Stand-Alone AI card

[wipeout140]

A new company called AIseek announced what it describes as the world's first dedicated processor for artificial intelligence. Called the Intia Processor, the AI chip would work in conjunction with optimized titles to improve nonplayer character AI. Similar to the way in which physics accelerators can make a game's environment look much more realistic, Intia would make the NPCs act more true to life.

Would you pay extra for Stand alone AI Card like the Ageia's PhysX card, could it be the future of games.

What do you think???

Extra Read on new stand alone AI Card

[OTZ]

You bet I would!

[Incinerate_IV]

I would if its under $50, and it actually make a difference.

For some reason I can't view the demo.

[krazykarl]

It seems as if the hardware industry has finally realized that hardware is better. You can get accelerated 3d graphics, 3d audio, physics, hardware network cards, storage processors and now AI chips. I think we are going to need a new form factor soon, with more expansion card slots.

[Lee1026]

Ok I do know I should be more tolerant of my fellow man and all that stuff, but really... this is just damned foolish.

Imagine the conversation that led to this...

-misty flashback fade-

Marketing Guy : Oh man, gaming is ready for a revolution!
Technical Guy : It's called a Wii now

Marketing Guy : Huh? We now what? -shakes head- I mean these gamers, they buy top end stuff, they have money to burn!

Technical Guy : Not really, they buy slightly under the curve and tweak up and overclock mostly
Marketing Guy : No no I read in a magazine that all gamers have more common sense than money
Technical Guy : -sigh-

Marketing Guy : These Ageis guys really whipped up a lot of frenzy about a new type of add on card.
Technical Guy : Yeah it's supposed to make the gamers run better by adding physics processing but the demo..
Marketing Guy : And they are making money hand over fist!
Technical Guy : Well, actually...

Marketing Guy : And it's so easy to make specialty stuff!!
Technical Guy : But their demo runs the same even without the card!

Marketing Guy : Wait, Wait, I got it! We'll make a card that adds more CPU power!
Technical Guy : Well dual cores add lots of CPu power that has yet to be tapped by games
Marketing Guy : No wait, even better, we'll make it special! That's what made the Ageis guys rich!

Technical Guy : Listen, the Ageis guys are not selling much, you might not want to...
Marketing Guy : We'll add better AI! That's IT!

Technical Guy : Better AI?
Marketing Guy : Yeah, we'll sell a card that makes the games run better!
Technical Guy : How's that work?
Marketing Guy : We'll umm, make it able to process AI commands like a graphics card processes graphics commands.

Technical Guy : But Graphics Commands are standardized, so they can optimize for that.
Marketing Guy : We'll get them to standardize AI commands.

Technical Guy : -twitches- But, every game has different needs from AI
Marketing Guy : So we'll make it flexible, generic, so it can do anything

Technical Guy : If it's generic processor design, it's the same as a regular CPU.
Marketing Guy : Exactly!

Technical Guy : But then what is it's advantage?
Marketing Guy : Haven't you been listening? It'll make games play BETTER!

[Lee1026]

It seems as if the hardware industry has finally realized that hardware is better. You can get accelerated 3d graphics, 3d audio, physics, hardware network cards, storage processors and now AI chips. I think we are going to need a new form factor soon, with more expansion card slots.

And none of them aside from graphics are actually selling. And none of them aside from graphics and audio actually does anything.

EDIT: Well, the raid controllers do something, but it can be more cheaply implemented with a additional core.

[HorseArcher]

I came across this website while I was reading around today: www.aiseek.com

Intelligence for New Worlds
A Whitepaper
AIseek Ltd.
7 HaBonim St.
Ramat Gan, 52462, Israel
www.aiseek.com

Introduction

Games today are better than at any time in the past. Thanks to the latest powerful
hardware platforms, artists are creating near photo-realistic environments, game
designers are building finely detailed worlds, and programmers are coding effects more
spectacular than ever before.

Unfortunately, this leap in graphics quality and richness of detail has not been matched
by a similar increase in the sophistication and believability of artificial intelligence (AI).
In the words of one leading developer, “characters still act like cardboard cut-outs”.

Naturally, leading studios are working hard – through sophisticated programming – to
enhance the behavior of non-player characters (NPCs) and other computer-controlled
agents. However, as we shall see, these efforts are severely limited by hardware
performance constraints. Thus the vision of a “living, breathing” world is yet to be
fulfilled – a world where thousands of NPCs move and act intelligently at all times,
within a large, constantly changing environment.

