Different block can not communicate with each other, right?

Accelerated Computing CUDA CUDA Programming and Performance
1

For current GPU, different blocks can not directly communicate with each other? They have to store the value to global memory and then sync and can share the information, right? (I heard H100 can communicate between shared memory, but it is not available now.) Thanks!

2

That’s correct. Details of Hopper’s Distributed Shared Memory can be found here.

1 Like
3

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.