Performance is variable,
I've got an I7 8 core laptop. Tensor flow runs one the 'standard' handwritten number recognition rests in around 20 minutes on all 8 cores. This is a bit faster than on my ancient gx720. Caffe runs the same test in about 25 minutes on one core out of the box and a noticeably faster when compiled for the chipset and runs 5c cooler too. I didnt try recompiling Tensor because it didnt tell me it might be a good idea.
Now the GPU on the RaspberryPis is rated at 24Glops and my I7 is rated at 42 or so I'm looking forward to seeing that working ( or rather a box of Pi zeros working - in theory the PiZeros work out about twice as expensive as the cheapest Titan so they are value for money for up to a box load).
I must play with Snap and see if it is that much faster.
I cant think of any reason why Google would release slow software on the world tho...