We argue that acceleration of low-level AI routines can address this fundamental
challenge. In doing so, AI acceleration will usher in the creation of entirely new game
worlds, offering a vastly enhanced playing experience.



Game AI Today

The interest in game AI is greater than ever. For example, in the pre-release excitement
surrounding a much-anticipated sequel, one leading RPG developer singled out a newly
built AI engine as one of the game’s hot features. The proliferation, in recent years, of
third-party AI middleware solutions is another indicator of this trend.
The AI Hierarchy. These in-house and third-party development efforts address a
variety of AI capabilities.
These capabilities form a clear hierarchy, with more
sophisticated functions depending on lower level building blocks. Thus, we will
distinguish between “high-level” and “low-level” AI. In general, high-level AI is
concerned with how NPCs ought to behave in various circumstances. This includes the
ability of an NPC to choose appropriate behaviors and goals in a constantly changing
world. In contrast, low-level AI focuses on performing the chosen actions and providing
input for higher-level behavioral decisions.

For example, consider an NPC seeing a city at a distance, deciding to go to the city to sell
an item, and then traveling to the city through the woods. The various mechanisms
mostly closely surrounding the taking of the decision (i.e., “I want to sell this item”) we
call high-level AI. On the other hand, the external acts of seeing the city (perception) and
moving appropriately (pathfinding) we call low-level AI. This example also illustrates the
hierarchical nature of AI; our NPC’s sophisticated decisions depend on lower-level
functions (pathfinding and perception). A similar analysis would apply to the case of an
NPC seeing an enemy, deciding to attack that enemy, and finally moving in the right
direction to attack.


Higher-level AI tends to be game specific and is created using many different methods
and tools. On the other hand, lower-level tasks – including movement, sensory simulation
(perception), and terrain analysis – are common across many different games and game
genres. Moreover, it is often these lower-level actions that are most obvious to gamers
and that most affect the believability of the game experience. If an NPC moves in an
awkward or aimless route from one point to another, stumbles into a static or dynamic
object or fails to notice an obvious change in its environment, the illusion of a life-like
world is lost.

The NPC Trade-off. Unfortunately, many of these lower-level tasks require CPUintensive
algorithms – the pathfinding algorithms used for movement are a classic
example. Hence, game developers are forced into what we call “the NPC trade-off”. This
compromise involves three dimensions: the number of NPCs, the intelligence of the
NPCs, and the complexity of the environment. Typically, a significant increase in one of
these dimensions will require constraints in one or both of the other dimensions. For
example, in order to populate a city with thousands of NPCs, all moving and behaving in
an intelligent fashion, we will be forced to constrain the complexity of our world.
Constraining world complexity means limiting ourselves to a rather small world, or to a
fairly static world, or both. Of course, with today’s sophisticated physics engines
introducing continuous, unpredictable changes into the game environment, the fiction of
a stable world is a thing of the past. This makes the NPC trade-off all the more acute.
Sophisticated developers may succeed in maintaining believability within the bounds of
the NPC trade-off, yet by doing so they are ultimately limiting game play and the game
experience itself. In order to create a “living, breathing” world, where thousands of NPCs
move and act intelligently at all times, within a large, constantly changing environment,
the NPC trade-off must be overcome.


We now review in greater detail the major algorithms and techniques in use today for
low-level AI. The areas considered include movement and sensory simulation. We also
discuss the problem of AI level-of-detail.
Movement. The current methodology for simulating movement is to use a pathfinding
algorithm (e.g., A*) to plan a path, another algorithm to smooth the path found, and
finally a steering algorithm to move a unit along the path. A* uses a heuristic to reduce
processing requirements, which essentially amounts to trying more promising routes first.

As a heuristic algorithm, A* must be tailored to specific levels and maps, yet even then it
may fail to find the right path within the allocated time. Moreover, for large, complex
maps, simple A* is not good enough and a hierarchical A* is used.
To accelerate computation at game time, much of the path data required by A* is
calculated before the game starts (e.g., waypoint graph, navigation mesh, tiered approach
etc.).
This optimization in the interest of performance seriously limits game play, since
the game map is constrained to change minimally and infrequently (if at all).

The steering algorithms used also present specific problems. One of the objectives of
these algorithms is to avoid collisions with obstacles and other agents along the preplanned
path. Highly dynamic games with a high object count use rudimentary collision
detection techniques, causing some objects to cross each other or to maintain
unrealistically large distances between them. With the imminent integration of high-end
physics engines, the shortcomings of these algorithms when dealing with a large number
of moving obstacles will require even greater design compromises.

