mirror of
https://github.com/NVIDIA/nccl-tests.git
synced 2026-01-14 02:47:21 +08:00
Clarification
This commit is contained in:
parent
e00cb1f1c4
commit
eb4c43ff3d
@ -54,7 +54,7 @@ Considering that each rank has a bandwidth to the outside world of _B_, the time
|
||||
`t = (S*2*(n-1)) / (n*B)`
|
||||
|
||||
Indeed, we have _S_ elements, 2*(n-1) operations per element, and _n_ links of bandwidth _B_ to perform them.
|
||||
Reordering the elements, we find that
|
||||
Reordering the equation, we find that
|
||||
|
||||
`t = (S/B) * (2*(n-1)/n)`
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user