I may not understand the question well but well, I'll try to answer.

First of all you need to clarify what you mean by "the CUDA core". If a multiprocessor, then there's no (simple. obvious) way to do what you want to do: a CUDA job is distributed among multiprocessors automatically by the driver/hardware. One trick I can think of is to launch a kernel that will surely completely occupy one multiprocessor (you can force it by allocating in it 32K of shared memory) and then launch your kernel on the remaining multiprocessors in a separate stream. Nasty solution.

If you mean "cuda core = scalar processor, 32 of which make up a warp", then you can artificially reduce the occupancy, eg. by allocating some extra shared memory that you don't actually need. You're almost sure to see a great decrease in your kernel performance as the occupancy decreases. This solution is not as ugly as the first one, but the question still remains: what do you need such tricks for? The scaling you're after is very unnatural for CUDA. Occupancy is the proper parameter to focus on.

I want to compare Cuda with the Intel Mic for which you can readily produce scaling of performance against resources used untilyou saturate the memory throughput.

On CPU you have two numbers: number of cores(x) and number of threads(y). If you want to use less cores than maximum then you should have x>y and hope than OS scheduler distributes one thread per one core.

On GPU you have four numbers: number of SM(x1), cores per SM(x2), number of thread-blocks(y1) and number of threads per each block(y2).

So under some conditions. GPU executes y1*y2 threads on x1*x2 cores

Since one thread-block can be executed on exactly one SM, if you want to use less cores than maximum then you should use lower value of y1, but the precise number of utilized cores is slightly more difficult to compute.

You can change the launch configuration, which specifies the number of cores and block.

Hello, Paul! I think that those who answered before me may have helped you earlier. However, what I could say is that the number of a GPU's core is not variable. All the threads are scheduled by the scheduler. Usually it is already optimized, being necessary only the specification of the number of blocks and threads per block.

Best wishes and good luck with your research!

Please read first few chapters of this document. surely you can get good knowledge about CUDA