Translation of Ageia's physics engine to CUDA (Nvidia)

I think it is a very good news for us
http://hardware.slashdot.org/hardware/0 ... 6231.shtml
http://www.tomshardware.com/2008/04/14/ ... g_almost_/

BTW: For the people that didn't know, Intel's Nehalem demo uses Ogre3d so perhaps someday we could have some benefit there too if they release an Ogre modified version optimized for multi-threading...

GPU based physics will be a stunning addition to PhysX and eventually working with Ogre.

But it's muggins here who will have to write the wrapping.

But it's muggins here who will have to write the wrapping.
They'll...
This change should be harmless to Physx library users.... I think the library itself will decide to run on GPU if it can.
In order to run a CUDA app (besides an 8xxx series card) you need specifics nvidia drivers so it will take some time be adopted by everyone.

Tell me if I am wrong, but how I understand CUDA is that it's a one way process; CUDA being a way to upload code to the graphics card and let it run, which the results are displayed on the screen as a form of polygons or textures.

How can those results be passed back to the CPU? Isn't there a massive bottleneck from reading from the GPU?

With that bottleneck present PhysX using CUDA can only be used for special effects only and nothing that needs attention or knowing what it's actually doing.

Tell me if I am wrong, but how I understand CUDA is that it's a one way process; CUDA being a way to upload code to the graphics card and let it run, which the results are displayed on the screen as a form of polygons or textures.
not exactly, with CUDA you can almost run any arbitrary code in the GPU and get the result in your app. In fact there is/was a BLAS/FFT library implemented in CUDA by Nvidia.
But of course that uploading only one float and calculate the square root and send back the result to the CPU wont be faster than doing it directly using the CPU.... and the reason is what you said: bottleneck
But when you have to do a very intensive calculus or a simple calculus but in a large ammount of data the bottleneck isn't a problem.

Perhaps, things like fluid simulations could take advantage of GPUs...