I grew up on assembler, cycle counting, and computer graphics. You're right, a lot of the old-timey optimizations are not applicable on today's hardware, but this is just the march of technology. I didn't have to worry about generating graphics scanline-by-scanline like those poor Atari 2600 programmers. They didn't have to worry about instruction pairing on the Pentium though. Each generation has their tricks.

As for 'manual' management, consider script engines that need to temporarily store the AST while compiling/interpreting a piece of script. These typically allocate each tree node by pointer-bumping from a large pre-allocated block of memory, and free the entire thing once done by resetting the pointer. Basically this is the same idea as a nursery in a generational collector, with the difference that you already know beforehand that you're not going to promote any of the elements.

What GC algorithm is best (mark/sweep, generational, soft real-time, hard real-time) depends on your needs I guess. In that light, it's too bad MSR canned Singularity: http://research.microsoft.com/pubs/69431/osr2007_rethinkings...