No idea what this means for folding, until there is an actual product to test. My suggestion, wait to see what it actually offers.

notfordman

Scarlet-Tech
No idea what this means for folding, until there is an actual product to test. My suggestion, wait to see what it actually offers.

^^+1
There was a time when AMD cards were better folders, seems like a couple years ago. 

Yep, the Pascal GPU changed that for the Most Part.
My GTX 1080 Cards runs circles around my AMD R9 290X Cards.

Probably safe to assume similar to RX 480 as the architecture is similar. Just higher like a 1070 to a 1080 difference. As to that half precision stuff, would have to depend on the project and would need the developer to target it specifically. Most crunching apps want higher precision not lower. Most AVX and higher allow similar packing schemes.

https://foldingforum.org/viewtopic.php?f=24&t=27749
 
ah, interesting to see that they did move to OpenMM with their new "core X" series, which seems to be something similar to OpenCL. there goes that reason for why AMD has higher compute scores than nvidia but have worse F@H performance. 

So Vega is released this week.  So who wants to sell me their 1080ti if Vega beats them :P haha.  Apparently they are priced insanely high though 

Well I looked at the dropbox for folding, and apparently the Vega Frontier sometimes beats the 1080, sometimes gets destroyed by a 950..http://ppd.fahmm.net/  Which has to be wrong, like so wrong.
 

chap34
So I don't know the specifics of how folding works, other than it works better in GPU than CPU. After watching the stream on Vega today, a new technology got me wondering if it will be a folding game changer.
 
"Rapid Packed Math will increase the calculations on Vega GPU. The demo was using TressFX benchmark where Vega would offer more than half a million more strands (of hair) per second. This is available through half-precision (FP16) shaders."
 
Thoughts?

Foldin@Home works with FP32 and FP64, I don't work with molecular dynamics but I believe FP16 are completely useless for that kind of workload

notfordman

Scarlet-Tech
No idea what this means for folding, until there is an actual product to test. My suggestion, wait to see what it actually offers.

^^+1
There was a time when AMD cards were better folders, seems like a couple years ago.

When? I can't really remember any AMD GPU that was faster than NVIDIA one in Folding@Home

Chris21010
even though AMD has always had much better compute performance over NVIDIA their cards are also hampered by the fact that F@H is coded around CUDA, an NVIDIA proprietary code. so until stanford creates code around an AMD API you'll always have the overhead losses of transcripting NVIDIA code to run on AMD hardware. that is why i have avoided their cards for folding as they just aren't nearly as efficient per watt.

The compute performance advantage AMD GPUs had over NVIDIA is just on paper, for example AMD claimed the HD 4870 delivered over a TFLOP which is the same value NVIDIA declared for the GTX 480/580 which, instead, are much faster than any HD 4000 series GPUs and on par or faster than the HD5870 that is supposed to deliver almost 3 TFLOP (same as GTX 680).
Simply speaking, AMD GPU have a resource occupancy deficit so while theoretically their GPUs should deliver a certain performance in a real world scenario they fail to reach that value as some of their compute unit can't be fully utilized.
CUDA is not the reason since F@H also uses OpenCL and OpenMM.

Tahiti might have been faster at folding than same-era NVidia cards *IF* the specific project needed heavy FP64 performance - the R9 280/280x were very very good at FP64, have only been beat by VERY HIGH COST workstation-specific cards since then and not all of THOSE.
 
 I suspect Folding uses a lot more FP32 though, and Tahiti wasn't particularly good at those ops.
 

Ranmacanada
Well I looked at the dropbox for folding, and apparently the Vega Frontier sometimes beats the 1080, sometimes gets destroyed by a 950..http://ppd.fahmm.net/  Which has to be wrong, like so wrong.
 

 Depending on the specific work unit, I've seen my 1080ti cards sometimes beat BADLY by the 1070.

Stefem

QuintLeo
 Depending on the specific work unit, I've seen my 1080ti cards sometimes beat BADLY by the 1070.

That's because they aren't processing the same work unit, the GTX 1080 Ti will be always faster on the same WU.

 Perhaps not the same work unit, but on certain projects the 1070 IS consistantly providing better PPD than the 1080ti and a MUCH lower frame time - 9414 is particularly bad that way and I've seen a LOT of those particular work units.
 
 I suspect either a driver issue or some sort of issue caused by the higher latency of GDDR5x vs GDDR5, as the 1080 cards I have also don't work well on those particular projects.
 
 Could also be something to do with the Intel CPU in that machine getting way overloaded, as the 1080ti cards I have are the only ones in an Intel-based machine (it IS a dual-core, the AMD-based machines are all at least quad core).
 
 I should have more data over the next week, as I plan to move at least one of those 1080ti cards into my "next build" machine that will have an octocore FX CPU in it as part of my "rationalization" work.

Vega is surprisingly good for folding. In my case I have the card undervolted and underclocked.
 
System setup:
GPU: RX Vega 64 Liquid
Core: 1552MHz (stock: 1750MHz)
HBM: 1050MHz (stock: 945MHz)
Voltage: 0.980v (stock: 1.200v)
Driver: 17.9.1
OS: Win10 64-bit
 
This is one of the better WUs for Vega at ~950K PPD. I have also seen some that do ~650K PPD and ~700K PPD.

 
The incredible thing is how low the folding power consumption is compared to gaming:

In games, I see the "GPU only Power Draw" reporting 180-195W with the same settings. (when GPU limited, no FPS cap or VSYNC) The slower WUs use a bit less, but I didn't save the exact numbers.
 
Keep in mind that AMD cards report power as vcore VRM output, not full card power like Nvidia. The system is pulling 275W from the wall, and I estimate that the full card is using ~180W. That's roughly equal to the GTX 1080's TDP.
For comparison, my titan X gets 1,360K PPD at around 220-230W power (2038MHz/1.05v) on P11409.

Chris21010
that is actually better than i thought but the 1080ti/titan Xp still edge it out on the PPD/Watt stat by just over 1K.

Yeah, I didn't expect Vega to dethrone Pascal in performance/watt.
 
I just got a P11431 on the Vega system, and it's even better:

1,131,405 PPD, using only slightly more power than P11409 at 140W for GPU core.
 
The thing that stood out to me is that it's more than twice as efficient (PPD/W) as Polaris, which is also 14nm. (RX 480/470/580/570)
RX 580 (1275MHz/1.00v) on P11409 gives ~350K PPD with 105W GPU power. (~140W full card power)

jedi95I just got a P11431 on the Vega system, and it's even better:

1,131,405 PPD, using only slightly more power than P11409 at 140W for GPU core.

Don't get too happy with that 11431 performance... Run 5 & Run 7 give the highest output for project 11431 by about 20 - 30% over the other runs of 11431 work units, at least on nVidia cards.

That is still pretty fantastic.  Specially with such immature drivers.  Hopefully as they mature Vega can compete, and come down in price.  Right now I am just sick and tired of getting nothing but 9414 and 9415 on my 1080Ti.  My card is being starved :( It can do so much more and yet it's getting no real work units.  I blame Vorsholk and his folding farm :P haha.  

bcavnaugh
I see 1222166 (8.64 WUs) PPD on P11431 ATM Link

And on a run 8 work unit to boot, nice!  I get 332,480 (+/- 1,000ppd) on run 8's from my 970.  


RE: Vega, nice to see AMD getting back into the game, strong competition is generally a good thing for consumers!!!