OpenCL vs. CUDA

Today's processors have undergone a huge transformation from those of just 10 years ago.  CPU manufacturers Intel and AMD (and up and coming CPU designer ARM) have increased processor speed via greater emphasis on superscalar execution, deeper pipelining, branch prediction, out of order execution, and fast multi-level caches.  This design philosophy has resulted in faster response time for single tasks executing on a processor, but at the expense of increased circuit complexity, high power consumption, and a small number of cores on the die.  On the other hand, GPU manufacturers NVIDIA and ATI have focused their designs on processors with many simple cores that implement SIMD parallelism, which hides latency of instruction execution [1].

While GPUs have been in existence for about 10 years, the software support for these processor have taken years to catch up.  Software developers are still sifting through solutions for programming these processors.  OpenCL and CUDA are frameworks for GPGPU computing.  Each framework comprises a language for expressing kernel code (instructions that run on a GPU), and an API for calling kernels (from the CPU).  While the frameworks are similar, there are some important differences.

CUDA is a proprietary framework. It is not open source, and all changes to the language and API are made by NVIDIA. But, some third-party tools have been built around the framework and it does seem to have a large following in academia.  Unfortunately, CUDA only runs on NVIDIA devices.  While it should be possible to run CUDA code on other platforms using Ocelot, this only works on Linux systems.
OpenCL is a standardized framework, and is starting to gain popularity.  Similar to NVIDIA's CUDA C++, OpenCL allows programmers to use the massive parallel computing power of GPU's for general purpose computing.  Unlike CUDA, OpenCL works on any supported GPU or CPU, including Intel, AMD, NVIDIA, IBM, and ARM processors. 
Does OpenCL make programming multiple platforms easier?  Is it as fast as CUDA, or does it sacrafice speed for diverse platform support?

Continue reading