Torvalds' argument wasn't that microkernels were slower though. It was that distributed algorithms are harder than shared memory algorithms.

Having not written a kernel, I can't judge those claims. It would be interesting to know if Minix has any "harder" algorithms than Linux because of this.

I think L4 already existed around that time, and got 10x or 100x improvement in IPC times, showing that microkernels could be fast enough.