Clock skew variation is the spread in clock skew between two flip-flops caused by process, voltage and temperature (PVT) variation, on top of the nominal skew the clock tree was designed for. It is estimated in three main ways: statistical models built from the delay spread of each clock buffer and wire, simulation across corners and Monte Carlo samples, and measurement on silicon. The estimate matters because every picosecond of skew variation either eats setup margin or causes hold violations, so a good estimate turns directly into frequency and yield.
Clock skew is the difference in arrival time of the clock signal at different sequential elements of a circuit. It affects timing, power and reliability, and it varies with process, voltage and temperature. In this post I go through the main methods for estimating that variation, their strengths and weaknesses, and how the estimate is used in timing sign-off.
Key takeaways
- Only the non-common part of two clock paths contributes to skew variation; the shared trunk cancels out, which is why tools apply common path pessimism removal (CPPR).
- Adding independent random variations as a root sum of squares (RSS) gives a much smaller number than adding worst cases, but only if the variations really are independent.
- Systematic effects (voltage droop, temperature gradients, layout-dependent effects) are correlated and must not be RSS-combined as if random.
- SPICE Monte Carlo on the critical clock paths is the reference for statistical estimates; silicon measurement is the final check.
- Clock meshes and balanced H-trees are used precisely because they reduce skew variation at the cost of power.
What causes skew to vary
| Source | Type | Effect on skew |
|---|---|---|
| Random device mismatch (threshold voltage, channel length) | Random, local | Different buffer delays on the two branches |
| Die-to-die or wafer-level process shift | Global, systematic | Mostly common to both branches, small net skew effect |
| IR drop and supply noise | Spatially correlated | Branches in a hot spot slow down together |
| Temperature gradients | Spatially correlated | Buffers and wires near hot blocks are slower |
| Interconnect width, thickness and coupling | Random and systematic | RC delay differs between routes, especially long, congested ones |
| Ageing (BTI, HCI) | Use-dependent | Gated clock branches age differently from free-running ones |
Method 1: statistical models
One of the simplest methods is a statistical model based on the nominal values and variances of the clock delay elements. Each buffer and wire segment on a clock path is given a mean delay and a standard deviation, usually from library or foundry variation data. The arrival time at a flip-flop is the sum of the segment delays, and skew is the difference between two arrival times.
If the segments are independent and normally distributed, the variances add. For two branches with only their non-common segments counted:
σ(skew) = √(Σ σi² of the launch branch + Σ σj² of the capture branch)
Worked example (illustrative numbers)
Two flip-flops share a clock trunk, then split into branches of four buffers each. Each buffer has a nominal delay of 25 ps and a random σ of 2 ps.
- Worst-case stacking: one branch all slow at +3σ, the other all fast at -3σ, gives 8 × 6 ps = 48 ps of skew variation.
- RSS with independent buffers: σ(skew) = √(8 × 2²) = 5.7 ps, so a 3σ bound is about 17 ps.
- The shared trunk adds nothing in either case, because any variation in it shifts launch and capture equally.
The difference between 48 ps and 17 ps is the reason statistical methods exist: worst-case stacking is very pessimistic when variations are truly random. The advantage of this method is that it is easy to implement and needs little information about the circuit. The disadvantage is that it may not capture correlations and non-linearities in the delay parameters, so it may overestimate or underestimate skew variation depending on the clock tree topology. A supply droop that slows all four buffers of one branch together is not random, and treating it as random gives a dangerously small number.
In sign-off tools this idea appears as on-chip variation (OCV) derates, advanced OCV (AOCV) tables that shrink the derate with path depth, and parametric OCV (POCV) or Liberty Variation Format (LVF) data that carry a σ for each cell arc.
Method 2: simulation-based estimation
A second method simulates the clock network under different PVT scenarios and measures skew in each. In practice this means:
- Corner simulation: run the extracted clock network at slow, typical and fast process with low and high voltage and temperature.
