Actually, I am not an expert in this field. But, I recently started to get into GPU computing and gathered some information for myself. I'd like to take it more from the perspective of OpenCL vs. CUDA.

CUDA has been around a little longer than OpenCL and at first it definitely was superior to OpenCL. But now OpenCL had time to mature and in some comparisons it gets comparable performance to CUDA on the same hardware. Although OpenCL is portable between different graphics cards vendors and even CPUs you will always have to optimize for the targeted architecture if you want to squeeze out every last bit of performance. There is no single program/kernel that will be optimal for all architectures.

Here are the main differences between the different architectures: On CPUs OpenCL will use SSE instructions for parallel computation, on AMD GPUs each processor can handle 4 floating points in parallel similar to SSE, and on NVIDIA GPUs each processor will always just handle a single floating point in one instruction. So, on AMD hardware you should consider to use arrays of 4 floats to boost performance. This is an architectural advantage of AMD over NVIDIA, but it only helps if you are programming accordingly.

I don't have any real world experience with different hardware. So, I can't give you a real comparison between AMD and NVIDIA cards. I hope that my view on this is correct and I'd like others with more experience to correct me where I am wrong.

I use CUDA too, and have only limited experiences with OpenCL, but it is obvious that the two environments are very similar. OpenCL development is a bit more uncomfortable, but the basic concepts (threads, block, shared memory, etc.) are usually the same or very similar (or sometimes the same, but named differently).

So, if you have some CUDA experiences, you can start OpenCL development without any problems.

thanks a lot,

However it seems that few people use AMD solution for scientific computation

and I was wondering if there is a specific reason behind it.

Marco

I think that most people are using CUDA for historic reasons. NVIDIA were the first to deal with GPU computing. People first tried to use triangles and textures to do scientific computations on a GPU. NVIDIA supported that approach and developed a first programming language which would make it easier to write this kind of software. CUDA was very restricted at first, but due to the lack of competitors it got widely adopted.

AMD tried to invent their own stuff, but the closest they came to copy CUDA was Brooke and that never left Beta status. Instead, Apple started developing OpenCL which AMD later adopted. But that is exactly the problem: a lot of people don't like things that are developed by Apple. And it also took a while until OpenCL became adopted as an open standard. Then, it took a while until OpenCL had been implemented on different platforms than just Mac OS X. Initially, there was very little support. Then, OpenCL had also to catch up to CUDA, especially performance-wise.

I would guess that AMD only caught up to NVIDIA about 2 years, maybe 3 years ago. By that time CUDA has been established as a standard in research. Many people already knew how to use CUDA. If you want to be comparable to previous research you probably need a CUDA implementation of your algorithm anyway.

I'm using CUDA and have only looked at OpenCL briefly. From what I know, CUDA now supports some features that are not available in OpenCL (yet), .e.g. dynamic parallelism and recursion (depending on the hardware of course!). That said, I think there are also a lot more supplementary libraries for CUDA, e.g. the Thrust library that is a great help for GPU programming. Nvidia also offers tools for visual debugging and profiling.

I have been using OpenCL a lot in my research. I believe that OpenCL power comes mainly from supporting heterogeneous computing. The same OpenCL code can easily run in different compute resources i.e., CPU, GPU, APU, etc. without modification. However, if you care about performance, you have to adopt different optimization techniques for different compute resources.

Returning to AMD GPU, it offers peak performance close to (and sometimes more than) that offered by NVIDIA GPU. To get the best performance from AMD GPUs, you have to optimize your code using AMD-specific optimization strategies. Please take a look at this paper , https://www.researchgate.net/publication/229034386_Architecture-Aware_Mapping_and_Optimization_on_a_1600-Core_GPU?ev=prf_pub,

that discusses more details about the relative performance of AMD to NVIDIA GPU for large-scale, molecular modeling application along with the different optimization techniques to adopt for each device. In short, the authors got 371- fold speedup from AMD GPU compared to 328-fold speedup from NVIDIA GPU.

Article Architecture-Aware Mapping and Optimization on a 1600-Core GPU

Although I believe OpenCL might provide rather similar performance to CUDA on similar hardware, three important facts remain: a) CUDA is optimized for a particular platform and therefore is has all the optimizations it needs b) there are many libraries available for it (thrust) that make programmer's work easier c) most scientific work I know is done with CUDA (of course Marwa's work and answer is an exception and therefore I would value her opinion better than mine).

I also believe that with a 300-400$ NVIDIA card (GTX760-770 or even 660Ti - 670) you can still get decent performance improvement http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-660/performance.

So, I would suggest to stick with CUDA.

To summarize:

1) Since CUDA is tightly coupled with Nvidia GPU architecture you can gain more performance (in comparison with OpenCL on the same HW) and can use more HW features (in comparison with OpenCL)

2) There are more tools, libraries, books, guides for CUDA than for OpenCL

3) Nowadays supercomputers use GPUs as an accelerators. Majority of these GPUs are Tesla-based nodes. I have never heard about an large AMD-based GPU server.

4) In OpenCL due to its generality, the programming effort (overhead) is higher

5) For every architecture some performance optimizations are needed, but these optimizations are different for multicore CPU, Intel MIC, Nvidia GPU, and AMD GPU ! So, unfortunately an efficient (on all supported platforms) and straight-forward OpenCL code can't exist!

Dear Ivan and Alexandros

I get your point, I guess the question the relative cost of an NVIDIA solution could push the balance of this comparison in the other direction.