Lastly, group movement remains a challenge. The movement of groups is often faulty
because the basic pathfinder does not take the paths of other units into account. Thus,
group movement solutions are generally programmed on a per case basis, resulting in
very limited group and formation behaviors.

Sensory Simulation. All games require a means for deciding what NPCs see to
allow them to interact with their surroundings. The methods used depend on the specific
genre and the concessions made possible by gameplay considerations. A common method
is to model each NPCs vision by casting a few rays in the virtual world and then checking
against other objects’ bounding boxes. Since only a few rays are cast many problems
occur, such as failing to understand partial cover. This means that it is not uncommon for
an NPC to see things that the player knows it should not, or to fail to see things that it
should.

Furthermore, these calculations are usually not fast enough for the needs of more
advanced algorithms, such as tactical unit placement and tactical pathfinding. These
advanced algorithms require many line-of-sight calculations for all possible NPC
destinations and for all possible covers. The alternative that many developers choose is a
pre-computed solution within a static environment. In extreme cases, genres that feature a
high NPC count, such as RTS games, have vision systems so rudimentary as to be almost
non-existent.

More sophisticated capabilities and human-like reactions must be programmed on a per
case basis. These include “camouflage” (the ability to hide on a similar-colored
background) and “attention getting” (bright or fast moving objects should catch the eye
and be detected faster).

AI Level-of-Detail. All the problems discussed above are aggravated by what is
known as “AI level-of-detail” (AILOD). Put briefly, this means that NPCs not
immediately on-screen are allocated less CPU time. This compromise can easily damage
the game’s believability. Consider a player entering a room, interacting with its (nonplayer)
inhabitants and exiting the room. Upon re-entering a short time later, the player
may find that nothing has changed, that the NPCs are exactly as he or she left them (or
performing random activities), in marked contrast to what can be expected in the real
world. This is a classic example of an AILOD-induced problem.


The Intia Processor: The AI Acceleration Solution

To address today’s AI challenges, and to enable the creation of entirely new game
worlds, AIseek developed the Intia processor. The Intia processor is the world’s first AI
acceleration chip. Computationally speaking, AI problems are typically search-intensive,
translating into large graphs with a high degree of branching. The Intia processor tackles
these specific processing challenges with a technological breakthrough – the Graph
Processing Core (GPC). The GPC lies at the heart of the Intia processor, delivering fast
and optimal solutions to graph-based computations.

The Intia processor is accompanied by a software development kit (SDK) for game
developers. Using the SDK, studios can empower their AI modules and games with
accelerated AI. The Intia processor’s accelerated AI functionality includes movement,
sensory simulation and terrain analysis. We will briefly review each one of these areas.


Movement.

The Intia processor features pathfinding functionality that is both optimal
and extremely fast. Unlike today’s software-based approaches (e.g., A*), the Intia’s
pathfinding uses no heuristics, thereby guaranteeing that the optimal path will always be
found. This optimality also means that the Intia processor avoids the common pitfalls of
A*, including failures to find a path when one exists, and the generation of “artifacts”
(e.g., weird, unrealistic paths). If a path exists, the Intia processor will always find it.
The Intia’s pathfinding is not only optimal, it is also extremely fast. Processing time for
each 100 nodes of path depth is only 10μs, making the Intia processor about 100-200
times faster than A*.

Most importantly, this very significant speed increase gives the Intia
processor the adaptability to support large, dynamically changing maps. Unlike
traditional approaches, which use relatively static, pre-processed maps, the Intia
processor supports maps that change continuously, with no pre-processing required. By
removing this limitation, the Intia processor enables the creation of new game worlds that
are based on large, rapidly changing environments.

The optimality of the Intia processor’s pathfinding includes excellent support for tactical
considerations. Thus, in finding the required path, the pathfinding algorithm can take full
account of any tactical information, such as the need to find a path that passes through
certain locations (e.g., hiding points, enemy positions).

To facilitate the integration of the Intia processor with existing studio AI modules, all
common game graph formats are supported. These formats including grids, navigation
meshes and waypoint graphs.


Sensory Simulation. The sensory simulation capability of the Intia processor features
fast, highly accurate line-of-sight functionality. Unlike current methods, the Intia
processor does not rely on simple point or box approximations for the bodies viewed.
Instead, the actual visible area is computed. Again, as in the case of pathfinding, this
increased accuracy eliminates “artifacts” (e.g. seeing through walls, failures to see an
object that should be seen, etc.).

