Kepler support is required, I can run it on my 680 GTX for IR, but there are some 7xx series cards that don't support Kepler so it's not just GPU generation. However I can't run it for final renders, due to the 2Gb memory on that card. Naturally, the newer and more powerful the card, the faster the AI Denoising is. So, the GPU processor will determine speed, and the GPU memory will determine render size (regardless of GPU processor).