Slashdot Mirror
World's First Physics Processing Unit
Duane writes "Gamers Depot has an exclusive interview with the team behind Ageia - the maker of the world's first Physics Processing Unit (PPU) - which was just announced today.
"Sure we've all heard about the CPU and GPU - that's old hat by now and as most hardware reviewers will tell you, it's about time we got something that's truly revolutionary. Yeah, Pixel shaders are cool, and can do a lot of really nice things; however, pale in comparison in scope to what the PhysX chip from Ageia has the potential to bring to gaming.""
GS: How will Epic's Unreal Engine 3 incorporate Ageia's technology?
TS: We've been collaborating with Ageia since their inception and Unreal Engine 3 thoroughly exploits the Novodex physics API; when the first Unreal Engine 3 based games begin shipping in early 2006, they will really up gamers' expectations. The combination of next-generation graphics, next-generation physics, and content-rich games goes way beyond current games, both qualitatively and quantitatively.
mmhmm. I guess epic 3 the v4p0rw4r3
Someone explain to me why I need to purchase a *slower* processor with less ram to do it for me?
For the same reason people purchase graphics cards with slower clock speeds and less RAM to compliment their blazing fast processors. As the article explains, the CPU is a general purpose chip. A PPU will be fully dedicated to physics, and therefore likely far more efficient. By your logic, all processes in the machine should be handled by a single chip, which while elegant, is probably not the most efficient solution. I predict we will see more specific-purpose chips being developed, not fewer.
It looks like we're heading there already. Havok has already developed a mature software physics engine which is used in many popular games. I think in this case, developers are willing to give up a little control on physics to have better looking effects. This PPU sounds like it's designed to hook right into Havok, and could really prove useful as Havok becomes more popular.
Your video card's GPU runs at a slower clock rate than the CPU, but because its pipelines are completely optimized for T&L and triangle filling, it can do those tasks faster than your CPU ever could. Likewise, a physics processor is optimized for simulating the dynamics of a mechanical system.
I would believe that most developers have their own method for dealing with physics - from simple collision to ragdoll and the like.
Basic collision detection methods are bounding planes, spheres, capsules, and axis-aligned boxes, along with Binary Space Partitions, Quadtrees and Octrees combined with particle systems. It would be fairly straight forward design an instruction set to perform these operations between the simple primitives (spheres, planes). But BSP Trees, Quadtrees and Octrees would require a high level data format.
If all the collision testing could be done within a single thread within the time limit of a single frame, it would be no different from the player-missile and sprite graphics implemented on early home PC's (Atari computers could do hardware base per-pixel collision detection). Although, it would probably seem easier to have additional vector processors like Sony's Cell processor.
Vintage computer adverts: http://www.vintageadbrowser.com/computers-and-software-ads
Mmmmm. Apparently there is an open source lib for physics.
::shudder::
http://ode.org/
But your point about a standard like OpenGL not existing is true. We'll probably have a rehash of the early graphics library incompatabilities again.
You think people would learn. Open standards help a new technology to expand and to become accepted. It helps *everybody* in the industry.
"Ignorance more frequently begets confidence than does knowledge"
- Charles Darwin
This matters at the physics level. If you are going to fully implement the ballistics you are going to have implement the motion of the bullet, the atmospheric drag on the bullet, the gyroscopic stabilization, the effect of gravity on the bullet ("bullet drop") not to mention the effects of the individual specifications of the bullet itself, and perhaps some entirely random factors (the world isn't perfect).
And if you are implementing a game where players can fire an assault rifle full-automatic (600-700 rounds a minute or more, depending on too many factors to list - which might need to be implemented and calcuated by the computer, of course...) you can see that the CPU is going to start needing some help to work it out.
And that's just the bullets.
The gun example is just an example of the sort of jobs a co-processor might be required to do in an FPS environment. To cut a long story short, if you are going to be simulating life, even a small approximation of life, accurately, you are going to need to be calculating an awful lot of physics.
