I have to say this is a very good book to learn more about cuda, especially for a novice like me, who take interest in parallel computing and is not content with cuda programming guide.
There are several important features of cuda included in this book I want to share with readers:
1、cuda employs single instruction, all data model(SIMD), which is same as that of CPU;
2、All threads in a warp will have the same executive time, that is why cuda programming guide suggest that we use 32x threads.
3、Each stream processor(SM) can execute only one instruction for a small subset of warps;
4、Latency tolerance enables GPUs to save much area dedicated to cache memories and branch prediction mechanism, which is quite different from CPU;
5、It also introduces a proper way to schedule the number of threads and blocks, taking into account the computability.