On x86, you'll get piss all performance increase. Let's face it, today's latest x86 CPU's are still little more than overclocked 386's with additional instruction sets thrown in. (such as SSE{,2}, MMX, 3DNow, etc)

It's still very much the same register-starved architecture it was back in 1990. This isn't the case with other architectures.

On MIPS, take, say, an Indy, chuck on Debian/MIPS, and have a fiddle. When you're done, throw on Gentoo/MIPS, and compare.
Debian/MIPS is compiled for MIPS1-class CPUs, with a little MIPS2-class code thrown in.
Gentoo/MIPS is compiled for either MIPS3-class or MIPS4-class CPUs -- depending on the stage tarball you download.

I've found my MIPS boxes much more responsive under Gentoo than under Debian, largely for that very reason.

What's the difference between MIPS1 and MIPS3/4? Well, more instructions implemented in hardware (as per the RISC philosophy), and also MIPS3, MIPS4, MIPS5 and MIPS64 are all 64-bit, whereas MIPS1, MIPS2 and MIPS32 are all 32-bit.

Yes, the Pentium4 has more instructions than the 386 -- many of them implemented in microcode, which breaks down to RISC-like instructions in the end. So you're just replacing a bunch of compiler-generated instructions in a program, with just one instruction in the program, that ends up translating to the same microcode program in the end.