|
- Gt300 is going to suport user debug (i.e. user can pause the execution of kernel and get feedback of current states).
- CUDA library is getting more enhancements, which should be saving user optimization efforts.
- from my experience with CUDA, I typically testing hardware with different thread block sizes combined with memory access patterns, etc to find a optimized parameters for my applications. The testing calibration itself is rather straightforward. |
|