clarry: The polling isn't done by the CPU, it's done by the usb host controller, which then interrupts the CPU if there is input. The protocol has very little impact latency; the slowest polling speed you get with USB is 125 Hz (that's 8 ms max), but goes up to 1000 Hz (1 ms). CPU usage is unaffected unless you type fast enough to actually generate 1000 key press / release events per second.
Regardless of the chosen wire protocol, both types of keyboards poll the keyboard matrix to detect pressed keys. Key switches also require some time to debounce. These are the primary source of latency.
erephine: For keyboards at any rate, PS/2 is purely interrupt based whereas USB is polled by the system.
Not sure about advertised polling rates and how they translate to response times, but at least as far as I've seen tests seem to fall into the ballpark mentioned above.
This is a good look at various keyboards in a controlled environment, the USB keyboards are significantly slower than their PS/2 counterparts (Table 2, Table 3).
Interrupt vs polling is inconsequential as long as the polling rate is high enough. Which USB can definitely achieve. The results from that 2001 paper do not show what causes the latency in the USB keyboards. But it isn't some problem inherent to USB.
