Foundations — Metastability and synchronizers

This page assumes you have seen nothing. We build each symbol from a picture, in an order where each one depends only on the ones before it. Then the parent topic becomes readable line by line.

0. The map of what we must build

Read it top to bottom: heights → levels → the danger middle, and separately ticks → edges → windows. Those two streams meet at the knife-edge; we then name its speed $\tau$, watch it grow exponentially, and finally the exponential, the rates, and the time budget $t_r$ all feed the final quantity, MTBF.

1. Voltage — a number drawn as a height

Everything digital is really analog underneath. A wire carries a voltage: an amount of electrical "push", measured in volts (symbol $V$).

In the figure, look at the two tanks: the left one is nearly empty and the purple double-arrow measuring its water height is short — that short arrow is a low voltage. The right tank is nearly full, so its purple arrow is long — a high voltage. The key thing to notice is that the water could sit at any height between empty and full, not just those two extremes.

Why the topic needs it: metastability is defined by a voltage being stuck in between two allowed heights. You cannot understand "in between" until you can see voltage as a height that can take any value, not just two.

2. Two valid levels — what "logic 0" and "logic 1" mean

Digital circuits agree to only read two ranges of that height:

3. The thresholds — the forbidden middle

Between the low range and the high range is a middle band. A downstream gate reading a voltage in this band cannot honestly say "0" or "1". We need a symbol for the top of the low range and the bottom of the high range.

In the figure, notice the two dashed lines: the lower dashed line is $V_L$ and the upper dashed line is $V_H$. The coral band below $V_L$ is the safe "logic 0" zone, the mint band above $V_H$ is the safe "logic 1" zone, and the butter-yellow band between the two dashed lines is the undecided strip. Watch that middle strip — a metastable output is a voltage trapped inside it.

Why the topic needs it: the deviation must grow until the output crosses out of that middle strip and reaches a valid level (up to $V_H$ for a 1, or down to $V_L$ for a 0). The thresholds are the two finish lines of the race to resolve.

4. The clock — a heartbeat that says "look now"

Circuits don't read wires continuously; they read on a beat.

5. The clock edge — the single instant of decision

A flip-flop doesn't sample during the whole beat — only at the edge, the instant the clock jumps from 0 to 1 (a rising edge).

Why the topic needs it: metastability is caused by data moving at the same instant as this shutter click. No edge, no snapshot, no danger.

6. Setup and hold — the "hold still for the photo" window

A camera needs the subject to hold still slightly before and after the click, or you get a blur.

In the figure, the purple stair-step is the clock, and the dashed vertical line marks the edge at time $0$. Look at the coral shaded band straddling that line: the left coral arrow spans $t_{su}$ (before the edge) and the right coral arrow spans $t_h$ (after it). The whole shaded band is the "hold still for the photo" window — if the data wire changes anywhere inside that band, the snapshot is blurred.

Why the topic needs it: this window is the trigger condition for the whole topic. See Setup and Hold Time for the full treatment.

7. The metastable state — the knife-edge itself

Now the two streams meet. A blurred photo can leave the output stuck in the middle band (Section 3), balanced between valleys (Section 2).

8. The imbalance $v$ — measuring lean from the hilltop

Before any growth formula we must name the quantity that grows and fix its coordinate.

9. The time constant $\tau$ — how fast the lean grows

Why does the lean grow at all? A flip-flop's guts are a loop that amplifies its own imbalance — see Positive Feedback and Regeneration. Any tiny lean gets fed back and made bigger, again and again. We need one number for the speed of that feedback.

Now that $\tau$ is defined we may write the growth law.

10. The exponential $e^{t/\tau }$ — the shape of the escape

$|v(t)|=|v_0|e^{t/\tau }$

In the figure, follow the two rising curves: both start almost flat near the bottom (a tiny $|v_0|$) and then sweep upward faster and faster. The coral curve starts from a smaller $|v_0|$ than the purple one, so it reaches the mint dashed line ($|v|=\Delta v$) later — a smaller starting lean means a longer wait to resolve. Notice how the steepness itself grows as the curve rises: that ever-increasing slope is the visual signature of "growth proportional to current size".

11. The rates — $f_{data}$ and why events must coincide

12. The window constant $T_0$ — where it comes from

$T_0$ is not a fudge factor; it is a width of time rooted in the same setup/hold window you already met.

13. The resolution budget $t_r$ and assembling MTBF

Now the last symbol, then every letter is earned and we can derive MTBF.

Step A — rate of getting knocked onto the hill. From Sections 11–12 the number of dangerous near-coincidences per second is the product of the danger-window width and both event rates: $\text{knock rate}=T_0{f}_{clk}{f}_{data}.$

Step B — the distribution of the starting lean $v_0$ (stated and justified). Where does the data change lands inside the danger window is out of our control and evenly likely anywhere in it — the async signal has no knowledge of the clock, so no landing spot is preferred. That is a uniform distribution of the landing time across the window. Because the starting lean is proportional to how far from the exact centre the change landed ($v_0\propto$ landing offset), a uniform landing time makes $v_0$ itself uniformly distributed over a tiny interval around $0$. This is the one modelling assumption of the whole topic, and it is justified purely by "the async signal cannot favour any instant".

Step C — chance a knock is still unresolved after waiting $t_r$. A knock survives past our budget only if it needs $t_{resolve}>t_r$. From Section 10, $t_{resolve}=\tau \ln (\Delta v/|v_0|)>t_r$ rearranges to $|v_0|<\Delta v{e}^{-t_r/\tau }.$ So "still stuck after $t_r$" means the starting lean landed in a sub-interval of width $\propto e^{-t_r/\tau }$. Under the uniform $v_0$ of Step B, the probability of landing in that sub-interval is just its fraction of the whole window: $P(\text{still stuck after }{t}_r)=e^{-t_r/\tau }.$

Step D — failure rate, then flip it. A knock becomes a real failure only if it both happens and is still unresolved when sampled, so $\text{failure rate}=\underset{\text{Step A}}{\underbrace{T_0{f}_{clk}{f}_{data}}}\times \underset{\text{Step C}}{\underbrace{e^{-t_r/\tau }}}.$ "Mean Time Between Failures" is the average wait between failures = the reciprocal of that rate, which sends the exponent positive:

The real cause of asynchronous signals is Clock Domain Crossing (CDC), and a common structure that carries data safely across domains is a FIFO Design with synchronized pointers — both build directly on these foundations.

Equipment checklist

Test yourself — you are ready for the parent page when you can answer every line without peeking.