BRP6 Cuda 55
7 Sep 2016 17:23:20 UTC
Topic 200912
(moderation:
I have a GTX660 which gained a solid 19% improvement in output over the stock Cuda 32.
A month ago I switched a small Fermi card, GT430, over to Einstein. This card seems to be about 2% slower with Cuda 55 than Cuda 32.
What to do? Change the driver to the same one the GTX660 has, revert back to Cuda32 or let it run because it is helping the developers refine the Beta app.
The developers aren't really
)
The developers aren't really needing any more input on the Beta apps. They been stable for over a year. Not even sure why they are considered Beta apps even. They should have been moved to stock a year ago. The GT430 card is a very low performance card and doesn't have the hardware to really utilize the Beta app. I would switch it back to CUDA32 which was developed in the same timeframe as the release of the GT430 product. It would match better to the card's capabilities.
Thanx, I'll burn thru my 2
)
Thanx, I'll burn thru my 2 day cache of Cuda 55's and revert to stock unless someone else comes up with another idea.
I think the observations here
)
I think the observations here have been that CUDA55 benefit relative to CUDA32 is very substantial on cards of the generations about three back from where we are now, but not beneficial before that.
Specifically, Kepler and later--big win.
Fermi and earlier--don't bother.
Have we heard anything from
)
Have we heard anything from the Einstein staff regarding the memory leak issue with Pascal cards running the cuda55 beta of BRP6 on W10 machines? A number of tasks went bad for me last night due to the memory running out faster than I had anticipated. Not the first time this has happened, unfortunately. I'll have to change the reboot schedule on that machine.
Are we at the crossroads now?
)
Are we at the crossroads now? Server status looks like there's no more work to send for BRP6. Interesting times ahead