Bubble Trouble

A GPipe-style pipeline splits a model across p=8 stages and processes a batch as m=32 microbatches, assuming forward and backward pass time per stage are equal and no interleaving is used. Using the standard bubble time fraction approximation (p-1)/m, what fraction of the pipeline's total execution time is wasted on bubbles?

A

About 44%

B

0%, bubbles vanish once m > p

C

About 22%

D

About 8%

E

About 3%