You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The 'nki.language.all_reduce' function enables communication between cores within a single Neuron chip. It performs a reduction operation (like, sum, max) across these cores, then shares the result with all participants ( Not chip2chip).
Think of it as internal coordination for parallel processing on one chip.
The 'nki.language.all_reduce' function enables communication between cores within a single Neuron chip. It performs a reduction operation (like, sum, max) across these cores, then shares the result with all participants ( Not chip2chip). Think of it as internal coordination for parallel processing on one chip.
Now i understand this api is for core2core communication, but there is another problem, I did not see any hardware link between cores in the architecture description. So is this api implemented by swapping data via HBM? or if there is a hardware link, what is the bandwidth?
There is a function
all_reduce
in nki document, but i donot understand the usage of this function.Is it used for chip2chip communication or core2core communication?
The text was updated successfully, but these errors were encountered: