Metastability and synchronizers: Worked examples — Hardware deep dive

This page is a drill. The parent note built the physics; here we hammer every kind of number the topic can throw at you. Before the examples, we lay out a scenario matrix — a checklist of every case-class — then work examples until every cell is covered.

Everything on this page uses one master formula. Let us re-state it so no symbol is unexplained.

Figure — Metastability and synchronizers

The figure above is the mental model for the whole page: MTBF plotted against slack $t_{r}$ on a log vertical axis. Because $ln (MTBF)$ is linear in $t_{r}$ , the curve is a straight line — every nanosecond of $t_{r}$ you add climbs a fixed number of decades. Steeper line = smaller $τ$ = a faster-resolving flip-flop.

The scenario matrix

#	Case class	What makes it tricky	Covered by
A	Baseline single-number MTBF	plug-and-chug, get units right	Ex 1
B	Add one flip-flop (2→3 FF)	$t_{r}$ gains a whole clock period	Ex 2
C	Change $t_{r}$ by a small amount	exponential sensitivity	Ex 3
D	Double the clock frequency	$f_{c l k}$ up and $t_{r}$ down — two effects	Ex 4
E	Solve backwards for $t_{r}$ given a required MTBF	invert the exponential with a logarithm	Ex 5
F	Degenerate: $t_{r} = 0$ (no slack)	limiting value, worst case	Ex 6
G	Degenerate: $f_{d a t a} \to 0$ (signal never changes)	zero input → infinite MTBF	Ex 6
H	Real-world word problem (button press)	translate English → symbols	Ex 7
I	Exam twist: compare two designs	fast-slow-FF trade-off, ratio $t_{r} / τ$	Ex 8

Every cell A–I gets hit below. Let us go.

Tool check: why a logarithm shows up

Several examples must undo the exponential — given an MTBF target, find the $t_{r}$ that produces it. The exponential $e^{x}$ answers "start at 1, grow by factor $e$ each unit of $x$ — where do I land?". Its inverse, the natural logarithm $ln$ , answers the reverse question: "I landed here — how many units of growth did that take?". So whenever $t_{r}$ is the unknown hiding inside $e^{t_{r} / τ}$ , we take $ln$ of both sides. That is the only new tool on this page; the rest is multiply/divide.

Ex 1 — Cell A: Baseline MTBF

Worked example Statement

A 2-FF synchronizer runs at $f_{c l k} = 50 MHz$ , with $f_{d a t a} = 1 MHz$ , $τ = 0.25 ns$ , $T_{0} = 0.5 ns$ . The slack is $t_{r} = 15 ns$ . Find the MTBF.

Forecast: Guess an order of magnitude first — seconds? years? longer than the universe? Write it down before reading on.

Step 1 — Convert everything to base units (seconds, Hz). $f_{c l k} = 5 \times 1 0^{7}$ , $f_{d a t a} = 1 \times 1 0^{6}$ , $τ = 2.5 \times 1 0^{- 10}$ , $T_{0} = 5 \times 1 0^{- 10}$ , $t_{r} = 1.5 \times 1 0^{- 8}$ . Why this step? The formula multiplies rates by times; mixing ns and Hz silently gives a wrong power of ten. Fix units once, up front.

Step 2 — Compute the exponent $t_{r} / τ$ . $\frac{t _{r}}{τ} = \frac{1.5 \times 1 0 ^{- 8}}{2.5 \times 1 0 ^{- 10}} = 60.$ Why this step? This dimensionless ratio ("how many resolving time-constants of grace we allow") is the only thing inside $e$ . It is the single most important number on the whole page.

Step 3 — Compute the prefactor $T_{0} f_{c l k} f_{d a t a}$ . $5 \times 1 0^{- 10} \cdot 5 \times 1 0^{7} \cdot 1 \times 1 0^{6} = 2.5 \times 1 0^{4} s^{- 1} .$ Why this step? This is the raw entry rate into the danger window (failures/sec before any resolving help). Units: $s \cdot s^{- 1} \cdot s^{- 1} = s^{- 1}$ ✓.

