"All told, training the GameGAN for Pac-Man took a quad GV100 setup four days, over which time it monitored 50,000 gameplay sessions. Which, to put things in perspective of the amount of hardware used, 4 GV100 GPUs is 84.4 billion transistors, almost 10 million times as many transistors as are found in the original arcade game’s Z80 CPU. So while teaching a GAN how to be a Pac-Man is incredibly impressive, it is, perhaps, not an especially efficient way to execute the game"
Note that it took 4 days to *train*, not execute. A more appropriate analogy would be development time... aka X months of meatbags on typewriters.
I do wonder what it takes to run at 50 FPS. The plans for a demo suggest you don't need a DGX box that costs more than a house, but can it execute on, say, an 8GB RTX 2070?
Make 6150 bucks every month… Start doing online computer-based work through our website. I have been working from home for 4 years now and I love it. I don’t have a boss standing over my shoulder and I make my own hours. The tips below are very informative and anyone currently working from home or planning to in the future could use this website. WWW. iⅭash68.ⅭOⅯ
Well, not really from scratch when it takes 50,000 repetitions of observing the original game to do it - it's basically the neural net equivalent of a savant copying a painting, except in motion and playable.
The code and the sprites are there, the AI is just taking a guess at the former and copying the latter from what it sees on screen.
I'd be much more impressed if it really created Pac Man from scratch using no more than a description of game play.
I'll be even more impressed when it can meet and surpass the efforts of ReactOS in reverse engineering Windows (at this point they may as well give up).
No way in heck a trained neural net runs faster or in less space than a couple K of Z80 ASM. Especially not if it has to use modern IO. A minimal USB stack is more code, and takes more processing time, than all of Pac-Man.
Yeah, Pac-man was... honestly larger than I thought. 16 kilobytes of ROM and 4K of RAM. Attached to a 3-MHz Z80. I'd say they're all popcorn chips today, but popcorn chips are an order of magnitude more capable.
The thing I find most amusing about this is that it failed to identify the simple algorithyms that constitute ghost behavior. The AI thinks "don't kill Pac-Man" is a game rule. Even in the brief snippet of play shown, the AI failed to learn that the ghosts flee Pac-Man when he is energized.
Quite astounding, even if it is early first steps. Perhaps there'll come a time when we see an AI generating something like Elder Scrolls 28: Return to Valenwood ;)
NVIDIA has come a very long way. They are constantly evolving. People also strive for excellence and continuous development. Try Paperell https://paperell.com/buy-term-paper. During distance learning, this is perhaps the best choice.
There are vacuum cleaners of different size and plan as customary new thing thinks about the market and the model gets superseded. The vacuum cleaners were earlier with a progress pack which amassed the improvement in the wake of cleaning at any rate it was required to purchase the improvement sack again and again. https://bestvacuumcleaner.us/ Nowadays the vacuum cleaners are bagless and are sure. The vacuum cleaners go with various sizes and part so guarantee you picked the cleaner that draws in your necessities.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
20 Comments
Back to Article
brucethemoose - Friday, May 22, 2020 - link
"All told, training the GameGAN for Pac-Man took a quad GV100 setup four days, over which time it monitored 50,000 gameplay sessions. Which, to put things in perspective of the amount of hardware used, 4 GV100 GPUs is 84.4 billion transistors, almost 10 million times as many transistors as are found in the original arcade game’s Z80 CPU. So while teaching a GAN how to be a Pac-Man is incredibly impressive, it is, perhaps, not an especially efficient way to execute the game"Note that it took 4 days to *train*, not execute. A more appropriate analogy would be development time... aka X months of meatbags on typewriters.
I do wonder what it takes to run at 50 FPS. The plans for a demo suggest you don't need a DGX box that costs more than a house, but can it execute on, say, an 8GB RTX 2070?
