Integrated Chipset not Property Reported

Accelerated Computing CUDA CUDA Programming and Performance
1

I am using a unibody MacBook Pro with the 9600M-GT and the 9400M GPUs.

Neither GPU reports itself as Integrated when querying the device properties. Is this correct? My impression is that both of these are integrated. I’m trying to use that field to decide when to use mapped memory or not. When I use pinned memory in this case, I’m getting a 20% performance improvement on my code, but in other machines where a discrete graphics card, it seems to hinder performance.

Is there any other way to make the determination on when this will be the best option?

Thanks,
-Brett

2

9400M should be detected as integrated, 9600 should not. Let me try to track down a new MBP.

edit: what CUDA version are you using?

3

I’m using the general release of 2.3 at the moment. I think this was released more than once, so it would be the later release. To improve my understanding of the situation: Why is the 9600M not considered integrated? (My impression had been that integrated simply meant that it was on the system board rather than a PCIe card).

-Brett

4

Integrated means that CPU memory and GPU memory are the same pool, not necessarily “not an add-in card.”

5

I own a MacBook Pro 17" 2.8Ghz (actual generation) with GeForce 9400M and GeForce 9600M GT and both are recognized, in Leopard and in Snow Leopard as CUDA-enabled GPU, I even use them in parallel to test for asymmetrical multi-gpu code and dynamic load-balancing :-)

You’d check to be in performance mode, not in maximum battery life (or you won’t see the GeForce 9600M GT)