NVIDIA's CUDA GPU compute API could be making its way to practically every PC, with an NVIDIA GPU in place, or not. A chinese freelance developer has coded a means to get CUDA work as a middleware on OpenCL. This move lets CUDA work on ATI Radeon GPUs that support OpenCL, as well as x86 CPUs, since OpenCL specs allow the API to run on CPUs for development purposes. The implications of these are many:
- Letting CUDA-accelerated software such as Badaboom make use of ATI GPUs
- Letting PhysX run on ATI GPUs as PhysX middleware uses CUDA for GPU acceleration
- Possibly better scaling of PhysX on multi-core CPUs (over OpenCL), as the regular PhysX CPU acceleration is infamous for bad multi-core scaling in performance
The software works as a translation layer, exchanging calls between CUDA and OpenCL or the CPU if OpenCL is not available. It comes in the form of a loader application that injects itself into the executing process. To get PhysX to run, one needs to install older versions of PhysX System Software (version 8.09.04 WHQL
being the latest) from its standalone installer (installs PhysX libraries without looking for NVIDIA GPUs).
DOWNLOAD: CUDA Emulator Loader
This development could also have its implications on the industry, as not very long ago developers at NGOHQ.com successfully ran
PhysX on ATI Radeon GPUs. Something NVIDIA didn't object to
, seeing it as an opportunity to propagate PhysX and maybe highlight better performance on GeForce GPUs. AMD cold-shouldered that development and later announced its own plans to develop GPU physics processing with Havoc. The developer wishes to remain anonymous till such legal issues are ironed out. It was originally posted at jishuyi.com.cn
, a Chinese techsite, which doesn't seem to be reachable outside of the PRC - probably by China's protective networks. We will shortly publish a comprehensive review of CUDA performance on ATI Radeon GPUs.