CUDA C is a programming language with C syntax. Conceptually it is quite different from C.
The problem it is trying to solve is coding multiple (similar) instruction streams for multiple processors.
CUDA offers more than Single Instruction Multiple Data (SIMD) vector processing, but data streams >> instruction streams, or there is much less benefit.
CUDA gives some mechanisms to do that, and hides some of the complexity.
CUDA is not optimised for multiple diverse instruction streams like a multi-core x86.
CUDA is not limited to a single instruction stream like x86 vector instructions, or limited to specific data types like x86 vector instructions.
CUDA supports ‘loops’ which can be executed in parallel. This is its most critical feature. The CUDA system will partition the execution of ‘loops’, and run the ‘loop’ body simultaneously across an array of identical processors, while providing some of the illusion of a normal sequential loop (specifically CUDA manages the loop “index”). The developer needs to be aware of the GPU machine structure to write ‘loops’ effectively, but almost all of the management is handled by the CUDA run-time. The effect is hundreds (or even thousands) of ‘loops’ complete in the same time as one ‘loop’.
CUDA supports what looks like if branches. Only processors running code which match the if test can be active, so a subset of processors will be active for each ‘branch’ of the if test. As an example this if... else if ... else ..., has three branches. Each processor will execute only one branch, and be ‘re-synched’ ready to move on with the rest of the processors when the if is complete. It may be that some of the branch conditions are not matched by any processor. So there is no need to execute that branch (for that example, three branches is the worst case). Then only one or two branches are executed sequentially, completing the whole if more quickly.
There is no ‘magic’. The programmer must be aware that the code will be run on a CUDA device, and write code consciously for it.
CUDA does not take old C/C++ code and auto-magically run the computation across an array of processors. CUDA can compile and run ordinary C and much of C++ sequentially, but there is very little (nothing?) to be gained by that because it will run sequentially, and more slowly than a modern CPU. This means the code in some libraries is not (yet) a good match with CUDA capabilities. A CUDA program could operate on multi-kByte bit-vectors simultaneously. CUDA isn’t able to auto-magically convert existing sequential C/C++ library code into something which would do that.
CUDA does provides a relatively straightforward way to write code, using familiar C/C++ syntax, adds a few extra concepts, and generates code which will run across an array of processors. It has the potential to give much more than 10x speedup vs e.g. multi-core x86.
Edit – Plans: I do not work for NVIDIA
For the very best performance CUDA wants information at compile time.
So template mechanisms are the most useful because it gives the developer a way to say things at compile time, which the CUDA compiler could use. As a simple example, if a matrix is defined (instantiated) at compile time to be 2D and 4 x 8, then the CUDA compiler can work with that to organise the program across the processors. If that size is dynamic, and changes while the program is running, it is much harder for the compiler or run-time system to do a very efficient job.
EDIT:
CUDA has class and function templates.
I apologise if people read this as saying CUDA does not. I agree I was not clear.
I believe the CUDA GPU-side implementation of templates is not complete w.r.t. C++.
User harrism has commented that my answer is misleading. harrism works for NVIDIA, so I will wait for advice. Hopefully this is already clearer.
The hardest stuff to do efficiently across multiple processors is dynamic branching down many alternate paths because that effectively serialises the code; in the worst case only one processor can execute at a time, which wastes the benefit of a GPU. So virtual functions seem to be very hard to do well.
There are some very smart whole-program-analysis tools which can deduce much more type information than the developer might understand. Existing tools might deduce enough to eliminate virtual functions, and hence move analysis of branching to compile time. There are also techniques for instrumenting program execution which feeds directly back into recompilation of programs which might reach better branching decisions.
AFAIK (modulo feedback) the CUDA compiler is not yet state-of-the-art in these areas.
(IMHO it is worth a few days for anyone interested, with a CUDA or OpenCL-capable system, to investigate them, and do some experiments. I also think, for people interested in these areas, it is well worth the effort to experiment with Haskell, and have a look at Data Parallel Haskell)