Step 4 — Assemble. $MTBF = \frac{e ^{60}}{2.5 \times 1 0 ^{4}} = \frac{1.14 \times 1 0 ^{26}}{2.5 \times 1 0 ^{4}} \approx 4.6 \times 1 0^{21} s .$

Verify: $4.6 \times 1 0^{21} s \div (3.15 \times 1 0^{7} s/yr) \approx 1.4 \times 1 0^{14}$ years — trillions of times the age of the universe. Units land in seconds (dimensionless numerator over a $s^{- 1}$ prefactor). Sanity: enormous, as a good synchronizer should be. ✓

Ex 2 — Cell B: Adding a third flip-flop

Worked example Statement

Same device as Ex 1 but now compare a 2-FF vs a 3-FF synchronizer. Each extra flip-flop adds one whole clock period $T_{c l k} = 1/ f_{c l k} = 20 ns$ of resolution time (recall the $T_{c l k}$ definition above). By what factor does MTBF grow going 2-FF → 3-FF?

Forecast: Guess the multiplier — 2×? 100×? Bigger?

Step 1 — Find the extra $t_{r}$ . Adding an FF adds $Δ t_{r} = T_{c l k} = 20 ns = 2 \times 1 0^{- 8} s$ . Why this step? The new flip-flop gives the previous one a full clock period to settle before anyone looks — that entire period is bonus $t_{r}$ .

Step 2 — Only the exponential changes. The prefactor $T_{0} f_{c l k} f_{d a t a}$ is untouched (same rates, same device). So the ratio is $\frac{MTBF _{3}}{MTBF _{2}} = \frac{e ^{(t_{r} + Δ t_{r}) / τ}}{e ^{t_{r} / τ}} = e^{Δ t_{r} / τ} .$ Why this step? Dividing kills every common factor — the messy prefactor cancels, leaving a pure exponential of the added slack. This is why "add a flop" is such a powerful, cheap fix.

Step 3 — Evaluate. $\frac{Δ t _{r}}{τ} = \frac{2 \times 1 0 ^{- 8}}{2.5 \times 1 0 ^{- 10}} = 80, e^{80} \approx 5.5 \times 1 0^{34} .$

Verify: The multiplier is astronomically bigger than the whole 2-FF MTBF was — one extra flip-flop turns "trillions of years" into a number with 30+ more zeros. Independent of the prefactor, which is exactly why designers reach for it. See Positive Feedback and Regeneration for why each period multiplies exponentially. ✓

Ex 3 — Cell C: Small change in $t_{r}$ , big change in MTBF

Ex 4 — Cell D: Doubling the clock frequency

Worked example Statement

Start from Ex 1 ( $f_{c l k} = 50 MHz$ , $t_{r} = 15 ns$ , $τ = 0.25 ns$ , prefactor terms as given). Now double the clock to $100 MHz$ . Assume $t_{r}$ scales down with the period: doubling $f_{c l k}$ halves $T_{c l k}$ , cutting $t_{r}$ by $7.5 ns$ to $t_{r}^{'} = 7.5 ns$ . Does MTBF change by a modest factor?

Forecast: "I doubled the clock so failures maybe double" — guess the real MTBF ratio.

Step 1 — Track the outside factor. Doubling $f_{c l k}$ doubles the prefactor $T_{0} f_{c l k} f_{d a t a}$ , which divides MTBF by 2. That is the small, linear effect. Why this step? $f_{c l k}$ sits outside the exponent, so it can only push MTBF around proportionally.

Step 2 — Track the inside factor. $t_{r}$ falls from $15 ns$ to $7.5 ns$ , i.e. $Δ t_{r} = - 7.5 ns$ . The exponent changes by $\frac{Δ t _{r}}{τ} = \frac{- 7.5 \times 1 0 ^{- 9}}{2.5 \times 1 0 ^{- 10}} = - 30, e^{- 30} \approx 9.4 \times 1 0^{- 14} .$ Why this step? This is the exponential term — a catastrophic collapse, dwarfing the factor-of-2.

