Stand-alone AI card: is it viable?
by Ross Miller 
{ Sep 6th 2006 at 5:55PM }
Theoretically, the card is a great idea in the vein of Ageia's PhysX card -- who doesn't hate the lackadaisical AI found in today's games? But, as Engadget points out, it suffers from the chicken/egg dilemma: no customer will buy the card until games are made that utilize it, but no developer will make a game utilizing Intia unless it already has an installed base. The best bet for AIseek would be to lobby console makers to get its chip included in the next generation -- some of the earliest 3D cards found success because of their inclusion in PlayStation and Nintendo 64.
No one would purchase an AI processor out of support for the idea alone; the technology is great, but what high-profile developer would take the risk and program excess code for a small, possibly nonexistent, audience?
[via Engadget]
Reader Comments (Page 1 of 1)
Tim @ Sep 6th 2006 6:13PM
I agree that it's probably best suited for consoles. Good AI is really mandatory these days and maybe more complex, more realistic AI will be the next leap in gaming (graphics have really reached a point of saturation). Hopefully this chip will make this possible.
www.itradevideogames.com
exy @ Sep 6th 2006 6:11PM
Sure, it might be feasible for consoles... but can I put it into my toaster to make it dance? How about I put it into my lawn gnome and make it patrol my yeard.
These are the kind of things I want from this chip. Somebody please make it happen!
Matt @ Sep 6th 2006 6:16PM
The problem of decent AI is as simple as cost. It costs too much for decent AI. Computation that could go to graphics or physics or other whiz-bang features is neccesary to do AI of this order. Let's say that you could either have wii-level graphics with this type of AI, or crappy AI with whiz-bang specular lighting and whatnot. 9 times out of 10, the audience (nerds) and the investors will say, "pump up the graphics, fake AI is good enough."
Offloading that cost to a dedicated processor doesn't change the fact that the cost is there, and at this point, it's not a cost that the gaming market is ready to pay.
Kanpai @ Sep 6th 2006 6:18PM
Regardless of the product's viability, this is a step in the right direction. Games need better AI; they need it so very badly. Innovations in AI are near stagnant when compared to the advancements in graphical detail, so when anyone makes a move towards improved realism in characters, i'm heartened, even if this is a crap product that no one will buy.
Jason Oliveira @ Sep 6th 2006 6:23PM
uh, the N64 had custom-built hardware from SGI. SGI's video cards were not only limited to their workstations, but they never even got a boost (financially or technologically) from the hardware in the N64.
I won't even go into the fact that the Playstation's 3D hardware had nothing to do with any 3D card.
the first system to use 3D hardware that actually came from a 3D card was the Dreamcast, as it used NEC's PowerVR2. but even that did nothing for NEC, as PowerVR2 tanked due to the overwhelming competition from Glide and OpenGL (and later, DirectX).
maybe if you had read the wikipedia link correctly, you would have seen that it stated something quite different than you thought.
codemule @ Sep 6th 2006 6:46PM
Aegia's card is already a doorstop and it hasn't even shipped.
The same will happen here.
Bill @ Sep 6th 2006 7:00PM
You guys are retarded.
We HAVE incredibly fast processors (compared to what we did ten years ago) do we have better AI? No.
AI isn't possible because computers are STUPID.
I'd venture AI has changed very little from PSX, to PS2, to Xbox 360. Why? Because it doesn't matter how much computing power you have..AI isn't possible.
Have you ever heard Carmack talk about how the original doom creatures had these incredibly simple AI routines, yet people convinced themselves they were doing deep strategies, flanking, sneaking up behind, etc? And they were controlled by simple two or three lines of code.
His point was the same as why he's skeptical of physics in games (which he's been right about to date). Which is, it's not clear how better physics necessarily make a game better. In fact for example, physics based puzzles can easily break a game (example, if these realistic rocks can roll around and possibly block my way..then if that happens I cant continue in the game. Whoops). He also said you can spend tons of power on complex AI, and as the example of the Doom creatures, not necessarily get a better result than simple AI.