In terms of speed, the line-of-sight checks are several orders of magnitude faster than
current methods: 512 agents can be checked against 512 agents in 0.02s. Once again, by
leveraging this immense speed, very dynamic maps are naturally supported.
Alongside the speed and accuracy of its line-of-sight checks, the Intia processor makes
available a sophisticated vision model. Under this model, the probability of detection
depends on the size and characteristics of the area actually visible. Thus, as in the real
world, small objects or objects that are farther away have a smaller probability of being
detected.

On the other hand, bright or fast-moving objects have a higher probability of
being seen. Excellent support for hiding and camouflage is also included.

Terrain Analysis. One area of AI that is particularly underexploited is terrain
analysis. In today’s games terrain analysis is usually performed statically during the game
design process; strategically important locations are identified once for the benefit of
NPCs and remain fixed throughout the game. However, the advent of dynamically
changing maps calls for a leap in terrain analysis speed and sophistication. The Intia
processor answers this challenge with advanced terrain analysis capabilities that can be
run in real-time to enable truly adaptive decision making.


To support such strategic planning at both character and game-wide levels, the following
built-in topology analysis functions are provided:
• T-connectivity: the area consisting of all points that can be reached at a given cost T.
• Critical Points: strategic locations, sometimes called “choke points”, that form the
bottleneck for getting in or out of an area.
• Controlling Regions: areas that permit strategic domination of the terrain.

Moreover, via the SDK provided, the terrain analysis abilities can be easily extended to
encompass custom functionality (e.g., wall building, ambush points etc.). In particular,
tactical information may be overlayed procedurally on the game map.
Summary

In the ongoing quest to build better games, gameplay, graphics and physics are
continuously being enhanced. However, one important area that has remained largely
untapped is AI – until now. With the Intia processor, developers can at last populate their
games with intelligent life on a scale impossible previously.

Moreover, game worlds can finally come alive with physics and other dynamic effects without compromising the
intelligent behavior of the virtual inhabitants of these environments.

Indeed, by freeing studios from crippling design constraints, the Intia processor enables the creation of
entirely new game worlds. Be it urban warfare with thousands of soldiers fighting
between exploding buildings, epic battlegrounds, or a life-like city where every man,
woman and child appears truly alive, accelerated AI promises a vastly improved gaming
experience.

About AIseek Ltd.
AIseek provides innovative technologies to power tomorrow’s computer games. The
company’s flagship product, the Intia processor, is the first dedicated processor for
artificial intelligence (AI). By accelerating and optimizing behavioral computations, the
Intia processor allows developers to populate games with intelligent life and to build
entirely new game worlds.

from this pdf

So, what do you think? If they will be able to pull it off and create a market for this, would this really help?

Look at this way.
Before graphics cards, the graphics was handled mostly by the CPU.
So we made graphics cards

Before sound cards, the sound was made by the PC "beeps" and stuff
So we made sound cards

the physics are handled by the CPU
so we made physics cards Although not so succesfull yet and not really utilized too many games (?)

So we now come to the AI.
The AI was always handled by the CPU. I always thought that there is no better thing to handle the AI, but the CPU. So if we move the AI to a separate card or add-on board, what will be the function of the CPU in games? Just an information Junction?

Do you think, with the not-so-hot success (so far?) of the physics card, this entire project might never takes off?

Reading over the document I posted, It does sound very promising, but I just don't know anymore...Can't we just have a better CPU , faster, more powerful, which would be able to process the AI, just like this add-on from that company?

[Incinerate_IV]

Thread merged. (Comon Archer)

[HorseArcher]

Thread merged. (Comon Archer)

ahh crap...I should read around before I click the new reply button..

[Lee1026]

So we now come to the AI.
The AI was always handled by the CPU. I always thought that there is no better thing to handle the AI, but the CPU. So if we move the AI to a separate card or add-on board, what will be the function of the CPU in games? Just an information Junction?

The AI have always been handled by the CPU because the AI is what the CPU is designed for. This entire business sounds like a plan to get investor money and not much else.

Reading over the document I posted, It does sound very promising, but I just don't know anymore...Can't we just have a better CPU , faster, more powerful, which would be able to process the AI, just like this add-on from that company?

AMD agrees with you. That is why they are buliding quad cores. That will be infinitly more useful then this junk can ever be.

[Incinerate_IV]

AMD agrees with you. That is why they are buliding quad cores. That will be infinitly more useful then this junk can ever be.

I agree, game designers and gamers are much better off going with dual/quad cores, since those stuff are actually widely avilable.

