At my company we are right now working to speed up our calculations with GPU acceleration. We are using OpenCL and aim to buy a Tesla card in a few weeks. For now we started developing on a Quadro K600, a K1100M and a K2200 card. We observed that the same program which runs with constant memory on the K600 and K1100M has a large memory leak with the K2200 (around 10 MB/s). We use the driver version 376.84 which should be the latest enterprise driver and we already tried to switch to the latest consumer driver.
If we only write and read from the GPU without calling clEnqueueNDRangeKernel or if we reduce the number of CPU worker threads which queue the kernels on the GPU then the memory leak can’t be observed on the K2200. We have a simplified example available based on the NVidia oclMultiThreads example.
Is there a known issue with the OpenCL driver for the K2200? Or what is the recommended way to report that issue? We are worried that if we give our program to our customers, then they might face the same issue and the process is intended to run for hours and this memory leak would prevent that from happening.