A question about the CUDA's thread parallelization Options

[yzhao]

I am newbie of CUDA. I have a question about how much thread parallelization CUDA can actually provide.

On the chapter 5.1 of programming guide, it says:

The maximum number of threads that can run concurrently on a multiprocessor is 768.

But it also says: The warp size is 32 threads for 8800 series. And the definition of warp is (see chapter 3.2): Each block is split into SIMD groups of threads called warps. And SIMD means: at any given clock gycles, each processor of the multiprocessor executes the same instruction, but operates on different data.

So I got a little confused. Could anybody answer this question to straight it up:

If I have 768 threads which do the same job on different data and somehow I manage them to be excuted on one multiprocessor, then they will be excuted concurrently, or divided into 24 warps, and excuted in sequence?

Thanks,
Yong