Eight Cores Forty Trees

Neither parallelizes across trees; both must hold the full dataset in memory, so the difference is cache behaviour

Each forest tree is fit on its own bootstrap sample and depends on no other tree, so all 40 fit at once; boosting round $m$ needs the residuals of $F_{m-1}$, so its trees are strictly sequential and only the split search inside one tree parallelizes

Both parallelize across trees; the forest is faster only because bootstrap samples are smaller than the full training set

Boosting parallelizes across trees and the forest cannot, because its bootstrap samples overlap and would race on shared rows