Step 3 — Combine. $\frac{MTBF _{fast}}{MTBF _{slow}} = \frac{1}{2} \cdot e^{- 30} \approx 4.7 \times 1 0^{- 14} .$

Verify: MTBF didn't "halve" — it dropped by ~14 orders of magnitude. The exponential term (from lost $t_{r}$ ) buried the factor-of-2 (from higher $f_{c l k}$ ). This confirms the parent note's Forecast-then-Verify: faster clocks make metastability dramatically worse, forcing 3-FF synchronizers (Clock Domain Crossing (CDC)). ✓

Ex 5 — Cell E: Solve backwards for the required slack

Worked example Statement

A safety spec demands $MTBF \geq 1 0^{9}$ years. Given $τ = 0.3 ns$ , $T_{0} = 1 ns$ , $f_{c l k} = 200 MHz$ , $f_{d a t a} = 2 MHz$ , what is the minimum $t_{r}$ you must provide?

Forecast: A billion years is a huge target — will you need nanoseconds, or something impossibly long?

Step 1 — Convert the target to seconds. $1 0^{9} yr \times 3.15 \times 1 0^{7} s/yr = 3.15 \times 1 0^{16} s$ . Why this step? The formula lives in seconds; the spec is in years. Convert or the log misfires.

Step 2 — Rearrange the formula for $t_{r}$ . Start from $MTBF = \frac{e ^{t_{r} / τ}}{P}$ with prefactor $P = T_{0} f_{c l k} f_{d a t a}$ . Multiply up and take $ln$ : $e^{t_{r} / τ} = MTBF \cdot P \Rightarrow t_{r} = τ ln (MTBF \cdot P) .$ Why this step? $t_{r}$ is trapped inside $e$ ; the natural log is the exact tool that frees it (see the "Tool check" above).

Step 3 — Compute the prefactor. $P = 1 \times 1 0^{- 9} \cdot 2 \times 1 0^{8} \cdot 2 \times 1 0^{6} = 4 \times 1 0^{5} s^{- 1} .$

Step 4 — Plug in. $MTBF \cdot P = 3.15 \times 1 0^{16} \cdot 4 \times 1 0^{5} = 1.26 \times 1 0^{22},$ $ln (1.26 \times 1 0^{22}) \approx 51.0, t_{r} = 0.3 ns \times 51.0 \approx 15.3 ns .$

Verify: Forward-check — $e^{15.3/0.3} = e^{51} \approx 1.4 \times 1 0^{22}$ , divide by $P = 4 \times 1 0^{5}$ → $\approx 3.5 \times 1 0^{16} s \approx 1.1 \times 1 0^{9}$ yr $\geq 1 0^{9}$ yr. ✓ Reassuringly, a billion-year target needs only ~15 ns of slack — the exponential is generous.

Ex 6 — Cells F & G: The degenerate limits

Worked example Statement

Two edge cases the formula must survive: (F) zero slack $t_{r} = 0$ ; (G) a data signal that never toggles, $f_{d a t a} \to 0$ .

Forecast: For each, guess whether MTBF is zero, finite, or infinite.

Step 1 — Case F: $t_{r} = 0$ . Set the exponent to zero: $MTBF = \frac{e ^{0}}{T _{0} f _{c l k} f _{d a t a}} = \frac{1}{T _{0} f _{c l k} f _{d a t a}} .$ Why this step? $e^{0} = 1$ removes all "resolving help". You are left with the bare entry rate — the reciprocal of "how often you fall in the hole". This is the worst-case floor: MTBF cannot be smaller than this for the given rates. With Ex-1 numbers, floor $= 1/ (2.5 \times 1 0^{4}) = 4 \times 1 0^{- 5} s$ — a failure every 40 microseconds. Useless without slack.

