - 123.141

Atoms 64614

Edit -

In an effort to make a predictive ranking of GPU performance, I took the specs of all of the desktop GPUs since 2016 for both AMD and NVIDIA from here https://www.techpowerup.com/gpu-specs/ and placed them in tables sorted by FP32 TFLOPS at boost frequency. The calculation for the TFLOPS is TFLOPS = FP32 * Frequency * 2. The last two columns show watts/TFLOPS & dollars/TFLOPS at boost frequency. The prices are the lowest prices for a new video card I could find on Amazon or Newegg. This ranking assumes that memory bandwidth has negligible impact on FAH performance.

First, the full table of NVIDIA GPUs:


The full table of AMD GPUs:


I then removed any GPU that had lower TFLOPS and higher price than another GPU to create tables sorted by price and performance. It is worth noting that due to driver, OpenCL, and/or OpenMM implementation AMD GPUs seem to perform lower than NVIDIA GPUs at equivalent peak TFLOPS; hence my decision to keep these tables separate.
NVIDIA:


AMD:


What surprised me with the AMD table is that all of the Navi GPUs were eliminated!

Please let me know if you see any errors in my methodology. If there aren't any error, I think this is about as good as I can do from theoretical standpoint and I'll focus on gathering real-world data which should supersede these tables in the future.

NoMoreQuarantine wrote:In an effort to make a predictive ranking of GPU performance, I took the specs of all of the desktop GPUs since 2016 for both AMD and NVIDIA ...

I think you should at least go back one more generation. Several of the Nvidia 900 series cards would fall toward the middle of your list, and can be found cheap on ebay.

Kebast wrote:I think you should at least go back one more generation. Several of the Nvidia 900 series cards would fall toward the middle of your list, and can be found cheap on ebay.

I was afraid someone might ask for that Alright, I'll do that tomorrow since I have the day off.