In the demo, the each flag is made up of 50×50 vertices. It’s even worse, GPU PhysX getting slower than CPU PhysX!

As soon as you use a grid with more vertices (80×80 for example), there’s no difference between the GPU and the CPU. To get the max performance, the grid that shapes the flag must have around 2500 vertices (50×50) to fit in GPU shared memory. Here are some framerates for the GTX 780 testbed (Intel i5-4670K):Īnd for a GeForce GTX 660 with 5 SMX (then 5 flags) we have:īut there is a constraint: the number of vertices of the flag. The GeForce GTX 780, for example, has 12 SMX, that’s why the demo creates 12 flags to fully exploit the GK110, each cloth being simulated by a SMX (192 CUDA cores). To take advantage of GPU PhysX for clothes, a PhysX application should create more cloth instances than the number of SMX (streaming multiprocessors). But good news, a trusted source told me that GPU PhysX is also planned for Linux and OS X (the bad news: the release date is not yet planned…). And GPU PhysX is available only under Windows.

GPU PhysX in the current PhysX 3.3 SDK is only available for cloth and particle modules. On OS X, only the CPU mode is available…Īnd the Linux version ? It should be available shortly, as soon as I will have fixed a linking issue with the PhysX libs. On Windows, the flag simulation can be done with the CPU or the GPU (you can switch between both modes with the G key). This demo is available for Windows 64-bit and Mac OS X 10. The demo is coded in Lua and shows the use of Lua modules. The demo also displays some information about the PhysX device using functions available in PhysX 3 SDK. This demo, based on the latest beta version of PhysX 3.3, uses the cloth module of the PhysX engine and shows animated flags made up of 50×50 vertices. Here is a small PhysX demo for testing purposes, created with GLSL Hacker.

For mac download Geeks3D FurMark 1.37