Jeroen,

i think you've seen some good observation so far regarding the parametric and functional uncertainty that can result from over-clocking parts. i doubt you'll find many who vote that over-clocking is a good idea for a product you would vend or rely upon in any serious way.

However ...

I over-clocked parts quite often. I haven't seen '51 micros overclock to the degree you've hoped for but significant headroom does exist from certain manufacturer's notable not ATMEL.

I've seen and learned quite a bit about how parts perform and their sensitivities to not only clock frequency but also slope and drive. Heat is a very significant characteristic impacting overclocking as you might have guessed. Also when parts fail you'll often find architectural characteristics may rear their head such as multiplies falling apart before register moves do (not unlike power fail testing scenarios).

Since most parts today, include internal ROM, and implement this with ROM as FLASH, the top end of reliable execution rate most often is attributable to this and not the logic technology or topology. This is why you see many parts move to SRAM for program store.

regards,
p