Copying it 1k or 10k for each thread is wasteful, even though they are
lightweight copies.

This brings the runtime down from 133 to ~100 hours.