Why is on sensor PDAF drastically slower than traditional PDAF? I assume the technology is the same, doesn't matter if the module is on the sensor or located behind a mirror. But the performance is not even close.
The technology in theory is similar, but the implementation isn't exactly the same.
Without knowing exactly which model, or even which manufacturer made your camera, it is difficult to be very specific. There's a large difference in performance between different models and lenses.
It always starts with light, so let's begin there. The amount of light used by each pixel of an on-sensor PDAF array is much less than the amount of light used by each pixel of a dedicated PDAF array.
Having said all of that, I highly suspect another big reason for the difference lies in the design of the lenses attached to the respective cameras. With traditional DSLRs, PDAF speed can vary quite a bit on the same body depending on the lens being used. The focus motors on lenses designed for the typical on-sensor PDAF cameras are more comparable to consumer grade DSLR lenses, and even fixed lenses on bridge cameras, than to higher end DSLR lenses. Overall design, weight of the elements being moved, the size and type of motor used to move the lens, and the firmware that controls it all have an influence on the overall speed and accuracy of the lens' mechanical focus system. If the on-sensor PDAF camera is even capable of interchangeable lenses, there may be faster lenses (aperture and focus drive) available for a premium. Remember that aperture is figured on actual, not effective, focal length. Some of the cameras in question, like the Fuji X100S, have APS-C sized sensors. When this is the case the aperture sizes are comparable to lenses used on APS-C DSLRs. But other cameras like the Nikon 1 series have 1" sensors with barely half the linear dimensions and less than 1/3 the surface area of an APS-C sensor, not to mention that they are dwarfed by the 7.5X larger 36X24mm full frame sensors. When comparing the lenses for these cameras, the absolute size of the diaphragm is smaller for a given f-number and angle of view because the focal length for that same angle of view is shorter. The mass of the optics in such lenses is less than that of more typical faster focusing DSLR lenses, but the space available for the motor that drives the movement of those optics is also much less.
Thanks to jrista for the information in his comments regarding this answer. Much of the correct information on this heavily revised answer is his. All of the mistakes are mine. B-)
Traditional phase detect autofocus system runs mainly in an open loop configuration: take measurement, send focus distance to lens. Contrast detect autofocus is closed loop, it has to take a contrast measurement, move the lens, take a contrast measurement, move lens etc. which is clearly much slower.
The reason phase detect on sensor isn't nearly as fast as a dedicated phase detect unit is simply that it isn't accurate enough to use in open loop mode, so the camera has to take several measurements to fine tune the focus.
The reason for the difference is accuracy is that a dedicated phase detect system has a series of lenses to direct the incoming light to two pixel arrays that are wide apart in order to measure the difference in phase and therefore the angle of incidence which tells you how far out the plane of focus is.
Phase detect on sensor uses a series of special pixels that have have an opaque mask over one half. By measuring the difference in light reading from pixels which have the left side blacked out compared to those which have the right side blacked out you get a measure of the angle of incidence but it's not very accurate as the height of the microlens above the pixel is small, leading to a much shorter baseline. Baseline is key to accuracy, imagine shutting one eye and trying to judge the distance of something by moving your head, if you can only move your head a tiny amount it's very hard.