In consumer and prosumer solutions

KEPLER double precision performance is capped at 1/24 the single precision performance

MAXWELL double precision performance is capped at 1/32 the single precision performance

(the latter refers only to the new 750, so I can't be sure that applies to all)

AMD chips are not capped (usually it gives a rate between 1/2 and 1/4 the single precision performance). This makes AMD GPU double precision computing in consumer hardware INFINITELY faster than a comparable NVIDIA solution.

see http://www.tomshardware.co.uk/geforce-gtx-760-review-gk104,review-32717-21.html

Indeed this comparison changes in single precision, but this is not my case.

Marco

-Don't be lazy, Please rate the questions/answers you find useful-

@Ivan

I cannot entirely agree with your statements. And there are certainly other sides to the medal in many of the points you make.

1) Two or three years ago OpenCL finally caught up to CUDA performance-wise. Probably, CUDA is still faster on NVIDIA hardware, but it is only a very tiny percentage. In comparison to AMD's gaming hardware NVIDIA looses big time for double precision. But, actually this is the question that has been asked initially. Futhermore, OpenCL allows for extensions which then allows to access special hardware features if necessary. In the end both languages are equally powerful and the implementation of the respective drivers for CUDA and OpenCL decide about the actual performance.

2) There are different tools and libraries for CUDA. Most people only know the CUDA libraries; and quite frankly, these are very popular. But, with a quick search on Google I found the following two links:

http://comportability.org/forums/discussion/2/what-opencl-libraries-are-already-out-there

http://developer.amd.com/tools-and-sdks/heterogeneous-computing/amd-accelerated-parallel-processing-math-libraries/

I don't know about books and guides for OpenCL. But, I prefer quality over quantity. And there are certainly good OpenCL books out there. Fewer books also makes it easier to choose one ;)

3) The question seems to be not directed towards supercomputers. And there are many computers with AMD hardware. Considering the point of performance per dollar AMD is a viable choice compared to NVIDIA, especially if it is about double precision computation.

4) OpenCL only has an initial overhead during the setup phase. Usually, this is mostly Copy&Paste to get it working in the first place. The actual implementation for just one specific type of hardware does not have any overhead compared to CUDA.

5) The question is not about writing portable code. (And CUDA is not portable at all, so this is not a fair point.) When choosing only one target platform OpenCL is as easy to tune as CUDA.

Coming back to the original question, AMD is a viable choice. It is true that OpenCL might be slightly slower on NVIDIA hardware. But then again you wouldn't be switching from CUDA to OpenCL if you would keep using NVIDIA hardware. The only of Ivan's points that is applicable to your situation is that there is quite an overhead during initialization. But, this point can be countered by the lower cost for higher performance. In the end, you can save a lot of time with every computation with a higher programming overhead for writing the initial setup.

Simon makes many good points. (Just a remark on Simon's first post: I believe the 4-float vector aspect is no longer relevent on recent AMD hardware). A few words about my personal experience: I asked myself Marco's question some time ago, double precision being necessary to my work (finite-difference-based computational aeroacoustics).

I chose to try both AMD (ATI at the time) and NVIDIA hardware with OpenCL, and have also compared skeleton CUDA and OpenCL codes on NVIDIA cards.

On the whole I have been very happy with OpenCL code on AMD hardware, regarding both performance (excellent!) and price (cheap) and indeed for a recent article performed all computations on a 6 GB 6970 card.

On a side note, if you do want to look at what OpenCL looks like before buying graphics hardware, both intel and amd have SDKs which contain sample code which will run on your CPU. This is a fairly easy way to get a feel for what the language is like (it does *not* tell you how code will perform on a GPU, of course!).

Thanks, to all,

many points to consider then. I am evaluating a Radeon R9 280X (a nice fit for my price point) and I guess NVIDIA with maxwell definitely created a bridge between single and double precision computation, leaving the field to AMD in the prosumer  price target (maybe it is a too small niche to create interest)

I would like to comment that if you use MATLAB as your main data analysis platform the weights change again: MATHWORKS provides only CUDA accelerated toolboxes (... i am so sad!).

I forgot to mention the Xeon Phi accelerator , although the price/performance target is not very attractive for me , never the less Intel compilers seems to compile without significant code changes (at least for un-optimized code) really smoothing the learning process.

Nearly three years after the original post, old AMD cards on the used market continue to provide a win for price to performance with respect to double precision TFlops. This is great for researchers on a budget who want to test the waters in GPU programming. Older Cards have good driver support on Linux and can be found at online auction sites for deep discounts. Note that the more powerful AMD cards (7970/ 7990) can get very hot and one requires an above average cooling setup in order to let them function. On the other hand, it is hard not to envy recent products from AMD that sport larger high bandwidth memory architectures and integrated water cooling at a competitive price/performance ratio.

I have thought about this as I want to do GPGPU for scientific computing. The truth is that Nvidia has a huge user base and you get help with issues faster you have better documentation and if you're using these things for something you don't want to be the person finding all the bugs in some specific cases in not so well tested libraries. At least my experience with Intel ISPC has been dismal and I want a platform that is well documented and well maintained not one that is barely kept alive. If you're doing research and you write to Nvidia and ask for a card for research they might grant you a higher end card for testing. That might help you with the price point. In my opinion it's worth going to a well documented well maintained platform.

Or waiting to C++ multithread at GPU. Opencl is too difficult and have serious bottleneck that avoids efficient using its 100% power

I am currently learning to compile tensorflow for a NVIDIA Quadro K5000 which has a computing capability of 3.0 and so is no longer supported in the publicly distributed compilation as of tensorflow 1.11.

I noticed in the compiler flags that you can compile tensorflow for AMD.

I hope this might help. Good Luck. I’ve been worried about this issue.

We have just ordered an AMD dual socket EPYC Supermicro MoBo with twin RTX 2080Tis, so we're pretty excited to jump into the GPU computing sphere! My sense is that nVIDIA has just been around longer, but I'm glad to see AMD pushing it's OpenCL development forward.