I understand the point of scaling up programs. However, what is the point of scaling down programs? It seems that when you write programs, you program according to how many processors you have. Why would you want to reduce that?
This comment was marked helpful 0 times.
nkindberg
@akashr I don't think you always know exactly how many processors your program will be run on. As the last bullet point mentions, CUDA is designed to work on both high-end and low-end GPUs. If you publicly released CUDA code, for example, people with both high-end and low-end GPUs could be running your code. So if you test and optimize your code on a GTX 680, you would probably also want to see how it scales down to a GTX 650 which has a lot less processors than the 680 since you will very likely have people with low-end GPUs trying to run your code.
I understand the point of scaling up programs. However, what is the point of scaling down programs? It seems that when you write programs, you program according to how many processors you have. Why would you want to reduce that?
This comment was marked helpful 0 times.
@akashr I don't think you always know exactly how many processors your program will be run on. As the last bullet point mentions, CUDA is designed to work on both high-end and low-end GPUs. If you publicly released CUDA code, for example, people with both high-end and low-end GPUs could be running your code. So if you test and optimize your code on a GTX 680, you would probably also want to see how it scales down to a GTX 650 which has a lot less processors than the 680 since you will very likely have people with low-end GPUs trying to run your code.
This comment was marked helpful 2 times.