Lord of the Bored - Friday, May 22, 2020 - link
I'd hope that running Pac-Man on a 2070 at least equals the original 60FPS.jeremyshaw - Saturday, May 23, 2020 - link
Costs more than a house?Ah, you don't live in this nightmare world called Silicon Valley. Hopefully, remote work will allow people to think of $200k as a house, lol.
p1esk - Sunday, May 24, 2020 - link
Quad GV100 setup is ~$40k. Where did you get $200k from? Also, I don't think $200k will get you a decent house anywhere. Certainly not in Houston, TX.skavi - Tuesday, May 26, 2020 - link
They are referring to the above mentioned DGX boxalicebcao75 - Monday, June 8, 2020 - link
Make 6150 bucks every month… Start doing online computer-based work through our website. I have been working from home for 4 years now and I love it. I don’t have a boss standing over my shoulder and I make my own hours. The tips below are very informative and anyone currently working from home or planning to in the future could use this website. WWW. iⅭash68.ⅭOⅯeastcoast_pete - Friday, May 22, 2020 - link
So, how did the AI deal with the copyright/IP issues around PacMan? If it managed to find a loophole by itself, I'd really be impressed (:UltraWide - Friday, May 22, 2020 - link
LOLFortunately it was in collaboration with NAMCO for their 40th anniversary for PAC-MAN!
edzieba - Friday, May 22, 2020 - link
It created the entire game from scratch, no original code or sprites were copied.soresu - Friday, May 22, 2020 - link
Well, not really from scratch when it takes 50,000 repetitions of observing the original game to do it - it's basically the neural net equivalent of a savant copying a painting, except in motion and playable.The code and the sprites are there, the AI is just taking a guess at the former and copying the latter from what it sees on screen.
I'd be much more impressed if it really created Pac Man from scratch using no more than a description of game play.
I'll be even more impressed when it can meet and surpass the efforts of ReactOS in reverse engineering Windows (at this point they may as well give up).
surt - Saturday, May 23, 2020 - link
Funny, but no, there would be no copyright issues since it's demonstrable that no copying took place.ajp_anton - Friday, May 22, 2020 - link
When actually executing the game, would this be more efficient than the original one? And what about storage space?This kind of game would be a hell to debug though.
Lord of the Bored - Saturday, May 23, 2020 - link
No way in heck a trained neural net runs faster or in less space than a couple K of Z80 ASM.Especially not if it has to use modern IO. A minimal USB stack is more code, and takes more processing time, than all of Pac-Man.
Spunjji - Tuesday, May 26, 2020 - link
"A minimal USB stack is more code, and takes more processing time, than all of Pac-Man"Doof. That really puts things in perspective!
Lord of the Bored - Tuesday, May 26, 2020 - link
Yeah, Pac-man was... honestly larger than I thought. 16 kilobytes of ROM and 4K of RAM. Attached to a 3-MHz Z80.I'd say they're all popcorn chips today, but popcorn chips are an order of magnitude more capable.
Lord of the Bored - Saturday, May 23, 2020 - link
The thing I find most amusing about this is that it failed to identify the simple algorithyms that constitute ghost behavior. The AI thinks "don't kill Pac-Man" is a game rule.Even in the brief snippet of play shown, the AI failed to learn that the ghosts flee Pac-Man when he is energized.
GeoffreyA - Sunday, May 24, 2020 - link
Quite astounding, even if it is early first steps. Perhaps there'll come a time when we see an AI generating something like Elder Scrolls 28: Return to Valenwood ;)Ej24 - Sunday, May 24, 2020 - link
Nah 100 years from now we'll only be on elder scrolls 7.anitafox - Tuesday, May 26, 2020 - link
NVIDIA has come a very long way. They are constantly evolving. People also strive for excellence and continuous development. Try Paperell https://paperell.com/buy-term-paper. During distance learning, this is perhaps the best choice.vacumcleans - Tuesday, June 9, 2020 - link
There are vacuum cleaners of different size and plan as customary new thing thinks about the market and the model gets superseded. The vacuum cleaners were earlier with a progress pack which amassed the improvement in the wake of cleaning at any rate it was required to purchase the improvement sack again and again. https://bestvacuumcleaner.us/ Nowadays the vacuum cleaners are bagless and are sure. The vacuum cleaners go with various sizes and part so guarantee you picked the cleaner that draws in your necessities.