Step 2 — Case G: $f_{d a t a} \to 0$ . Treat it as a limit, not literal division by zero. Hold everything else fixed and let $f_{d a t a}$ shrink toward zero: $lim_{f_{d a t a} \to 0^{+}} MTBF = lim_{f_{d a t a} \to 0^{+}} \frac{e ^{t_{r} / τ}}{T _{0} f _{c l k} f _{d a t a}} = + \infty.$ Why this step? The numerator is a fixed positive number; the denominator marches steadily to zero from above, so the quotient grows without bound. Physically: no data transition ⇒ nothing can ever be caught mid-flight ⇒ metastability becomes impossible, so the mean time between failures diverges to infinity. Writing it as a limit (rather than "divide by 0") is the mathematically honest way to say "never fails".

Verify: Both limits match physical intuition: no slack → protection collapses to the raw entry rate (finite, tiny MTBF); no transitions → MTBF $\to \infty$ . The formula behaves sanely at both boundaries. A signal that is genuinely static across the domain needs no synchronizer — but see the parent's mistake: rare is not static. ✓

Ex 7 — Cell H: Real-world word problem (a button)

Worked example Statement

A human presses a push-button connected to a chip clocked at $f_{c l k} = 25 MHz$ . A person can't press faster than about $f_{d a t a} = 10$ presses/second worst-case bounce, so take $f_{d a t a} = 10 Hz$ . The button feeds a 2-FF synchronizer with $t_{r} = 39 ns$ , $τ = 0.3 ns$ , $T_{0} = 1 ns$ . Is one synchronizer "good enough"?

Forecast: Human presses are rare and slow. Guess: is MTBF centuries, or worryingly short?

Step 1 — Translate English → symbols. "Async button" → the press is unrelated to the clock, so it is a clock-domain crossing. Map: $f_{c l k} = 2.5 \times 1 0^{7}$ , $f_{d a t a} = 10$ , $t_{r} = 3.9 \times 1 0^{- 8}$ , $τ = 3 \times 1 0^{- 10}$ , $T_{0} = 1 \times 1 0^{- 9}$ . Why this step? The hard part of a word problem is spotting which English phrase is which symbol; a slow human still counts as $f_{d a t a}$ .

Step 2 — Exponent. $\frac{t _{r}}{τ} = \frac{3.9 \times 1 0 ^{- 8}}{3 \times 1 0 ^{- 10}} = 130, e^{130} \approx 2.6 \times 1 0^{56} .$ Why this step? 130 time-constants of grace — thanks to a slow clock the period is long, giving huge $t_{r}$ .

Step 3 — Prefactor and assemble. $P = 1 \times 1 0^{- 9} \cdot 2.5 \times 1 0^{7} \cdot 10 = 0.25 s^{- 1},$ $MTBF = \frac{2.6 \times 1 0 ^{56}}{0.25} \approx 1.05 \times 1 0^{57} s .$

Step 4 — Convert to years and conclude. $MTBF = \frac{1.05 \times 1 0 ^{57} s}{3.15 \times 1 0 ^{7} s/yr} \approx 3.3 \times 1 0^{49} years .$ Why this step? Seconds are unreadable at this scale; years make the verdict obvious. Conclusion: one 2-FF synchronizer is massively good enough — in fact overkill. The slow $f_{d a t a} = 10$ and the long clock period both push MTBF absurdly high.

Verify: Order-of-magnitude check — the exponent $130$ dwarfs the tiny prefactor ( $0.25$ ), so MTBF is dominated by $e^{130}$ . Compare with Ex 7's answer against Ex 4's collapse: the danger cases are fast async data, not slow buttons. ✓

Ex 8 — Cell I: Exam twist — fast FF vs slow FF, the ratio decides

Worked example Statement

Two candidate flip-flops, same $T_{0} = 1 ns$ , and the same design rates $f_{c l k} = 100 MHz$ , $f_{d a t a} = 5 MHz$ (so the prefactor is identical for both):

Design P (fast FF, tight slack): $τ = 0.1 ns$ , $t_{r} = 4 ns$ .
Design Q (slower FF, generous slack): $τ = 0.4 ns$ , $t_{r} = 20 ns$ . Which is more reliable? The exam wants you to see that only the ratio $t_{r} / τ$ matters in the exponent.

