New processor to handle AI

#1 Hmm.. I think it's a common misconception that AI willed be "solved" by more CPU cycles. It's not the case. Running neural networks, for example, is pretty easy but it's still hard to get them to do useful things because it's the way a problem is posed and bounded that controls how effective they are, not how much processing power they have to play with.

#2 I do agree.

And a "new" processor won't make games with poorly programmed pathfinding better. To dedicate AI to a processor you need to have a decent AI in your game.

Another marketing rubbish.

STef

#3 Even if all these new dedicated chips are great - the AI chip, the physics chip, the ray-tracer chip - and work as advertised they all have the same chicken and egg problem. No one writes code for them because no one has them and no one buys them because no one writes code for them. As much hype as they manage to generate more technologies have come and gone simply because they never gained an audience.

Also - insanely better images and AI and particle accurate physics are not going to make your game fun to play. As the newest generation of consoles are proving - the "fun" of game doesn't live in the technology and bigger, shinier, louder and more accurate makes a good tech demo but doesn't make a good game.

=Tod

#4 It's interesting.. never say never. Keep an eye on this one. Remember Bill Gates said no one would ever need more then 256 Kilobytes of RAM once....

#5 Actually, Randy, that should be 640k, and whether he actually said it is anybody's guess. ;)

#6 640 kilobytes of Conventional memory, mind you. ;) RAM is another beast, and could be alot higher.

#7 Ah yea 640. Sorry it was late..

#8 Bill Gates never said that (check the urban myths site for references).

Anyway, AI does have a cycle and resource problem. There exists interesting AI algorithms, today, that aren't really feasible for real-time games on a regular PC (much less for squads, or even hordes of AIs in a large online game). Neural nets, by and large, are not the biggest area of research right now; there's still work to be done in things like knowledge representation :-)

AI could probably be accelerated with associative memory (like really fast hash tables). Other than that, it's really too branchy in most of its applications to parallelize well like the graphics cards. A CPU aimed towards rules matching (prolog style), with associative memory, could probably do neat things, but the question is how many neat things more than a regular CPU thread.

#9 I'm not really saying "never". But I am saying is that for every 100 flashy "the-next-big-thing" technology announcements - always being made by the parties that have something to gain - I see maybe one that actually succeeds in the market place. You have the whole chicken<->egg problem I outlined above and then you have the steadily rising tide of general computing power. While your chip may give you 10x the raytracing/AI/Physics power of the current PC within a couple years the PC has cut that advantage to more like 3-5x the power and advantage is suddenly less compelling. Even John Carmack admitted that when they ship a new engine -eg Quake 4, Doom 3 - it is basically targeted at the last generation video cards because those cards were the bleeding edge target when they started coding. At some point you have to pick a target and shoot for it. If you're constantly choosing new targets you'll never ship. (Insert link to Duke Nukem Forever here.)

I would love to own the box with all these things in it. But if I'm trying to ship a game I'm not going to spend time learning about how to integrate the chip, take advantage of it and program/debug for it when no one has it yet. And I really not going to program a game that depends on specialized ray-tracing (or whatever) and assume that the market of 1Million machines I need to get my 100K in sales to break even are going to exist when I'm done. I'll re-evaluate when I ship and maybe my next game will support it... but with 2+ year game cycles that means that people are going to have to buy the new technology betting everyone's next game is going to use it. Only game consoles get around this by delivering an entire package - not just a component - and promising millions of installed units as a market for when you finish your game even when they haven't started shipping units yet.

I'm not trying to be a hater, I've just seen alot of tech demos turned products never reach critical mass for these reasons.

=Tod