- Monte Carlo SPICE: sample local and global variation many times on the critical clock paths and build a skew distribution.
- Mixed scenarios: apply an IR-drop or thermal map from power analysis so that spatial correlation is modelled.
This approach can capture correlations and non-linear effects such as slew-dependent delay, and gives more accurate estimates than a simple statistical model. The disadvantage is cost: it needs more compute time, and it may not cover every possible combination of variations. Teams usually reserve Monte Carlo SPICE for the clock spine, the mesh drivers and a handful of critical branches, and use it to calibrate the derates used in static timing analysis.
Method 3: measurement on silicon
A third method measures skew on fabricated chips under different conditions, using on-chip sensors or external instruments. Common on-chip structures include delay-line or vernier time-to-digital converters between two clock branches, ring oscillators placed near clock buffers to track local speed, and scan-based tests that sweep clock delay until a capture fails.
This gives the most realistic view of skew variation because it reflects actual behaviour. The disadvantages are cost and effort, the area of the test structures, and the fact that it only covers the points that were instrumented. It also arrives late: measurement validates the design and calibrates models for the next one, rather than guiding the current one.
Comparing the methods
| Method | Accuracy | Cost | When used |
|---|---|---|---|
| Statistical model / OCV derates | Good for random effects, weak on correlated ones | Low | Every STA run, early planning |
| Corner and Monte Carlo simulation | High on the paths simulated | Medium to high | Clock spine, mesh, critical branches, derate calibration |
| Silicon measurement | Highest for the points measured | High | Test chips, bring-up, model correlation |
How the estimate is used
Clock skew variation estimates improve performance and robustness in several ways:
- Timing constraints and margins: the estimate sets clock uncertainty and derates, so the design avoids timing violations without wasting slack on excess margin.
- Frequency and voltage tuning: knowing the skew spread lets designers choose an operating frequency or voltage with the right trade-off between speed, power and reliability.
- Finding the sources: the analysis points at the branches causing variation, such as routing congestion, long unbalanced wires, process-sensitive buffers or temperature gradients, so they can be fixed.
- Choosing the clock structure: a clock mesh shorts many branch ends together, which averages out local variation and gives low skew variation, at the cost of higher power. A tree with a mesh at the leaves, or a multi-source clock tree, is a common compromise.
Common mistakes
- Counting variation on the common clock path, which double-counts pessimism (CPPR removes it).
- RSS-combining effects that are correlated, such as a regional IR-drop.
- Using the same derate for short and deep clock paths; deeper paths average out more random variation.
- Forgetting that hold checks on the same edge are the most sensitive to skew variation, because there is no clock period to absorb it.
- Characterising skew at one corner only; skew can reverse sign between slow and fast corners when branches use different buffer types or wire lengths.
Interview questions on this topic
- Why does variation on the common part of two clock paths not affect skew?
- Four buffers per branch, σ 2 ps each: what is the 3σ skew variation with RSS, and with worst-case stacking?
- When is RSS combination wrong?
- Why does a clock mesh reduce skew variation, and what does it cost?
- How would you check your skew model against silicon?
FAQ
Is skew variation the same as jitter?
No. Skew variation is a static, spatial difference that depends on the die and its operating point. Jitter is a cycle-to-cycle timing change at one point, caused by the clock source and supply noise. Both are included in clock uncertainty, but they are modelled differently.
Why is skew variation worse for hold than for setup?
A setup check spans a clock period, so skew is only part of the budget. A hold check compares two edges at the same instant, so even a few picoseconds of extra skew can cause a failure, and a hold failure cannot be fixed by slowing the clock.
Which method should I use?
All three, at different stages. Statistical derates in every timing run, Monte Carlo SPICE to calibrate them on the critical clock structures, and silicon measurement to confirm the models. The right balance depends on accuracy needs, complexity, cost and what data is available.

Your comments will be moderated before it appears here.