--- paulmillscv <

paulmillscv@...> wrote:

> Yes, firmware (a combination of algorithms in hardware) is very

> underexplored for practical implementation of algorithms. So it

> remains now to define a unit of work in mathematics/computer

> science. I imagine something like the 16 bit signed multiplication

> is the unit?

If you know that you will be spending all your time doing bignum

multiplies, then you will look at the most optimal bignum

multiplication algorithm at the firmware level. You can do bignum

multiplies in almost constant time, but it all depends on how much

silicon you want to devote to it. If you know that you're unlikely to

go over 512 bits, then you might want a 512*512 multiplier unit for

example. Again, you have choices of whether you want to minimise

latency or not. If you can pipeline large quantities of operations,

then you don't need a low latency. If everything must be serialised,

then you've got to go for a more expensive low latency design.

(Either that or you explicitly work on two or more different problems

in parallel, but that in turn requires some smarts...)

Phil

__________________________________________________

Do You Yahoo!?

Yahoo! Games - play chess, backgammon, pool and more

http://games.yahoo.com/