|
||
|
|
|
              |
|
 |
 |
|||||||||||
|
 |
|
|||||||||
By William Van Winkle |
|||||||||||
The Power of Physics In a virtual 3D space, objects are usually made to adhere to a set of rules that emulate everyday Newtonian physics as closely as possible. Things have weight, hardness or fluidity, properties in how they break, and ways in which they move past one another. Everyone who finished college Physics knows that there are mathematical ways of determining how objects behave in space, but modeling these rules into a computer-based context is no small task. Once those rules are coded, the CPU historically has had the job of applying those rules to virtual objects in real-time. Game designers usually only provide for about 500 objects in their CG worlds because that's all today's upper-mainstream hardware can process at once without stumbling under the concurrent demands of background applications. Everything else you see is painted in graphically. It may look good, but it's a picture, not an object subject to manipulation. Think about when a car drives through thick fog. In the real world, the fog swirls as the car passes through it. In today's games, the fog remains static because there is no physical interaction between the car and the moisture particles. "Your eye is a very good filter," says NVIDIA's Andy Keane, general manager of visualization applications. "When you look at something, you won't know why it doesn't look real, but you know instantly that it's not. Something is missing. Physics is the cue your eye picks up to make things look real."
"We're talking about introducing titles with over 10,000 or 20,000 objects," adds ATI's Godfrey Cheng, director of marketing for multimedia products. "The whole gameplay changes. That's why it's taking these guys a lot longer to develop physics-tied titles. It's an order of magnitude different in the level of realism. The question now is how do you have 10,000 objects flying from an exploding wall without having it look cheesy." We've seen this before. When you need more horsepower and want to offload work from the CPU, you bring in another processor unit of some sort to shoulder the load. Fabless semico company AGEIA was founded in 2002 and spent three years silently developing what would become PhysX, the world's first physics processing unit (PPU). Two early PhysX designs from ASUS and BFG surfaced last spring, and market reception was lukewarm. The only serious game with the requisite PhysX support available at launch was Tom Clancy's Ghost Recon Advanced Warfighter, and while there was obviously more debris flying from explosions and falling from trees, actual realism was barely enhanced because debris objects still looked blocky and unnatural, and objects still vanished shortly after hitting the ground so the host system wouldn't have to keep track of them. Demos floating about showed better promise from fluid dynamics, but proof is in the FPS, which often suffered a terrible hit in moments of sudden, high-quantity object movement. AGEIA continues to revise its drivers, but competition is coming up fast from the two graphics titans, and with ATI and NVIDIA both backing the widely used middleware physics engine from Havok, AGEIA's prospects for 2007 seem shaky.
According to Andy Keane, NVIDIA swung into action on physics as soon as AGEIA starting dropping press releases, and by last March the company was ready to demo its first physics title, Hellgate: London, at the Game Developer's Conference. And where to find a co-processor? Why, one of the SLI GPUs, of course. Actually, NVIDIA's documentation shows that one GPU (GeForce 7600 or higher) can run both graphics rendering and physics. You don't see this publicized much, most likely because such an approach sells less hardware. But going forward, if you have a buyer who wants a value-rich gaming experience on a low budget, a single-card solution is perfectly viable. With two cards, the second GPU gets impressed for physics duty, and if no physics processing is required at the moment, it's circuitry gets reassigned to rendering—a thoughtful double-duty trick. The configuration NVIDIA is promoting most today is with three GPUs, meaning one GX2 card plus another single-GPU unit. The GX2 performs SLI rendering while the third chip does physics work. In the case of two GX2 cards (four GPUs), the fourth GPU would sit idle, an odd situation NVIDIA seems likely to address in the future. Unlike multi-GPU rendering, there are no cases where one can argue whether going with off-CPU physics yields a noticeable benefit. NVIDIA rendered a scene with 15,000 boulders on a 3.46 GHz Pentium 4 Extreme Edition 955, then ran the scene again on a 2.6 GHz Conroe chip with physics crunching inside of a 7900 GTX. The GeForce chip showed a 12X performance gain over the CPU. On the P4EE platform running one 7900 GTX GPU versus two, the dual configuration showed a 1.7X gain over the single graphics approach. Late in 2005, ATI also cozied up with Havok in order to put CrossFire on the physics job. Under all the corporate posturing, ATI's approach is very similar to NVIDIA's. While the initial development code from Havok varies from NVIDIA