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

https://techxplore.com/news/2024-10-integer-addition-algorithm-energy-ai.html
A team of engineers at AI inference technology company
BitEnergy AI reports a method to reduce the energy needs
of AI applications by 95%. The group has published a
paper describing their new technique on the arXiv preprint
server.
As AI applications have gone mainstream, their use has
risen dramatically, leading to a notable rise in energy
needs and costs. LLMs such as ChatGPT require a lot of
computing power, which in turn means a lot of electricity
is needed to run them.
As just one example, ChatGPT now requires roughly 564 MWh
daily, or enough to power 18,000 American homes. As the
science continues to advance and such apps become more
popular, critics have suggested that AI applications might
be using around 100 TWh annually in just a few years, on
par with Bitcoin mining operations.
In this new effort, the team at BitEnergy AI claims that
they have found a way to dramatically reduce the amount
of computing required to run AI apps that does not result
in reduced performance.
The new technique is basic—instead of using complex
floating-point multiplication (FPM), the method uses integer
addition. Apps use FPM to handle extremely large or small
numbers, allowing applications to carry out calculations
using them with extreme precision. It is also the most
energy-intensive part of AI number crunching.
. . .
   The default use of floating-point really took off when
   'neural networks' became popular in the 80s. Seemed the
   ideal way to keep track of all the various weightings
   and values.
   But, floating-point operations use a huge amount of
   CPU/NPU power.
   Seems somebody finally realized that the 'extra resolution'
   of floating-point was rarely necessary and you can just
   use large integers instead. Integer math is FAST and uses
   LITTLE power .....
   I did one or two apps long back using a sort of "fuzzy
   logic". All the books had examples showing the use of
   floating-point for dealing with the 'fuzzy' values.
   However I quickly figured out that 32-bit ints offered
   more than enough resolution and were very quick - esp
   on micro-controllers.