Re: Well DUH ! AI People Finally Realize They Can Ditch Most Floating-Point for Big Ints

On 10/16/24 6:56 AM, Richard Kettlewell wrote:

"186282@ud0s4.net" <186283@ud0s4.net> writes:

   BUT ... as said, even a 32-bit int can handle fairly
   large vals. Mult little vals by 100 or 1000 and you can
   throw away the need for decimal points - and the POWER
   required to do such calx. Accuracy should be more than
   adequate.

 You’re talking about fixed-point arithmetic, which is already used where
appropriate (although the scale is a power of 2 so you can shift
products down into the right place rather than dividing).
 

   In any case, I'm happy SOMEONE finally realized this.
>
   TOOK a really LONG time though ......

 It’s obvious that you’ve not actually read or understood the paper that
this thread is about.

   Maybe I understood it better than you ... and from
   4+ decades of experiences.
   But, argue as you will. I'm not too proud. Many are
   better than me, many more are worse.
   IF this was just about the power reqs of various forms
   of fixed/floating then there'd be little point in the
   article. Breaking the FP tradition as much as possible
   and going to (wide) ints really CAN save tons of power
   and time. With current AI systems this is a BIG deal.
   There was a period where I had to do some quasi-AI stuff
   for micro-controllers. Crude NNs/Fuzzy mostly. Not too
   sophisticated, yet the approach DID make 'em better.
   Now DO check into what's needed for FP on a PIC or 8051.
   It's nasty. By seeing beyond the usual examples in books
   and articles - which all used FP for "convenience" -
   I found the vast advantages of substituting ints instead.
   Easy to FAKE a few decimal-points of precision using
   ints. That's usually more than good enough.
   You can splice an NPU into most any kind of processor
   BUT the steps to do FP don't really change, still suck
   up power. Just SEEMS trivial because it's faster.