Support for ATI Radeon GPUs
Message boards :
Wish list :
Support for ATI Radeon GPUs
Message board moderation
Author | Message |
---|---|
Send message Joined: 1 Jan 14 Posts: 302 Credit: 32,671,868 RAC: 0 |
We have now two volunteers who try to make OpenCL app for AMD cards. Next I'm so busy at work for now. If both of them failed to do OpenCL app, I will do one. That is GREAT NEWS!!!! |
Send message Joined: 24 Aug 13 Posts: 112 Credit: 31,804,229 RAC: 2,023 |
We have now two volunteers who try to make OpenCL app for AMD cards. Next I'm so busy at work for now. If both of them failed to do OpenCL app, I will do one. Good to hear :) Teach him to program OpenCL ;) Team AnandTech - SETI@H, Muon1 DPAD, Folding@H, MilkyWay@H, Asteroids@H, LHC@H, POGS, Rosetta@H, Einstein@H,DHPE & CPDN Main rig - Ryzen 3600, 32GB DDR4 3200, RX 580 8GB, Win10 2nd rig - i7 4930k @4.1 GHz, 16GB DDR3 1866, HD 7870 XT 3GB(DS), Win7 |
Send message Joined: 23 Jan 13 Posts: 8 Credit: 13,121 RAC: 0 |
|
Send message Joined: 23 Jan 13 Posts: 8 Credit: 13,121 RAC: 0 |
Last modified: 28 Oct 2014, 17:32:53 UTC |
Send message Joined: 28 Apr 13 Posts: 87 Credit: 26,719,619 RAC: 186 |
And AMD, finally, starts to publish OpenCL 2.0 examples,....waiting for SDK Does that mean you're about to write an openCL app which should run on all platforms (including Intel with sw dp support and arm)? If yes, pm me your address, I'll send you a present when it's up and running ... Alexander |
Send message Joined: 26 Sep 13 Posts: 33 Credit: 6,274,255 RAC: 8,014 |
Last modified: 31 Oct 2014, 14:21:28 UTC Does OpenCL support start cheaper and earlier on the vast majority of ATI offerings? The ATI m7740 is $60 used and has OpenCL 1.0 with double precision and was released in 2009, for example. If so, then this might bring a lot of older hardware back online for the A@H project. My particular case has me steering to the ATI for price reasons. My Dell m6500 has a dying nVidia 2800m, which doesn't have double precision and not supported on A@H, so I'm looking for a replacement. Dell shipped the m6500 with two compatible GPU's to this project. The ATI m7820 has OpenCL 1.0 support and double precision instructions and runs $100 used, while the only Kepler nVidia card (Quadro 5000m) Dell shipped on the m6500 is $400 used. The Dell m6400 ATI m7740 is $60 used and could work in the m6500 slots and was OpenCL 1.0 in 2009 with double precision. I'd love to have ATI support after looking at the pricing options on the GPU's that are assured to work in the m6500. |
Send message Joined: 29 Jul 14 Posts: 6 Credit: 3,115,200 RAC: 0 |
POEM@home develop a new openCl application thanks to a community member. the new application seems to be very fast and can be tested by user right now. http://boinc.fzk.de/poem/ in the news published on their web site you can find also the name of the community member/developer |
Send message Joined: 1 Jan 14 Posts: 302 Credit: 32,671,868 RAC: 0 |
Last modified: 4 Dec 2014, 12:45:03 UTC POEM@home develop a new openCl application thanks to a community member. WOO HOO!!! That's GREAT news!! |
Send message Joined: 28 Apr 13 Posts: 87 Credit: 26,719,619 RAC: 186 |
We have now two volunteers who try to make OpenCL app for AMD cards. Next I'm so busy at work for now. If both of them failed to do OpenCL app, I will do one. Poem has released a new OpenCL app and has work available 7/24. GPUGRID has a beta app for AMD cards. If you wait too long all gpu's may have found their projects and noone will be here to crunch that ... |
Send message Joined: 15 Feb 14 Posts: 20 Credit: 8,563,792 RAC: 0 |
Last modified: 9 Feb 2015, 17:50:31 UTC One thing I found out is that AMD only guarantees its cards' accuracy to be equal to the OpenCL specification. The problem with that definition is that single precision floating point divides and square roots are not guaranteed to conform to IEEE 754 specifications according to the OpenCL specifications listed at https://www.khronos.org/registry/cl/ unless the OpenCL version used is at least 1.2, the GPU being targeted is capable of correctly performing divides and square roots according to the IEEE 754 specification, and the compiler flag "-cl-fp32-correctly-rounded-divide-sqrt" is passed to the compiler. (The flag does not exist in OpenCL 1.0 or 1.1.) This last flag forces the GPU to correctly perform divides and square roots if it is capable of performing those operations correctly, and terminate the program before attempting to run it if the GPU is incapable of doing correct IEEE 754 divides or square roots. Without the flag, a GPU is allowed to be off by up to 2.5 units in the last place for divides and up to 3 units in the last place for square roots in the standard profile. This inaccuracy can be crucial to graphics speed, but is not welcome in many scientific computations where accuracy is more important than speed. This matter is moot if your program does not use single precision divides or square roots. These operations are always correctly performed in double precision if double precision math is supported.
