hi, depending on how much memory your driver allows to allocate at once, dt will happily use whatever additional memory you have. that's always a plus, if you run out we revert to tiling (which means multiple passes, will be slow).
if you run darktable -d opencl, it will output stuff like [opencl_init] device 0 `GeForce GTX 670' has sm_20 support. [opencl_init] device 0 `GeForce GTX 670' supports image sizes of 32768 x 32768 [opencl_init] device 0 `GeForce GTX 670' allows GPU memory allocations of up to 1023MB which is important to us. sm_20: we use some inline ptx assembly for nvidia cards to support atomic floats in global memory. that should really be standard today if you get an nvidia card. image size: many of our buffers are images/textures. so the bigger the better i guess, depending on how crazy you go with your images. allocations: we can't usually alloc more than that in one buffer, even if the card has more memory in total (mine from the example above has 4G). that's still plenty here for interactive work, the equalizer for one will rather allocate 7 wavelet scale buffers @screen res than one gigantic one. hth, jo On Mon, Jan 16, 2017 at 12:04 PM, Christian Kanzian <christian.kanz...@gmx.at> wrote: > Hi, > > Am 2017-01-16 11:51, schrieb Michael Below: >> >> Hi, >> >> I think that line about AMD not being supported at all by current >> Linux distros is a bit misleading, AFAIK you need to install the >> AMDGPU-PRO driver available on their homepage (for Ubuntu and RH). >> NVidia offers a more streamlined installation process for a wider >> range of distros. > > > That's correct! As I've written, it seems that AMD drops support for older > models earlier. > I only want to point out that one has to check driver support carefully. > >> There was one previous report on the list about an OpenCL GPU being >> slower than a CPU, I think that involved an old NVidia card paired >> with a new i7 CPU. For myself, OpenCL does speed things up >> considerably (old AMD Phenom II X4 810 CPU, Nvidia 750Ti card). > > > There were a number of reports and discussions in conjunction with Nvidia > cards. > Some operations or better moduls like Equalizer and profiled denoise took > way longer > on GPU than CPU. In my case it was a Nvidia GT640 vs. i7-2600. I bought the > card for > darktable only. Different OpenCL settings didn't helped. I also had a > slugish ton curve. > > These are just personal experiences. > >> About the memory issue: I think darktable can use multiple buffers >> inside OpenCL. >> >> Cheers >> Michael > > > Christian > > >> >> -----Ursprüngliche Nachricht----- >> Von: dt-l...@stefan-klinger.de [mailto:dt-l...@stefan-klinger.de] >> Gesendet: Montag, 16. Januar 2017 11:32 >> An: darktable-user@lists.darktable.org >> Betreff: Re: [darktable-user] what graphics card to get? >> >> Here's a summary of the information gathered so far. It's not really >> exhaustive! >> >> Patrick Rudin: >> >> * Suggests NVidia 1050, 1050Ti or 1060. >> >> * AMD seems not to be supported *at all* by current Linux distros. >> >> * Would also like to know about memory requirement. >> >> >> Christian Kanzian: >> >> * cheap card can be slower than the CPU. [contradicts Guillermo >> Rozas below] >> >> * AMD Radeon R9 270X supported under Debian 8, but not any more >> under Debian 9, since AMD stopped support in proprietary driver. >> Planning to turn to NVidia. >> >> Guillermo Rozas: >> >> * runs DT on an intel i7-4720HQ notebook with an Nvidia 960M 4GB, >> under Ubuntu 16.10. CUDA is fully supported by the proprietary >> drivers, and DT uses the GPU without issues. >> >> * any supported GPU will give massive performance improvement. >> [contradicts Christian Kanzian above] >> >> * get as much memory as possible [contradicts following point] >> >> * DT cannot use buffers larger than 1.5GB [contradicts previous point] >> >> David Vincent-Jones: >> >> * There seems to be a problem with color management and NVidia. >> Using a more current version of DisplayCAL/ArgyllCMS seems to >> solve the problem. >> >> Maybe the devs want to add some more information? >> >> That was my original question: >> >> On 2017-Jan-13, Stefan Klinger wrote with possible deletions: >>> >>> Hello, >>> >>> what graphics card to get for using Darktable with OpenCL under Linux? >>> >>> I already have an Asus M5A97 evo R2 mainboard, with an PCIe 2.0 x16 >>> slot for the graphics card. I want to stay in the price range below >>> 200€. I do not play games, nor do I use 3D graphics software of any >>> kind. Darktable will be the most demanding task for this piece of >>> hardware. >>> >>> The Darktable manual mentions a minimum of 1GB of graphics memory to >>> be sufficient [1]. Is this up-to-date information? Does DT benefit >>> from more more? How much more? >>> >>> There seem to be different versions of OpenCL [2]. Is that an issue >>> when deciding which card to buy? What does Darktable need here? >>> >>> NVidia graphics cards seem to differ in “compute ability”, which seems >>> to be the CUDA version [3] they support. What is required here for >>> Darktable? Is there something similar to look out for with AMD Radeon >>> cards? >>> >>> The manual [4] mentions “recent” AMD Radeon HD7xxx to work out of the >>> box. According to Wikipedia [5] that series was introduced in 2011 >>> and had its latest update in 2013. Also, it seems to be difficult to >>> buy any of these now (I don't want to buy used hardware). Is there >>> more current information available? >>> >>> Does anyone have information on / experience with Radeon R7 or similar >>> (i.e., GCN family of cards [6])? They seem to be available for under >>> 100€ currently [7]. >>> >>> A list of NVidia GPUs with CUDA-support (I guess that's required for >>> OpenCL on NVidia) is here [10]. >>> >>> Most graphics cards I've seen are not actually built by AMD or NVidia, >>> but rather by 3rd party manufacturers (MSI, ASUS, Palit, Zotac, >>> Gainward, Gigabyte, PNY, ...) using their GPUs. Is there any >>> particularly good / bad one to choose? >>> >>> What else to pay attention to? >>> >>> Thanks for helping >>> Stefan >>> >>> >>> BTW: The link “darktable user mailing list” [8] on [9] is outdated. >>> >>> ____________________ >>> [1] http://www.darktable.org/usermanual/ch10s02s03.html.php >>> [2] https://en.wikipedia.org/wiki/OpenCL >>> [3] >>> https://en.wikipedia.org/wiki/CUDA#Version_features_and_specifications >>> [4] http://www.darktable.org/usermanual/ch10s02s06.html.php >>> [5] https://en.wikipedia.org/wiki/Radeon_HD_7000_Series >>> [6] https://en.wikipedia.org/wiki/Radeon#Graphics_Core_Next-family >>> [7] https://www.heise.de/preisvergleich/?fs=Radeon%20R7&cat=gra16_512 >>> [8] >>> http://www.darktable.org/resources/darktable-user+subscr...@lists.dark >>> table.org [9] http://www.darktable.org/resources/ >>> [10] https://developer.nvidia.com/cuda-gpus > > ____________________________________________________________________________ > darktable user mailing list > to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org > ____________________________________________________________________________ darktable user mailing list to unsubscribe send a mail to darktable-user+unsubscr...@lists.darktable.org