Forecast: "Fast FF must win" — is that right? Guess the winner before computing.

Step 1 — Compute each exponent (the deciding quantity). $P: \frac{t _{r}}{τ} = \frac{4}{0.1} = 40, Q: \frac{t _{r}}{τ} = \frac{20}{0.4} = 50.$ Why this step? Same prefactor for both (identical $T_{0}$ , $f_{c l k}$ , $f_{d a t a}$ ), so MTBF is ranked purely by the exponent $t_{r} / τ$ — the dimensionless "grace in time-constants".

Step 2 — Compare via the ratio. $\frac{MTBF _{Q}}{MTBF _{P}} = \frac{e ^{50}}{e ^{40}} = e^{50 - 40} = e^{10} \approx 2.2 \times 1 0^{4} .$ Why this step? Dividing the two MTBFs cancels the identical prefactor, leaving a pure exponential of the difference in exponents. The winner is whoever has the larger $t_{r} / τ$ , and Q leads by 10.

Step 3 — Interpret. The slower flip-flop Q is about $2.2 \times 1 0^{4}$ (≈22,000×) more reliable, because its generous $t_{r}$ more than compensates for its larger $τ$ . Speed alone is not the metric.

Verify: For a full sanity check compute both MTBFs outright. Prefactor $P = 1 \times 1 0^{- 9} \cdot 1 0^{8} \cdot 5 \times 1 0^{6} = 5 \times 1 0^{5} s^{- 1}$ . Then $MTBF_{P} = e^{40} /5 \times 1 0^{5} \approx 4.7 \times 1 0^{11} s$ and $MTBF_{Q} = e^{50} /5 \times 1 0^{5} \approx 1.0 \times 1 0^{16} s$ ; their ratio is indeed $e^{10} \approx 2.2 \times 1 0^{4}$ . This confirms the parent's mistake #2 ("just use a faster FF"): MTBF depends on the ratio $t_{r} / τ$ , not $τ$ alone. See Setup and Hold Time and Timing Analysis for how $t_{r}$ is budgeted. ✓

The figure above plots MTBF vs $t_{r}$ on a log axis for both designs — notice both are straight lines (log of an exponential is linear), and Q's operating point sits far higher than P's despite Q's gentler slope, because Q's larger $t_{r}$ carries it up the line.

Recall Quick self-test

A design gains 2 ns of slack; $τ = 0.4$ ns. MTBF multiplies by? ::: $e^{2/0.4} = e^{5} \approx 148 \times$ . You double $f_{c l k}$ and lose 6 ns of $t_{r}$ with $τ = 0.2$ ns. Dominant effect? ::: The exponential: $e^{- 6/0.2} = e^{- 30} \approx 9 \times 1 0^{- 14}$ , swamping the factor-of-2 from $f_{c l k}$ . $f_{d a t a} \to 0$ gives MTBF = ? ::: Infinite — a never-changing signal can't be caught mid-transition. Two designs, same prefactor: which wins? ::: The one with the larger ratio $t_{r} / τ$ . What is $T_{c l k}$ in terms of $f_{c l k}$ ? ::: $T_{c l k} = 1/ f_{c l k}$ — the gap between clock ticks.

Connections

Parent: Metastability and synchronizers — the physics these examples drill
D Flip-Flop — the element with the resolving constant $τ$
Setup and Hold Time — the violated constraint that triggers it
Clock Domain Crossing (CDC) — why async data (Ex 4, 7) exists
FIFO Design — multi-bit crossing where these MTBF budgets are applied
Positive Feedback and Regeneration — why resolution is exponential (Ex 2)
Timing Analysis — computes the slack $t_{r}$ used in every example