[Lee1026]

I agree, game designers and gamers are much better off going with dual/quad cores, since those stuff are actually widely avilable.

Yep. And the stuff from AMD is going to be easier to program for, as they use X86, and that is very well understood. No one is going to learn something new jsut to use those bloody cards when they can just use the 2nd or 3rd core.

[Lee1026]

Before sound cards, the sound was made by the PC "beeps" and stuff
So we made sound cards

These days, sound connectors are all that is left on the motherboard. All the work goes right back to the CPU. The trend now is that the CPU start reclaiming its roles, not the other way around.


Physics algorithms generally require lots of vector FP resources, so throwing specialized hardware at them makes some sense because mainstream CPUs -- even with their vector enhancements -- don't measure up. Most traditional AI algorithms, on the other hand, are extremely branchy and so require good prediction, caches, etc -- pretty much exactly what mainstream CPUs have and do very well at. It's hard to see how this can offer a real benefit, particularly when so many games are GPU- (rather than CPU-) bound already. Parallelizing AI behavior isn't trivial either because of shared state.

[HorseArcher]

These days, sound connectors are all that is left on the motherboard. All the work goes right back to the CPU. The trend now is that the CPU start reclaiming its roles, not the other way around.

I want that exactly. Since the giant increase in processing power, I just hope one day that all this crap I have to install in order to have a computer, should be all compressed into that CPU>

I meant, sound acceleration, and 3dgraphics.

I always thought that we are trying to go as small as possible, otherwise we would be going into a room filled with circuit boards and wires and call it a "PC".



..that is...not what I want.

Only TVs should be getting bigger, but lighter, everything else technological should be getting smaller.

[Peter]

Well its good if it becomes easier for the developers to create good AI, time will tell if this will be any good (how do we tell that the AI has been improved other then it "feels" improved?)

[Incinerate_IV]

The thing is when this card comes out, its probably going to be $200+ like the Phyx card, so almost no one is going to buy it. Why should developers waste effort to satisify 0.0001% of the customers who actually have that card? I think I'm with lee on this one, everything should be done on the CPU, theres nothing special about AI. Phyx card actually makes more sense, but that can be done with CPU as well. The newest version of Havok actually does everything the Phyx card can do. (including flowing liquid and cloths tearing)

[krazykarl]

The thing is when this card comes out, its probably going to be $200+ like the Phyx card, so almost no one is going to buy it. Why should developers waste effort to satisify 0.0001% of the customers who actually have that card? I think I'm with lee on this one, everything should be done on the CPU, theres nothing special about AI. Phyx card actually makes more sense, but that can be done with CPU as well. The newest version of Havok actually does everything the Phyx card can do. (including flowing liquid and cloths tearing)

The main issue with physX is that it has its own SDK, whenever M$ decides to integrate the physX sdk right into directX and programmers have to make far less effort to add the support for the chip in, it will take off. Same thing with this AI stuff, if they can get some actual silicone to market, we will still have to wait for M$ to put the correct API's into DX to get it widely adopted.

[Shaun]

Well, a stand alone AI card is certianly more useful than a physics, no sorry, Physx card. But for those of us with dual core, are they really necessary?

[krazykarl]

The best option would be to have the API run in either hardware or software mode, that way a game can support the hardware calls via software if the hardware is not present. This will allow a game using the technology to support a very wide range of systems, and not force people to spend money on hardware, whilest at the same time providing an option to those whom wish to purchase the hardware.

[Lee1026]

The main issue with physX is that it has its own SDK, whenever M$ decides to integrate the physX sdk right into directX and programmers have to make far less effort to add the support for the chip in, it will take off. Same thing with this AI stuff, if they can get some actual silicone to market, we will still have to wait for M$ to put the correct API's into DX to get it widely adopted.

The second problem with physX is that they are not much more efficient in terms of flops per transister. This means that AMD and Intel will be able to beat them at price for performance by simply adding more cores. And while programming more cores is hard, it is a lot easier to program for a X86 core connected by a extremely fast bus then it is for some strange stuff that no one have a lots of experience working with. While adding APIs to direct X would help improve things, they are still harder then to program the X86 core.

In addition, the thing that determines AI performance is cache, branching, and number crunching, which is precisely what the CPU is designed.

The best option would be to have the API run in either hardware or software mode, that way a game can support the hardware calls via software if the hardware is not present. This will allow a game using the technology to support a very wide range of systems, and not force people to spend money on hardware, whilest at the same time providing an option to those whom wish to purchase the hardware.

The best option would be to add more cores, so that different people can use them for different things.