I have two GPUs in a PC here. One Nvidia GTX470 and one AMD Radeon HD5870. I run OpenCL on both cards, using AMD APP SDK 2.3.
The problem is that while the timings of the benchmarks I run on the GTX are really accurate, the same timings on the HD5870 are not. And this seems to increase as workload increases. The standard deviation of the timings are at least 10 times as large on the AMD. I use these cards in an auto-balancing scenario where timings are critical to achieve optimality, so this is a critical problem in my case.
I am/We are going to publish some results from this, and of course, I know the Radeon can perform better than it does now (because of wrong workload, because of faulty timings). I strongly believe this is a problem with the driver or runtime.
Is anyone on here acquainted with this problem, or similar behavior?