Imagine two workers on an assembly line who work in sync to the ring of a periodic bell (the clock signal). The first worker finishes their task and hands the part to the second worker over a long conveyor belt (the combinational logic). If the conveyor is too slow and the process is delayed, the part will not reach the second worker in time for them to get ready for the next ring (a Setup violation). On the other hand, if the conveyor is insanely fast, the first worker may fling the next part and land it right in the second worker's hands exactly while they have not yet finished dealing with the previous task (a Hold violation). Correct, smooth timing requires a calculated data transfer – not too slow, but by no means too fast either.
Timing design in sequential circuits requires meeting two critical constraints to guarantee reliable and stable data transfer between flip-flops throughout the whole system. The first constraint is the maximum delay (Max Delay/Setup Path), which makes sure that the data manages to cross the critical, longest logic path and settle at the input of the next flip-flop early enough before the clock rises. The equation \[T_c \ge t_{pcq} + t_{pd} + t_{setup}\] describes this constraint, and from it we derive the maximum operating frequency allowed for the system \(\left(f_{max} = \frac{1}{T_c}\right)\).
On the other hand, the minimum delay constraint (Min Delay/Hold Path) is meant to prevent a "race" in which data rushes too fast through the shortest logic path and overwrites the previous data before the required hold time of that same clock edge has passed. For this we require the equation \[t_{ccq} + t_{cd} \ge t_{hold}\] to hold, which guarantees a sufficient contamination delay.
An essential difference between the two is that while a Setup violation can easily be worked around even after the chip is manufactured, by lowering the clock frequency, a Hold violation does not depend on the frequency at all (the variable \(T_c\) does not appear in it) and is a fatal failure. Fixing a Hold violation requires intervening in the hardware design itself, usually by adding artificial buffers on the short paths, which increase the path's \(t_{cd}\), without affecting the long path or the processor frequency.
On every clock pulse it fires two data balls: one into the long path and one into the short path.
Samples the data on every clock pulse. Each of its inputs has a laser door and an electric floor.
A winding orange route crowded with nodes (logic gates). Its propagation delay \(t_{pd}\) is the longest.
A straight, short cyan highway. Its contamination delay \(t_{cd}\) is the shortest.
The head arrives after \(t_{ccq}+t_{cd}\): the moment the data starts to change. The end of the trail arrives after \(t_{pcq}+t_{pd}\): the moment the data settles.
Closes \(t_{setup}\) before the next clock pulse. A ball whose trail has not passed it in time crashes into it.
Turns on with the clock pulse and stays on for \(t_{hold}\) after it. A ball that touches it during that time evaporates.
Every zigzag on the short path is a buffer that adds delay: both \(t_{cd}\) and \(t_{pd}\) of the path grow.
A button above the short path: faster gates over the same distance, so the \(t_{cd}\) of the path shrinks.
The clock wave. Before every pulse there is a red band \((t_{setup})\) and after it a blue band \((t_{hold})\). Cyan dot: arrival of the short ball's head. Orange diamond: the long ball settles.
Above each path its equation appears with the numbers. Green: the constraint holds. Red: it is violated.