I'd say there can be good AI, Halo being the best example, or perhaps FEAR, but it's not even that much AI as doing something different.
I guess the bottom line is, more power DOES NOT equal better AI. It's not a problem like graphics that you can just throw power at.
And back to physics..the Agiea card has been a complete disaster, and again, Carmack pretty much called it exactly. The Ageia card actually slows down games (one of the exact things Carmack mentioned being very worried about), in GRAW the "improved" physics provided by the card are a joke (stupid shit like a little bit more more debris from explosions, that isn't very impressive, all looks the same, and disappears seconds after it hits the ground).
Then somebody hacked some of Ageia's demos and found you dont need the PPU at all to run one or more of them! They worked fine with a normal CPU only! I think Ageia in response then re-jiggered the demos some more, but the damage was done.
But back to AI. Do you know what the turing test is? It's a famous AI test where to pass, a CPU using AI behind a wall or otherwise not seen, would "fool" a human who could ask it questions via text, into not knowing whether it was a real human or a computer.
Well obviously no computer, not the biggest supercomputer in the world, has ever been able to pass that test, by fooling a real human. What does that proof? Among other things it proves: Computers cannot do AI because guess what, they're not intelligent beings, and two, throwing power at the problem means nothing. You could have a supercomputer that takes up a whole rooom, guess what, it still couldn't pass the turing test, no computer to date has (nor ever will).
Brian Sexton @ Sep 6th 2006 7:10PM
They should make one of these for the Xbox 360 and provide a patch for Dead Rising so fewer survivors will fall to the most dangerous thing in the zombie-infested mall: their own stupidity.
Randomoriginality @ Sep 7th 2006 10:39AM
I think standardized processing will get us through the next generation, but specialized processing will be the next big thing. Instead of a triple core 3.2ghz processor with 2 hardware threads per core, or a cell base power PC chip with 7 Synergistic Processing Units the future of console games will be specialized. There will be a Main CPU for running and organizing game code, a GPU for graphics, a PPU for Physics, a NPU for Net code, a AIPU for AI processing and a I/OPU for processing complex 3-d and virtual reality inputs. All of the Processors will have a unified memory infrastructure and can all cross communicate their needs to each other. Each Processor will be built with threads that communicate with each other processor with the Main CPU running the show. The next Generation of Processors combined with 5 specialized next Gen processors could steal the show, however, with all that heat and power being sucked up, not to mention the cost...we would have to be at a point where these things are cheaper to develop...but I can imagine that in 10 years, the games that wow us now, will no longer wow us...and the story continues.
Omega697 @ Sep 6th 2006 7:30PM
Bill, you are spot on.
I actually do AI research. Computing power is not the problem. The problem is a lack of good AI algorithms for games. AI in games is 20 years behind the state of the art. With the exception of NERO (a game built by AI researchers), most game AI is a case-based "expert" system and some flocking behavior to make things look more realistic. How many games out there use machine learning to adapt to the player? How many use reinforcement learning? Neural networks? Genetic algorithms? Evolutionary methods? Basically none. Game companies hire artists, graphics programmers, game engine programmers, but not AI guys. Why? Because they are so behind the times. If game companies would invest some cash in AI research aimed towards games (the closest thing going on right now is DARPA investing money in combat simulator AI), they might wind up getting some good AI. But seeing as no one can get any grant money or make any $$$ researching AI for games, games is stuck in the AI ghetto.
Manoel Neto @ Sep 6th 2006 7:33PM
Developers don't slap cheesy action-reaction AI on games nowadays due to lack of power. The real reason is that it's pretty hard to code good AI for something as complex as an action game.
I'd say doing "good" AI for games is far harder than doing academic AI, due to a major issue: there isn't a clear and obvious objective for game AI.
If you start coding an AI system for stitching multiple photos taken at random directions together to build a panoramic image, you want your AI to be as good and as precise as possible, and you can give it clear feedback for improvement.
