Concurrent writes to global memory


I’m looking for some information about the behavior of concurrent memory writes to the global memory.
I found somewhere on the web that one of the transactions overwrites all the preceding ones (with undefined order), but I wanted to be sure.

In my program, I’m doing concurrent writes with the same data : in this case, the behavior is well defined ?

I didn’t found anything in an official “NVIDIA paper” !

Thank you !

Read the CUDA programming guide very carefully. It states that when when warp makes multiple writes to the same memory locations, at least one of the writes will sucessed. Which one that is in undefined.