Best best is to probably run both the SMP and the GFX card. There are a few options for SMP. Stanford say you should run only a single SMP client, and state that the quick turn around is better for science. This is the better approach for science. You will get more points if you run 2 SMP clients at the same time, and even more if you fiddle with the core affinities. This is less good for science (according to stanford), but will net more ppd (quite a bit more). Finally, you could run Linux on a virtual machine, then run the SMP client on there for the best ppd figures, but this obviously involves quite a bit more effort. Depending on your clock and you exact config, you should get a few thousand ppd off the SMP client. Dont forget to add the -smp flag to the client, else you will be stuck with regular, non SMP units!
The SMP client should get your more ppd than the gfx card (ATI 4870s dont give much at the best of times, let alone a 3670). You need to set the priority of the GFX card to be higher than the SMP client else you will not get enough cpu time for the gfx client to run. I dont know what sort of figures you should expect (nVidia card here), but maybe an additional 1000ppd?
