How much of a GPU would you need for this to be practical?
Doesn't really matter. What you would have to do to attempt a Finney attack is have a script that can spend your coins the instant that you find a block, withhold that block for half a second, and then hope that no one else solved that same block within that half second. Faster hashing just means that you don't have to wait as long to try it.
That's just it though. If you can rip off an eCommerce provider by 300BTC (10 * 30 BTC transactions) every 30 minutes because your GPU is capable of some 33% of the entire hashing power of the network, that's worth the effort. If you can rip off an eCommerce provider by 300BTC once every 6 months then it's no longer worth it, so I think hashing power is critical in the practicality of this attack.
Perhaps, but I would think that if this were to start happening, something could be done about it. I can think of a couple of different ways that repeatedly doing this kind of attack could bring the attention of the rest of the network to the attacker's IP address. It's not that an individual node cannot be discovered with the help of the majority of the rest of the network.