I just upgraded to CUDA 3.1 on my Mac Pro with a GTX 285 card. Everything seems to work fine, except that deviceQuery always reports a clock speed of 1.48 GHz, even when it’s obvious by the execution times on my application that it is running at a lower speed.
The “CUDA Clock Rate” application posted last fall in the NVidia GPU computing forums that allows you to increase the clock speed also always reports 1.48 GHz. I can still use this application to force the card into the highest clock speed, and I do see a corresponding drop in the execution times of my application, but the speed is always reported as 1.48 GHz
The drivers reported by the CUDA Preference pane are CUDA Driver Version 3.1.10 and GPU Driver Version 1.5.49.0 (18.5.2f16). I am running OS X version 10.5.8 (Leopard).
I am on 10.6.4 and actually deviceQuery ALWAYS show me the peak clock frequency of any of my GPU, no more the actual clock rate they are running. It’s troublesome when you are programming “warm-up” code to push the GPU at full speed before launching your real kernel (I write some benchmarks, so I need GPU to be a peak frequency for accuracy).
Thanks for responding with your experience. I found another thread that said that this behavior started with version 2.3a, although tmurray from NVidia did agree that this is a bad thing. Maybe this thread will remind NVidia of the problem.
I would like to be able to pre-compute peak performance on any GPU found, to pre-balance load, and to know when warm-up succeed, to reduce this phase when launching benchmarks kernel.
I would like to be able to pre-compute peak performance on any GPU found, to pre-balance load, and to know when warm-up succeed, to reduce this phase when launching benchmarks kernel.