Link:

Where PhysX allows physics interactions with a greater number of objects than software physics, AIseek says its Intia processor supports a greater number of intelligent non-playable characters than standard general-purpose processors.

To accomplish this feat, the Intia processor accelerates low-level AI tasks including movement, path-finding, terrain analysis, and line-of-sight sensory simulation. According to AIseek, Intia can run path-finding calculations 100-200 times faster than a standard processor running a software algorithm. Such speedups might not bring huge improvements to close-quarter first-person shooters akin to Doom 3, but AIseek says real-time strategy games and RPGs that often have many characters on-screen would greatly benefit from AI acceleration.

Whilst the idea sounds pretty good, I think if these guys release it as a stand alone card it will fail quickly. No ones going to go out and buy a graphics card, a physX card and an AI card unless you want to show off your epenis at a lan. Ultimately a quad core graphics card would be the go i rekon, dual core for graphics, 1 core for physX and 1 for A.I.

To get an idea of what the AIPU can do, I recommend checking out the Tank Battle videos on the AISeek website. I think a game like the upcoming Supreme Commander would rock with something like this implemented. promoted forum item

as you said its a good idea but buying 3 cards? seriously who would do that? and why couldnt fps games get a benefit from the acceleration? it just depends how well the developers created the game. im sure im not the only one who thinks that all the ai out in games now is just bulls***. sure they run when they are low on hp and take cover and such but other than that, it is no diffrent than what half life 1 ai was. we need to see some real challange in the fps genre. playing the hardest difficulty only changes how hard you get hit and how accurate they are. not their tactics. would be so great to see AI that would set traps for you and do the whole sam fisher thing, throw cans around corners and such to divert your attention then they jump out at you and attack.

last edited by DM at 16:57:19 06/Sep/06

What's the cost? If the price of a physics and AI cards was pretty resonable and greatly improved the potential of games I'd get them in a flash.

as you said its a good idea but buying 3 cards? seriously who would do that?

People also said that a seperate GPU was retarded. We now call those people morons.

People also said that a seperate GPU was retarded

Do you have anything to back up that claim?

GPUs have been around for a long time, long before they became mainstream due to GLQuake and Unreal. The required architecture is pretty much straightforward; lots of independent floating-point matrix/vector operations that can be accelerated with more pipes.

But with AI, the kind of calculations you would expect are very similar to what a regular CPU would execute (branch-heavy), with some dependent floating-ploint math mixed in. You might as well just dedicate yourself a core on a multi-core processor if you want to run some serious AI algorithms.

All in all, I think a specialised AI processor that you would have to go out and buy is a crappy idea. Good luck to them.

last edited by parabol at 17:29:29 06/Sep/06

Makes sense, AI processing is ridiculously slow/inefficient with regular processors given what's required. Dedicated hardware to do this makes perfect sense.

Do you have anything to back up that claim?

If only I could dig up archives from Rivazone.com

The S3 Virge SX/DX, Rendition Verite V1000 and the NEC PowerVR got slammed when they first were released, with everyone claiming they were useles. The argument against the Verite initially was "oh big deal, so ONE GAME [Quake] runs faster".
3Dfx suffered the same treatment for quite some time with Voodoo [1].

Then Verite V2000, V2100 and V2200, and nVidia Riva 128 came out... Enter the Microsoft vs nVidia war which almost sent nVidia bankrupt again over which would be used - Pixels or Voxels. (Microsoft had previously already driven nVidia into the ground when they were in the audio hardware acceleration business in the early 90s).

We will eventually see specialised hardware for parallel processing, but it's the chicken or the egg problem. I think they need to make the hardware more of a general solution that can be used for a wider variety of useful tasks. So far the Aegia card really hasn't impressed many people and for obvious reasons.

Do you have anything to back up that claim?

GPUs have been around for a long time, long before they became mainstream due to GLQuake and Unreal. The required architecture is pretty much straightforward; lots of independent floating-point matrix/vector operations that can be accelerated with more pipes.

It’s nearly a decade ago since I first saw voodoo cards at A3. I remember companies like Microsoft betting on software over hardware acceleration. At the time v1 cards where around $400 dollars extra for an already expensive machine and non gamers didn’t really see the point; many of them didn’t think it would succeed.

But with AI, the kind of calculations you would expect are very similar to what a regular CPU would execute (branch-heavy), with some dependent floating-ploint math mixed in. You might as well just dedicate yourself a core on a multi-core processor if you want to run some serious AI algorithms.

Generally speaking publishers don’t consider AI a real game selling attribute, if they did we would see a lot more of a focus on AI. In modern development we see developers implement AI last and as such, they often only get dregs of processor cycles that remain from the rest of the game. Basically the end result if often nothing more than really basic AI because that is all they can afford too do.

The argument of distributing load off of the CPU and onto specialised hardware means that developers will always have some sort of cycles that are dedicated to AI. By having more cycle time with a specialised processor AI developers can make very complicated AI algorithms into development and not have to worry about destroying gameplay.

