Previous | Next --- Slide 53 of 78
Back to Lecture Thumbnails
jmackama

I'm having a hard time keeping terminology straight. In CUDA, are warps the sub-problems, or the worker threads?

anamdev

I believe it depends on the implementation but usually warps would be the sub-problem. Each warp would go onto a single block of the GPU so in terms of communication it would make a lot more sense and would be more practical for all the threads in that warp to perform tasks in the same region (of an image) or on the same data.

ferozenaina

The warps can be thought of as groups of subproblems. The 1080 can have 64 simultaneously active warps but only a few of these warps(4?) run concurrently. These running warps can be thought of as parallel worker threads.

Each warp itself consists of 32 threads (tasks).