Yes I've read the paper. Advances in CPU microarchitecture alone have noticeably shrunk that gap. There have also been significant algorithmic advances in GC. For instance, a few simple prefetching hints speed up marking easily by 30%, and parallel marking scales almost linearly with core count, to mention but two simple performance improvements.

So like I said, maybe in 2005 a 3x memory overhead only incurred 17% slowdown, but 3x overhead is at least at par, and possibly even a little ahead, of manual memory management.