i wouldnt worry about 1 or 2 invalids, especially with such a small sample size (really can't make too many conclusions about just one or two invalids).
no app or device is capable of 100% valid processing. there will be invalids occasionally on even the most stable of devices or applications. the einstein validator also seems fairly strict, so similar devices are more likely to validate against each other than to an outlier due to ever so slight differences in final result. this has been seen in many cases with einstein as a whole across many of their searches/applications/devices, so what you're seeing isnt without precedent.
Yeah I get them here and there on my RTX 3060 system (processing GPUGRID at the moment). Which has an EPYC 7443P and the first EPYC system I’ve seen to have this problem.
didn’t you have a theory that it was slower GPUs that had the problem? What Turing GPUs exactly are in that “daily driver” system?
The only other Turing gpus are the two 2080 Ti's you sold me that are in the Epyc hosts.
I've had only a couple of the flushing cache errors on the 2080 Ti's though. None showing currently as the errors are dominated by the rash of bad tasks the project generated lately in the past few days with permanent download errors.
I used to get the flushing cache errors predominately on the 5950X hosts with 2080 Ti or 2070 Super cards.
Now down to only 3 2080's and 2 2080 Ti cards cumulative in my hosts.
A tad late to the party, but every bit as impressed with this on my 1660 Ti as I was with Petri's SETI app which I finally installed only a few months prior to SETI halting new work.
With the stock app, 4021 tasks take approximately 450 seconds when the system is otherwise mostly idle. With V0.95 of the special sauce (if that term is over played from SETI days, please kick me back in bounds) WUs run approximately 355 seconds - a reduction of a little over 20%.
HUGELY impressed.
Will run for a few days and then install V1.0 and see if that makes any additional difference, understanding this is among the lower performing cards supported by the app.
Thanks to all who provided the installation instructions. Worked like a charm.
I just read the whole thread, I haven't yet tried this app, but I have one major concern: If this is an app that was written outside E@H, and differs so massively in calculation time, how can one be sure that it delivers the right scientific results? I'm here for the science and I would not want to deliver wrong results and skew the scientific outcome because of an urge to calculate faster.
Checking my hosts
)
Checking my hosts results: https://einsteinathome.org/workunit/648154078
stock iGPU vs 2 NV3070 with Petri's app.
iGPU considered invalid one. Known issue?
i wouldnt worry about 1 or 2
)
i wouldnt worry about 1 or 2 invalids, especially with such a small sample size (really can't make too many conclusions about just one or two invalids).
here's an example of one of my systems validating against another iGPU just fine: https://einsteinathome.org/workunit/647036007
no app or device is capable of 100% valid processing. there will be invalids occasionally on even the most stable of devices or applications. the einstein validator also seems fairly strict, so similar devices are more likely to validate against each other than to an outlier due to ever so slight differences in final result. this has been seen in many cases with einstein as a whole across many of their searches/applications/devices, so what you're seeing isnt without precedent.
_________________________________________________________________________
Wish I could figure out why I
)
Wish I could figure out why I continue to get errors on my Turing cards only with the Einstein app.
This daily driver continues to get the flushing cache errors. I've tried everything under the sun to determine the cause and eliminate the errors.
The only solution I've found is to move to Ampere cards where the error has never occurred.
Yeah I get them here and
)
Yeah I get them here and there on my RTX 3060 system (processing GPUGRID at the moment). Which has an EPYC 7443P and the first EPYC system I’ve seen to have this problem.
didn’t you have a theory that it was slower GPUs that had the problem? What Turing GPUs exactly are in that “daily driver” system?
_________________________________________________________________________
The daily driver has three
)
The daily driver has three 2080 EVGA hybrid gpus.
The only other Turing gpus are the two 2080 Ti's you sold me that are in the Epyc hosts.
I've had only a couple of the flushing cache errors on the 2080 Ti's though. None showing currently as the errors are dominated by the rash of bad tasks the project generated lately in the past few days with permanent download errors.
I used to get the flushing cache errors predominately on the 5950X hosts with 2080 Ti or 2070 Super cards.
Now down to only 3 2080's and 2 2080 Ti cards cumulative in my hosts.
Yeah weird. My 7-GPU system
)
Yeah weird. My 7-GPU system (EPYC 7402P) has 3x 2080Ti and no errors (no flushing errors)
_________________________________________________________________________
A tad late to the party, but
)
A tad late to the party, but every bit as impressed with this on my 1660 Ti as I was with Petri's SETI app which I finally installed only a few months prior to SETI halting new work.
With the stock app, 4021 tasks take approximately 450 seconds when the system is otherwise mostly idle. With V0.95 of the special sauce (if that term is over played from SETI days, please kick me back in bounds) WUs run approximately 355 seconds - a reduction of a little over 20%.
HUGELY impressed.
Will run for a few days and then install V1.0 and see if that makes any additional difference, understanding this is among the lower performing cards supported by the app.
Thanks to all who provided the installation instructions. Worked like a charm.
Regards,
Fred
No the "special sauce"
)
No the "special sauce" moniker is still applicable and in use by GPUUG team members.
Happy crunching and your welcome for your holiday present.
Thank you Keith.
)
Thank you Keith.
I just read the whole thread,
)
I just read the whole thread, I haven't yet tried this app, but I have one major concern: If this is an app that was written outside E@H, and differs so massively in calculation time, how can one be sure that it delivers the right scientific results? I'm here for the science and I would not want to deliver wrong results and skew the scientific outcome because of an urge to calculate faster.
my blog about raspberry pis and diy life