On mercredi 6 décembre 2017 08:02:52 CET Michael Rasmussen wrote: > On Wed, 06 Dec 2017 07:38:59 +0100 > > Remco Viëtor <[email protected]> wrote: > > https://www.phoronix.com/scan.php?page=article&item=darktable-opencl-gpu&n > > um=1 might be interesting as well. And there the 1050(TI) doesn't do that > > well, you'd have to go to at least a 1060 (which I did, and it works very > > well, even things like equalizer or profiled denoise are virtually > > instantanous) > You also need to factor in the CPU. To gain the advantage of the most > expensive and superior GPU's you most also have the most expensive and > superior CPU's (i7 or Ryzen 7) so for most people with ordinary > equipment a 1050ti will max out their equipment's potential. Of course the rest of the system is important as well. But the 1060 is not top of the line, it's one step up from the 1050. And OP didn't give any more details of his system.
> You should also remember that most of the benchmarks you read on the > net is gaming benchmarks which differs the workload of image processing > to a great extend. And that's why I posted the Phoronix link, which at least has the advantage to be close to the workload considered here. > Reading eg. the conclusion from this benchmark > (https://www.pugetsystems.com/labs/articles/Photoshop-CC-2017-NVIDIA-GeForce > -GPU-Performance-899/) shows there is little gain given after GTX 1050 and > comparing this to the price difference you surely get the best price > performance ration by sticking with GTX 1050. Rather surprising results there. I noticed a remark about CUDA in one of the comments, and that's not used by Darktable. And the 1050 and 1050 Ti might have a problem with OpenCL (given the results shown). That could be a driver issue (Phoronix used version 375.10, current is 384.98 or higher, note that you will need the proprietary drivers to use OpenCL or CUDA). Remco ____________________________________________________________________________ darktable user mailing list to unsubscribe send a mail to [email protected]