to ATI, the expectation is that the codes will coincide by the fourth quarter as both companies make the final preparations before Havok FX-enabled games arrive for the holidays. This is made possible through Havok FX operating as programmable code under Shader Model 3.0 (and later) rather than being tied to this or that hardware platform. But no, that doesn't mean FX is expected to run adequately under SM3-enabled IGP cores, such as the GMA X3000 in Intel's G965, or even entry-level cards like the GeForce 7300 or Radeon X1300. Moreover, both vendors are targeting first-gen implementations with one or two GPUs doing rendering while another one GPU does physics. ATI's shorthand for this is 1+1 or 2+1. (No word from ATI yet on having one GPU tackle both tasks.) The main difference between NVIDIA and CrossFire physics is that whereas NVIDIA is presently working within a two-slot construct, ATI is pushing for three slots right out of the gate. On one hand, this makes ATI's job a bit harder, because only one Intel board now exists sporting three slots, although ATI maintains that "four or five" other partners are pursuing three-slot designs. On the other hand, ATI is providing a more budget-friendly upgrade process for those who want to start with one GPU, because dual-slot SLI NVIDIA buyers will eventually be faced with purchasing a dual-GPU card if they want to run SLI rendering alongside physics. To NVIDIA's credit, though, $650 for the 7950 GX2 doesn't sound bad compared against two $500 7900 GTX SKUs. One lingering worry remains: Shouldn't the AGEIA PPU, designed from the ground up explicitly for the task of running physics code, prove more scalable and suited to physics work than an ATI or NVIDIA GPU that has more or less been shoehorned into the task? This is a big deal, because no reseller wants to urge buyers into a solution that's going to be ridiculed throughout the market in under 12 months. Yet NVIDIA claims that, months before official release, it is already capable of processing more objects than AGEIA.
"Yes, GPUs were designed for rendering, but rendering has become an increasingly programmable process," says NVIDIA's Keane. "It's like, was a CPU designed to run a database? No, it was designed to run instructions. It's the same with a GPU. It runs instructions. Now, the CPU and GPU are both very good at certain types of processing, but they're different. What we've done is take the parts of physics that the GPU is very good at and put it on the GPU. You're not shoehorning stuff into the GPU; you're mapping over what it's good at." We don't think the strong reseller story is in ATI and NVIDIA's top-end parts. Sure, some resellers manage to attract enthusiasts, gamers, and those who don't mind dropping three grand on a machine to play Tiger Woods PGA Tour. (Imagine the realism possible through physics processing when you watch someone smash out of a sand trap!) But a lot of that business goes through etail. Step over into a tier-one like Dell, which currently offers the AGEIA PhysX PPU (without specifying whose card it is—tsk, tsk—although the 128MB of memory probably pegs it as the BFG product), and you won't find options on a "gaming" PC below a 512MB Radeon X1900 XTX or dual GeForce 7900 GS cards. This all-or-nothing approach from the big players leaves a lot of gray ground in the middle, and this is where system builders can shine. "As developers use more and more GPU-based physics effects, you want a more powerful GPU," says NVIDIA's Keane. "In simple games, the difference between a $99 card and a $399 card will be nothing. But it's just like graphics. Both cards will run the game, but if a developer targets a feature that's not in the $99 card, you don't get to see it. There's a scalability in physics just like graphics. That said, I think the first round of physics-enabled titles will target the mid-range."
That would be the mid-range that the OEMs are ignoring in a gaming context. The sweet spots here are currently the GeForce 7600 and Radeon X1600 parts, which span from $99 to $199 street prices. A user could get three X1600 XT cards for the price of one CrossFire Edition X1900 card, and we'd love to put that little head-to-head in action on a test bench. (Note that ATI won't support three GPUs working together on rendering for several more months.) The move to parallelism may just reshape our preconceptions about performance-per-dollar. And if mainstream cards ultimately do deliver a stronger value proposition for multi-GPU rendering plus physics for mid-range buyers, then channel system builders stand best poised of all to capitalize on it. "We think the titles that are about to come out will allow the channel to get to market quicker than the OEMs and deliver more flexible configurations," says ATI's Cheng. "When you start adding physics on a third card or a fourth card, OEMs are not going to want to qualify that. So system builders will be able to run circles around the OEMs and address whatever the phat, exciting needs are of the moment." ...more |
|||||||||||
 Back to top
|
|||||||||||
Copyright © 2007 RAM Magazine. All rights reserved.
Do not duplicate or redistribute in any form. |
|||||||||||
