To your first post, yes, you’ll have to use bumblebee or the new native support on newer kernels, although I believe those lack power management (for now, anyway).
Actually bumblebee is not a requirement to use CUDA at all. You can let the Intel GPU drive your display all the time and just install the nvidia drivers without creating an xorg configuration file and just modprobe the driver whenever you want to use cuda… see: https://wiki.ubuntu.com/Bumblebee
In regards to your other question, see:
This is from a laptop with an i7-4700MQ processor, and the bandwidth of the card is PCI-E 3.0 x8: https://devtalk.nvidia.com/default/topic/546357/cuda-programming-and-performance/sounds-like-gk208-laptops-cards-will-support-most-sm_35-features/post/3875737/#3875737
This is from an ASUS laptop with an i5-4500U processor:
(The video card’s bandwidth in that case is capped at PCI-E 2.0 x4)
My own Toshiba P50-ABT2G22 laptop with an i5-4200U processor also is capped at PCI-E 2.0 x4 speeds. So it is possible that (most?) U processors with NVIDIA chipsets are connected to a lower bandwidth PCI-E lane. Other than that, the clocks (and pipelines) on identical GPUs should remain the same across a same model card (e.g. GK208 GT740s should all have a clock around ~980 MHz and 384 cores)