News Nvidia Recreates Pac-Man Using AI


I wonder what the file size difference is between a Human Generated Pac-Man game and an AI Generated Pac-Man game?
My guess is somewhere in the range of thousands to millions of time the size.

Deep neural networks can be very large, while the original Pac Man game fit in only 16 KB of ROM for all the maps, sounds, characters, and game logic.

Moreover, Pac Man ran on a 3 MHz 16-bit CPU, while I'm sure Nvidia's version requires a GPU millions of times faster.

Last edited:
Reactions: digitalgriffin


Pac Man is nice game but generated game is with too low resolution !!
Their point isn't that it's good, but rather that it's possible.

You could think of it a bit like people who overclock CPUs using LN2 or who build dragsters that reach over 300 MPH in a quarter mile - they don't do it because it's a viable way to operate a computer or build a street car, but just to see how far they can push the technology.

Nvidia is simply showing how deep learning can tackle ever more problems than it was originally conceived to be suited for. It's a little bit like a solution that's searching for problems, but worthwhile if they pioneer new techniques in the process. Hopefully, others will eventually employ these same techniques to do something a lot more useful.
Reactions: JarredWaltonGPU