In H100 whitepaper , when describe 'Asynchronous Transaction Barrier`, it uses this image
What is the difference between ‘spin’ and ‘sleep’?
A spin loop is constantly executing instructions to see if the condition is met. In this case it is spinning on shared memory. In GH100 waiting threads can wait on the condition on a sleep state. When the condition is met the threads will move back to active. The ability to sleep a warp until a condition is met reduces instruction issue and memory traffic allowing the warp scheduler to make forward progress with other warps.
1 Like
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.