|
Send message Joined: 28 Apr 13 Posts: 87 Credit: 26,719,619 RAC: 186 |
OK, might all be right, but: Milkyway, GPUGRID, Einstein, Poem, Seti, Collatz and some more do have fine working AMD gpu apps. And in addition, as you posted, these problems do not occur when running double precision. And AFAIK Asteroids is a double precision project. AMD is well known that they are faster with DP than nVidia. If you do not believe, look into the top hosts list @ Milkyway, a DP project. Just to give you an impression: the R9 280x is rated SP 4178 GFLOPS, DP 1044 GFLOPS. A GTX970 is rated SP 3494 GFLOPS and DP 146 GFLOPS !!! If the project does not need that crunching power - fine, more gpu's left for other projects. Speaking for myself, I'm happy with the cpu apps! |
Send message Joined: 15 Feb 14 Posts: 20 Credit: 8,563,792 RAC: 0 |
Last modified: 10 Feb 2015, 4:43:39 UTC I was not trying to disparage the AMD or ATI GPUs, but was providing a heads-up because other projects had developers who complained about buggy AMD OpenCL drivers which might be solved by passing the correct OpenCL compiler flags. Nvidia's GPUs from Fermi onwards default to IEEE 754 when running CUDA applications and degrade to gaming precision only if the programmer requests that. (Pre-Fermi Nvidia GPUs are stuck at gaming precision and cannot become IEEE 754 compliant in single precision.) Personally, I think that OpenCL should have defaulted to IEEE 754 compliance and should require developers to explicitly request gaming accuracy to get faster but less accurate results instead of requiring programmers to explicitly request IEEE 754 accuracy to get it, and feel that this is a fault of OpenCL standard instead of AMD. If I had written the standard, such flags would be required to run on hardware that is incapable of IEEE 754 compliance so developers have no illusions on what they are programming. |
Send message Joined: 28 Apr 13 Posts: 87 Credit: 26,719,619 RAC: 186 |
I was not trying to disparage the AMD or ATI GPUs, but was providing a heads-up because other projects had developers who complained about buggy AMD OpenCL drivers which might be solved by passing the correct OpenCL compiler flags. Nvidia's GPUs from Fermi onwards default to IEEE 754 when running CUDA applications and degrade to gaming precision only if the programmer requests that. (Pre-Fermi Nvidia GPUs are stuck at gaming precision and cannot become IEEE 754 compliant in single precision.) Hi Jesse, very interesting info, indeed! Reading statements like your's is one reason why I'm crunching different project, it's much much better than reading newspapers. I would like to point to a discussion @ GPUGRID, http://www.gpugrid.net/forum_thread.php?id=3922&nowrap=true#40152 about the new nVidia cards. Alexander |
Send message Joined: 29 Jul 14 Posts: 6 Credit: 3,115,200 RAC: 0 |
I am using my old ATI 280 on Milkyway@Home, at the moment. Milkyway@Home is a DP OpenCl project (as A@H). My old ATI 280 (bought for 98€) is still faster than any Nvidia card on that project. (10 time faster than a GTX760!) Because less computational time means less power used, I think that would be important to develop A@H also for ATI card, if you think that the carbon footprint of your project is as important as your results. best regards |
Send message Joined: 25 Mar 16 Posts: 5 Credit: 4,501,920 RAC: 0 |
|
Send message Joined: 12 Jan 13 Posts: 58 Credit: 13,655,270 RAC: 1,995 |
AMD bought ATI .... Here is a URL for AMD Driver Support https://www.amd.com/en/support Bill F In October 1969 I took an oath to support and defend the Constitution of the United States against all enemies, foreign and domestic; There was no expiration date. |
Send message Joined: 19 Sep 14 Posts: 13 Credit: 10,026,859 RAC: 0 |
|
Send message Joined: 25 Mar 16 Posts: 5 Credit: 4,501,920 RAC: 0 |
Last modified: 5 Sep 2020, 10:16:15 UTC Hello, I need to know the execution code of the tasks to convert them into Open CL, and be processed by AMD or Ati graphics cards. I would also like to use CUDA's Tensor flow, I think that calculations can be accelerated with new technologies. I would also like to adapt the code for old gtx560 graphics cards. Since they do not use the full clock frequency of the GPU and shaders of the card and it takes a long time to execute, since they have become obsolete, and it is not fair. I am a Python programmer, SciPy, NumPy ... Thanks |
Send message Joined: 13 Apr 20 Posts: 12 Credit: 27,678,205 RAC: 0 |
|
Send message Joined: 23 Oct 12 Posts: 18 Credit: 116,663 RAC: 79 |
|
Message boards :
Wish list :
Support for ATI Radeon GPUs