All in all, I think a specialised AI processor that you would have to go out and buy is a crappy idea. Good luck to them.

It really depends on one thing, does it improve the gaming experience enough to make it worthwhile. People spend thousands of dollars on rigs already and if one or two developers in high profile games add enhanced content because of these cards then I’d imagine that people would buy them.

Heck, if supreme commander said “We are going to develop an AI patch that requires this card and it will allow for no unit cap and AI so complex that it doesn't have to cheat to be practically unbeatable” I’d buy a card the same day I bought the game.

Really, it’s the same in any field; a group of highly specialised units that work together generally out perform a group of generic units. CPUs should be about distributing processing load to those specialisations and doing any generic processing that isn’t worth distributing out. AI is so complex and involved that it is worth putting on a specialised processor.

A brief read of the whitepaper shows exactly how powerful this card is. It’s 100-200 times faster in path finding than software solutions like A* and 512 agents can scan 512 agents in 2 hundredths of a second and allows AI processing to understand sensory simulation based off of its environment, surveying and understanding terrain.

It could actually be revolutionary for game development.

whitepaper

Sounds cool. All comes down to price and game compatibility though.

Wonder if we will eventually get specialised slots for these sort of cards? Or motherboards arranged with certain pci express configurations specifically for physics/ai/whatever else.

A new revolution in AI would take some of the focus off of multiplayer in some games I think, AI up to this point has generally been so poor that people are forced to switch to fighting other humans just for entertainment. I think this could be great in first person shooter games where the main focus is co-op with lot's of AI enemies and battles to fight in, or for various cross genre titles like battlezone.

most fps games hardly feature co-op anymore. at least like i remember it like in quake 2. that was fun. why did game developers suddenly decide that no one wanted to play a co-op mode anymore? imagine fear co-op. just scale the amount of bad guys to suit and there ya go

But with AI, the kind of calculations you would expect are very similar to what a regular CPU would execute (branch-heavy), with some dependent floating-ploint math mixed in. You might as well just dedicate yourself a core on a multi-core processor if you want to run some serious AI algorithms.

When the first GPU's came out the same argument could have been made. I've implemented a (very s***ty... i spent about 4 days on it)software graphics pipeline it's not that big a deal.

The idea of AI-focused boards is not new... people were kicking around the idea of specialist neural net hardware 10 years ago. About 6 or 7 years ago i heard of a person who was trying to build boards specifically to fend AI onto. Granted my understanding was it was basically an old generation cpu in an expansion slot but Intia is definately not that... if they've got specialist hardware that makes breadth first search plausible then thats just fricken awesome.

Ultimately this is likely to be only the first step... I can envision AI cards containing advanced agent based technology whereby you setup the "brain" with your choice of bayesian/neural/whatever algorithms(Think of this as similar to OpenGL's "list" mechanism and feed it perceptrons and receive information callbacks.

Of course the biggest barrier to this is that there hasn't been a commercial entity that has been working on AISeek for eons and eons and is in the middle of a standards war.

most fps games hardly feature co-op anymore. at least like i remember it like in quake 2. that was fun. why did game developers suddenly decide that no one wanted to play a co-op mode anymore? imagine fear co-op. just scale the amount of bad guys to suit and there ya go

Two reasons...

1) because co-op involves adding to the design of the game system and if for some reason i was marking the designs that it seems like some games have, god knows, i'd be handing out fails like they were going rancid. And of course adding to the design of the game system means that you're spending money on something that doesn't appease the ignorant masses that think a game is "s***" if it actually runs on a card that is more than 15 days old because it "looks like arse".

2) Funnily enough part of the reason is because of AI. The people pimping Heavy Gear 2 pretty much stated that the reason co-op was removed was they had to add "a rediculous number of enemies" after a few players simply because there was so few cycles left for AI (and also paradoxically, the more enemies you need the less time you've got to spend on each individual bot both because of rendering/physics overhead and because you now have to divvy the piece of the pie dedicated to AI amongst more mouths)

Maybe if they make it a standard feature on a chipset - as you say, game developers go for the visual bling before AI.

Yeah... Standardised AI APIs (OpenAIL?) would do a lot for the cause, that way you can implement the API and it'll default to destroying your CPU if you don't have a card. It won't be long after MS releases DirectAI that AI boards will be a fait accompli.

What we obviously need is a 'gaming card' which includes GPU, physics and AI all on one board!@#

Nah then you'd have GPU, physics and AI all competing for pipeline bandwidth and processor time on an expansion board.

Of course if they released the GPU/Physics/AI pack containing 3 individual cards with out of the box SLI so that the CPU doesn't even need to communicate translation and orientation of the models of physics affected items and computer agents... i'll be there with $1000 of freshly stolen pension money.

well explained, cheers.

Xyzzy wins