When I was interviewing with the DB2 team for IBM, they mentioned that the high-end IBM database servers implement special instructions (CPU-level) for accelerating certain DB operations. I don't know any more details than that, but I'd guess that's where Oracle is going to go with SPARC.
I'm not aware of any DB2-specific instructions, but POWER6 and above implement a decimal floating point instruction set, allowing things like regulatoriliy-mandated financial computations to be hardware accelerated instead of computed in software. That's clearly a nod to the same market that is buying DB2.
That might be as simple as having super-low level operations that they hand write in assembly for each supported CPU, and all other layers staying the same.
