Einstein@home has historically been pretty poor for Nvidia performance in comparison to AMD cards. I've always had a hunch that there is some poor optimization or some error in the coding for their app that artificially restricted performance on Nvidia. that turned out to be true.
 
for the last few months, I've been helping another user (petri33) develop and test some modifications to the Einstein@home Gamma Ray application for fixing the issues with the nvidia app. making a handful of rather simple changes, we boosted nvidia performance massively for Pascal/Turing/Ampere GPUs. Pascal cards see ~40-60% speedup, Turing sees ~50-65% speedup, and Ampere is over 100% faster (2x!), Maxwell and older GPUs saw no benefit. These were my observations on a Linux platform. while the performance is amazing, the method of our implementation was not very user friendly, and Linux-only. (I talked a few users through it, and it's not a good experience to the less tech-savvy).
 
we shared our findings with the project developers, and they were pretty receptive with making the changes to their app. they have now released a beta app v1.28 for the Gamma Ray tasks (plan class: FGRPopencl2Pup-nvidia) for all Windows, Linux, and Mac systems. It has the requirement that you must have OpenCL 2.0 or greater drivers (you need 465 or 470 branch nvidia drivers), and you must have a Pascal or newer GPU. users on both Windows and Linux have reported that the beta tasks are working as expected.
 
so install new drivers if you havent already, and enable beta tasks in the project preferences, and enjoy a massive boost in processing speed and associate boost to your RAC.
 
Gamma Ray tasks only. nothing for Gravitational Wave tasks... yet.
 
:)

actually it looks like they just released it from beta. so anyone with a Pascal and newer and 465 and newer drivers should be able to get this app now.

thanks for all the hard work to get this improvement to be implimented

Got some running as well.  Nice to see performance improvements

bill1024
 
I did have run BETA tasks and the tasks that failed were ver .28
If the new tasks are out of BETA and being sent out as normal working tasks, is the VII not going to work for 
Einstein anymore? That would be a shame, that GPU does very well in Einstein.
I sure hope there is a fix for this issue.
 

the AMD app is still in beta.
only the Nvidia app is out of beta and in production. 
 
several other RadeonVII users reported either errors, or no change in performance, or worse performance. I do not have any Vega or Vega20 cards that I can test. but I have a feeling that most issues are due to poor driver support for the Radeon VII. the new app requires certain features from OpenCL 2.0 and was compiled with OpenCL 2.0 target. so if the driver doesnt have proper support, it will produce errors. My RX 570 under Linux works ok with the 1.28 beta tasks, no errors, but performs marginally slower than the production application (which matches my testing with pre-release code on Polaris).
 
turn off beta tasks if you're getting errors on your AMD systems and you can't get newer drivers installed. At least one other user reported that it works on their Radeon VII (under Linux), but ran 10% slower than the production app. I haven't seen any success stories with Vega or older yet. AMD driver support is garbage, and especially bad for the Radeon VII it seems since it was dead on arrival and discontinued after like 6 months.
 
In pre-release testing, we did see fairly good gains with Navi GPUs, with roughly 20% speedup on an RX5700, but since this is an EVGA forum, I figured most users would be using Nvidia GPUs. the code change was aimed at nvidia primarily since they were the most affected. AMD cards didn't have nearly the speed issues that nvidia did. Navi on the 1.28 app seems to only be 3-4% faster than production, and I've contacted the devs to see what they did and if they missed some compiler optimization flags.
 
other projects work fine because all other projects are using legacy OpenCL 1.2 features which are well supported by nearly all drivers. the new features in this new app however require opencl 2.0 support. Nvidia has a much better handle on that than AMD it seems.

hey bill, are you out of Einstein jail yet?
 
I confirmed the issues under Windows with my RX570. I tried all three latest driver package, 21.8.2, 21.6.1, and even the pro/enterprise driver 21.Q2.1. all three gave the same error after ~10s. So you'll probably still see errors with Windows/Vega.
 
pretty sure it's a problem with the windows drivers with these older GPUs (since it seems to work with Windows/Navi)

oh OK, I meant specifically your Radeon VII Windows system that got errors yesterday on the beta app. I was wondering if the new drivers helped or if you still got errors. WITH the beta app. I know the normal app works fine. when you enable beta, just make sure to watch it and stop processing and/or abort the beta tasks to avoid crashing your cache again.
 
BTW, don't run multiple tasks at a time on nvidia. they just run slower than 1x. it provides a benefit with the AMD cards though.
 
I'm testing some new optimizations. not available from the project yet :)
 
3060Ti = 210 sec @190W
2080Ti = 190 sec @225W
3080Ti = 109 sec @300W
 

They failed right away with the beta flag checked and new drivers.
 
Sure hope they keep it so the AMD cards work and not have that AMD beta go mainstream.
Even if they have a beta2 that we can choose to run these AMD cards.
It still smokes these tasks for a older GPU.