The Battle for Physics
When NVIDIA and Havok FX announced their partnership at the Game Developer’s Conference in 2006, another company was rolling out a product that had been in development since 2002. AGEIA’s PhysX PPU launched into retail channels shortly after GDC 2006, and has been the ONLY hardware-based physics accelerator on the market ever since. When it launched, there was a great deal of skepticism in the enthusiast crowd. These doubts were centered more around the affordability of low-end NVIDIA GPUs that would supposedly be able to do exactly the same thing as the PPU, with the added bonus of being able to do normal graphics duty when not needed for physics. Although this NVIDIA/Havok technology has yet to break onto the scene in any sort of purchasable form, the whole concept that it exists has likely hurt AGEIA’s success with what is otherwise one of the most innovative products released in the last five years.

AGEIA’s PhysX PPU, for all intents and purposes, is a very useful product for calculating the game physics in applicable games. The problem is, however, that there are currently very few games on the market that truly make use of AGEIA’s PhysX technology. An even bigger problem is that the games that are currently out have only recently been released, meaning that the PhysX PPU was dominating a market that had no real demand whatsoever. It cannot be disputed that the PhysX PPU is an amazing performer in games like CellFactor: Revolution that were coded from the ground up using the PhysX API. However, it also cannot be disputed that the games that are really going to take the PhysX PPU or other forms of hardware physics acceleration into the limelight have yet to be released.

With NVIDIA’s recent launch of its Tesla family of products, a reasonable analyst has to think that NVIDIA’s first real step into the Havok FX world is right around the corner. Their latest GPU, the G80 core, can be considered one of the most programmable and powerful graphics cores developed to date. Its unified shader architecture, the basis for what NVIDIA calls CUDA, means that all of the stream processors on the G80 core can work together on a single task and complete it up to 100 times more efficiently than a traditional CPU. Nobody doubts the computational ability of the G80 core, and nobody doubts the ability of big green to push their development package into center stage, but we have already been through more than a year of no physics capabilities from NVIDIA hardware, and nobody really seems to be doing anything about it.

The big problem with this battle for physics is that it doesn't seem to be happening. After GDC 2006, we kind of expected one of the two main combatants - AGEIA or NVIDIA - to come out as a clear winner in the near future. More than a year later, that never happened. Now there are more players in the game. Although never really "out" of the game so to speak, ATI has now become a strong player in the GPGPU/GPU Physics game. Using three ATI graphics cards in a triple-CrossFire setup - something that is possible on many CrossFire chipsets with three x16 PCI-Express ports - should yield a strong CrossFire graphics setup with one GPU dedicated to physics calculations. It would be nice if we could see this concept in action, but its cool nonetheless. AMD/ATI's future plans for something called "Fusion" is also a potentially huge factor in the future of in-game physics. AMD's initiative to integrate the GPU architecture onto the same die as the CPU not only promises to speed up the communication between the two processors (obviously), but also to bring about x86 instructions as a standard for GPU applications. Finally, Intel and their recently formed graphics division will probably be a pretty strong contender for top honors in GPU performance by the end of next year. Since Intel started the whole x86 ISA, it is reasonable to believe they would like to put their GPUs to work on the instruction set as well.

In the end, there is a ton of potential for general purpose graphics processing units, but there is limited ability to use that potential. A hodgepodge of different architectures and development environments are the problems that currently plague the rapid development of GPGPU applications. Physics engines such as Havok FX, which touted more than a year ago its ability to run to full effect on NVIDIA graphics hardware, has yet to prove itself in the face of battle. Meanwhile, a whole new kind of specialized chip has been reaping the benefits of accelerated in-game physics when a GPU is supposed to be able to do it just as well. For now, nothing is certain about the future of GPGPU. One thing that is certain, however, is that we, as a collective community, need to get organized and get behind the drive to push GPGPU computing forward.

---

Sources:
GPGPU.org Homepage
Wikipedia - Graphics Processing Unit
Havok FX
NVIDIA Tesla