From http://www.word-detective.com/032602.html
Dear Word Detective: I work in the field of data processing and computers. Often I will come across new standards or proposed standards that have arisen from a company's latest "white paper." The term is also common in politics and government in referring to a formal declaration of policy or viewpoint. A friend at work tonight asked me if I knew the origin of the term. As a wild guess, I told him it might refer back to a time when paper was a relatively expensive commodity and rough drafts and notes were written down on odd scraps and cheaper, darker paper. When a final, formal version was to be drawn up, it was on the more costly bleached "white" paper. Am I even close? --
Close? Well, let's just say, "No cigar for you, buckaroo." But that's actually not a bad guess, since paper was at one time very expensive. The actual origin of "white paper," however, is a bit less logical than that.
Today, as you note, we use the term "white paper" to mean a formal statement of governmental or political policy that includes an extended explanation of that policy, usually accompanied by data and statistics compiled to make the case for whatever the policy is. The U.S. State Department, for instance, is fond of issuing "white papers" on various political "flashpoints" around the globe, usually shortly after the U.S embassy there has been torched.
As tedious as I'm sure governmental "white papers" may be, the term originally arose in the context of something apparently even more snooze-worthy. "Blue Papers" in the 19th century (so-called because of their blue covers) were humongous policy or legislative statements delivered by the British government for consideration by Parliament. But if a report or statement was too brief to be rightly considered a "Blue Paper," it was issued with white covers, and, with uncommon logic, called a "White Paper." Probably because these pithy "White Papers" were more directly useful than the bloated "Blue Papers," Americans adopted the term and have been using "white paper" since World War II to mean "background report," whether in the governmental or business realm.
You have no idea what you're talking about. Will multithreading make GPU's useless? No. This chip will have over 100 Gflops of processing power, and will handle floating point calculations in a way that a PC processor with around 10 Gflops of processing power, mostly focused on integer calculations cannot.
They came over to our office and did a demo a few months back.
There's no hardware yet so it was more like a software showcase of what COULD be happening. You have your basic explosion and rag doll stuff. The ones that showed best are a corridor flooding with liquid fire, a dense system of cogs and gears that worked flawless, and a tall building that collapse with thousands of pieces of debris bouncing off each other. The physics can be turned on and off in real time and will in turn generate different outcome.
We questioned the process of integrating the chip into the market. It will be the chicken and egg conundrum. The manufactures and developers will both wait for the other to create the demand for it.
The most straight forward solution seems to be to convince a console developer to include it in their next gen console.
- Collision detection. You have some representations in the world- are any two of them intersecting? The answer to this depends on how they are represented and can be fast (boxes, cylinders, spheres) or expensive (polygon soup) or anything in between. The triangle algorithm you describe is indeed simple, but it's usually not feasible to run it on every triangle pair in the world in every frame- which is where the scene graph and various methods for trivial rejection come in.
- Collision response. Once you have found an intersection, how do you correct the world state so that the objects are a) no longer colliding b) exhibit the correct behavior in response to the collision over the following state updates? A large part of this is quantifying the exact nature and magnitude of the overlap between objects, and for anything that isn't a primitive shape this can be complex.
The complexity of physics in games is constantly increasing, which is why more and more companies are turning to providers like Havok who work on these problems full-time (instead of trying to solve them and make the rest of the game simultaneously, as an in-house solution would require).That was a parody of Winston Churchill, moron.
The AACS key is NOT 0xF606EEFD628B1CA427BEA93A9CA9773F
I don't get what is so different between a GPU and this PPU thing. A GPU is mainly multiplying vectors and matrices and dot products and division, physics simulation is not very different, although the result is handled differently.
Quite differently. You are suggesting that one draws the matrix results and the other just stores the matrix results, but there is a more important factor than that. All the data that is pushed onto the grpahics card is essentially on a one-way trip. After going through the T&L pipeline, it guts pushed into a video buffer, drawn and then overwritten. I'm not even sure that there is capacity to write back data to general memory from the AGP cards, since this incurs a big performance hit. Any physics chip will need to be able to matrix math and then save the results for further useage, since you need to maintain things like mass, velocity, heading, etc.
Using a GPU chip as the starting point for a PPU wouldn't be an entirely bad idea, though. Interestingly enough, I think the rendering community would love to have a 3d accelerator that could use specialized GPU rendering speeds and the ability to write the image generated to the hard drive, so it's not like you are the only person thinking along those lines.
HA! I just wasted some of your bandwidth with a frivolous sig!