This is GC — yes, even with it only showing ~0% GC time.
eigen allocates an 800MB matrix to store its result right at the get-go. You’re doing that 8 times, all at once. It’s highly likely that Julia may think that — after allocating a few of them — it’s a good idea to do garbage collection. BUT: the GC does not do its collection multi-threaded-ly. Instead, it holds the thread until all threads hit a GC safepoint. But a few of the threads have already kicked off into a multi-minute BLAS ccall. So it’s gotta wait for those to return. In the meantime, all the other threads are just sleeping. That sleeping time does not appear as GC time in @time’s metrics.
That’s why it’s inconsistent — both in number of times you perform the benchmark (it may be that you’ve not hit memory pressure the first time you do it) and in the machines folks are reporting back (they may have so much memory they don’t feel any pressure).
See: GC-safe `ccall`s · Issue #51574 · JuliaLang/julia · GitHub