But on games, "perfect" AI isn't desirable. If an enemy employs perfect tatics and aims perfectly, the game will become unplayable, and the AI will be labeled "bad".
Good AI for games is different than good AI for academic purposes. In games you want the AI to look good, enemies to react differently and avoid doing stupid things without sniping you from a kilometer away automatically. This dubious goal is incompatible with the AI being seriously researched, and thus lots of hack jobs are done.
Frozenhead @ Sep 6th 2006 7:41PM
No.
CPUs are plenty fast to do all the things this company is talking about doing. Any game that did decide to support this card would have to implement parallel AI to run off the card for customers without it. If the behavior doesn't match, you now how to balance two different games. AI is gameplay, and you can't buy a card to magically make it better.
tactics @ Sep 6th 2006 8:24PM
seems to me, they should team up with Ageia and make a card that has both the PhysX and this new AI chip imbedded together. that, to me, would make sense. some people are already adopting the PhysX thing, right?
go partnership!
-"superfan" tactics.
greatslack @ Sep 6th 2006 8:25PM
with a dedicated graphics card, sound card, physics card, and AI card, what's left for the CPU to do?
Psaakyrn @ Sep 6th 2006 8:35PM
I don't quite understand why a dedicated AI card is needed though, unless they found an AI algorithm which works with constantly repeated predetermined functions (and thus, would work better when hard-coded).. unless we're still going to use the (severely outdated) AI algorithms..
Sub @ Sep 6th 2006 10:05PM
So they want me to buy another card for my computer? I have a simple answer for them - No.
GhostDoggy @ Sep 7th 2006 5:06AM
I would attribute lackluster AI as a result of lackluster programming and not processing ability, folks. These hardware 'solutions' are more snakeoil than anything else. Good marketing sold Bose and now sells these kinds of consumer-desperate outlets.
Waccoon @ Sep 7th 2006 6:15AM
Good AI requires a lot of scalar power for decision-making. That's what CPUs do best. Coprocessors are mostly for gruntwork that CPUs don't have the bandwidth to handle.
I can't wait for the coprocessor controller card! It's just a giant heatsink attached to an inlet tube, that conveniently sucks money directly out of your wallet.
Psaakyrn @ Sep 7th 2006 9:51AM
There is something else that I just thought of, which an AI card would perform better than your normal CPU, but I seriously doubt this is it, based on our current technology (and IF it is it, it might as well replace our CPUs). Perhaps they figured out a method to match commonly-threaded logic patterns and subsequently increase the efficiency of using said patterns (the increased efficiency method being hardware based and thus cannot be done by software alone)? Naturally this would do wonders for AI of any sort, but it would also do wonders for just about every non-fixed non-random processing required, unless they haven't found a way to make this applicable to all possible calculations.
Luke Wallace @ Sep 7th 2006 9:31AM
3D accelerator cards were not required for every game, you just used Software Rendering if you didn't have one. It was the improved graphics and faster framerates that encouraged people to get one, but people with high end machines didn't have to get one if they didn't want to. I would imagine it could work the same way with physics and AI, have a "Software Engine" for those without the dedicated card, but those with it would see an improved frame rate because the dedicated card would allow more power to be used for graphics.
I don't see why it isn't viable. Now, would it be as profitable and prevalent as graphics cards within 2 years? That's a different question. Provide some games where it becomes noticable, and you'll get a definite Maybe from me.
Jason @ Sep 7th 2006 6:19PM
Suppose programmers implement some AI that would bog down a typical CPU. How do you turn it off? With graphics you can turn down the resolution, turn down the color bit depth, etc. AI affects the way the game plays, so it's not scalable like that. The result is that the pool of machines that can run this AI would be much smaller than the pool of machines that can run the latest high-end graphics. Developers would then have to be far more conservative with this new AI than they are with graphics, which means players might not even notice the improvements.