`X.T @ X` is almost 2 times slower if `X` has any dependencies in the task graph about dask HOT 1 CLOSED

Thanks for opening this issue. This is indeed an interesting case. FWIW I have no idea what numpy is doing to be faster when it's a same-matrix multiplication but I guess that doesn't really matter.

Interestingly, this is somehow connected to task fusion and I wonder if there are any bigger implications here or if this is just affected this specific case.

if you set

import dask
dask.config.set({"optimization.fuse.active": False})

you should preserve for Z the same perf characteristics as for X. (BTW there is no need to run on so many columns. If you reduce the n_cols to smth like 1k, the entire benchmark will run much faster but will produce similar results)

from dask.