it's not a stability issue. It's memory management. GPU acceleration is a pipeline and the data decoded at the CPU is buffered into ram and then shadowed down to the GPU vram for GPU processing. This means Memory management is key to GPU acceleration performance and operation. The D700 cards have twice the Vram so the memory management will be vastly different with those GPU's versus the D500 cards. Allot of the memory management for GPU acceleration is decided by the Video drivers ie Open CL updates and driver changes along with the applications. The amount of system ram available is also a key factor. Buffers cannot be shared between cores in the GPU so data that is required for multiple threads have to be duplicated in new buffers. All of this is a fine dance and if the memory management is not maintaining synchronization then you are going to have issues. The more ram buffers created for the GPU processing the more data sent to the GPU at 1 time the exponential increase in data buffers that have to be duplicated. Adobe is still far behind with Open CL memory management profiles versus Cuda since Cuda was used years ahead of time by Adobe. This will be a growing process and 1 that will see many hurdles in the way because of Open CL changes by Apple with OSX and AMD drivers.
Eric
ADK
26
Replies
AdChoices