General-purpose computing on graphics processing units (GPGPU) is very popular right now, and NVIDIA has the lead in this arena with their Compute Unified Device Architecture (CUDA). While in the future, it looks like people will be moving from CUDA, which is proprietary to NVIDIA, to OpenCL, which should be available from a variety of vendors including NVIDIA. Right now, NVIDIA has the lead over rivals with CUDA.
However, there is a serious security bug in the NVIDIA CUDA Linux drivers that allows one user to read the files of another user. NVIDIA already knows about it, and has a patch available for it. It looks like the 270.18 beta drivers will have the patch, but you should want to check just to make sure.
We have recently found serious security breach in CUDA Linux drivers. The problem is related to cudaHostAlloc/cuMemHostAlloc API calls. In brief, driver maps pinned memory to user space but does not initialize it to zero. As an example, our simplest “proof of concept” program was able to read large fragments of files being written or read by other users.
via : Serious security issue with CUDA on Linux Serious security issue with CUDA on Linux
