If you compile with -arch sm_13 and then try to run on a Compute 1.0 or 1.1 GPU, then yes, I believe it will just fail to launch the kernel or maybe even earlier.
I’m using the latest production release of CUDA 2.0 [just released a few weeks ago]. I recompiled with the architecture specification, same result. The code actually runs, but simply fails to launch the kernel.