Slashdot Mirror
All GeForce 8 Graphics Cards to Gain PhysX Support
J. Dzhugashvili writes "Nvidia completed its acquisition of Ageia yesterday, and it has revealed exactly what it plans to do with the company's PhysX physics processing engine. Nvidia CEO Jen-Hsun Huang says Nvidia is working to add PhysX support to its GeForce 8 series graphics processors using its CUDA general-purpose GPU (GPGPU) application programming interface. PhysX support will be available to all GeForce 8 owners via a simple software download, allowing those users to accelerate games that use the PhysX API without the need for any extra hardware. (Older cards aren't CUDA-compatible and therefore won't gain PhysX support.) With Havok FX shelved, the move may finally popularize hardware-accelerated physics processing in games."
(Think! Why would NVIDIA waste expensive chip real estate for stream processors if they weren't useful for 99.9% of the applications running on these chips?)
...what will be calculating my 3D images, if the GPU is already working on the physics? It is not like there is so much spare capacity left over in modern games anyway... FTFAPhysics covers a lot, from gravity, inertia, particles, collisions, IK and various other bits and pieces. Not everything lends itself to acceleration. So what will be accelerated by this?
The RadeonHD driver *is* GPL'd, and all the specifications necessary for writing your own drivers from scratch are in the process of being released. Significant amounts have already been released after being checked out by AMD's lawyers.
And the closed-source, binary module is still making progress while all that other stuff happens.
Whereas for me, ATI hasn't had usable XV support since 8.35.5 (or thereabouts), and the 3D rendering is buggy as hell... which kinda defeats the purpose of using a dedicated GPU. Go look at the known issues in the release notes - it reads like an alpha dev-snapshot. I regret fitting ATI to my laptop for the sake of a supposed performance advantage over the Nvidia option, and my next machines will absolutely be NVidia Quadro (Sun workstation, yes Nvidia even provides drivers for Solaris x86) or Intel integrated (ultraportable notebook).
Forget thrust, drag, lift and weight. Airplanes fly because of money.
Not quite true. They said they'd release the specs so said driver could be created but they only released partial specs for the 2D aspects of their chips. They still haven't released updated doc's for 3D/Video rendering etc
I had an ATI card once. The open source drivers (no mode switching) worked better than the binary ones (black screen). And the binary ones complained that, despite buying a card in a box plastered with "RADEON" and the ATI logo, it was not an ATI card, and therefore unsupported. Fuckers.
I've been a happy owner of NVidia cards ever since.
On the CUDA forums, we've gone back and forth about this, and the diagrams that people base this statement on are backwards. There are 16 multiprocessors (to use the NVIDIA terminology), each with 8 stream processors per multiprocessor. The 8 stream processors on each multiprocessor run the same instruction at once, but on separate register files. Multiprocessors, however, are completely independent, so in principle, one could imagine partitioning the resources between physics simulation and 3D rendering. This sort of partitioning has not been made available through CUDA yet, but hopefully this means we will see it soon.
You are correct that these 128 stream processors (however you slice them) are the main compute engine. There is additional circuitry to do hardware accelerated video decoding, but NVIDIA has not exposed that functionality to 3rd party programmers, and it isn't used during 3D rendering.