TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

Clock Tree Synthesis Basics

The Editor's Desk
by
0
Share Facebook X LinkedIn WhatsApp Email

Clock tree synthesis (CTS) is the step in physical design that builds the network of buffers and wires carrying the clock from its source to every flip-flop. The main idea is to balance the clock arrival times, so the skew between endpoints stays small. The tree is built to meet three kinds of constraints: clock skew (the difference between clock arrival times), clock latency (the delay from the clock root to a clock leaf) and transition time (how fast the clock edges rise and fall).

Why the clock needs its own tree

After placement, a single clock net may have to reach hundreds of thousands of flops spread across the die. No single driver can do that: the edges would be far too slow and would reach nearby flops long before distant ones. CTS solves this by building a tree. The source drives a few large buffers, each of those drives a few more, and so on until each final buffer drives a small cluster of flops. The tool sizes and places these buffers, and chooses the branch lengths, so that every flop sees the clock at nearly the same time with sharp edges.

What the tool tries to achieve

  1. Low clock skew: small differences in arrival time between flops that exchange data.
  2. Controlled clock latency: the maximum delay from the clock root to a leaf, kept short enough to limit power and on-chip variation.
  3. Good transition time: sharp clock edges at every flop, within the library's limits.

How clock nets are treated differently

  • Clock buffers are usually larger, with shorter transition times and more evenly matched rise and fall times than ordinary buffers. Many libraries have dedicated clock buffers and inverters for this reason.
  • Clock nets are generally routed first, on higher metal layers, with minimum detouring, to give them the highest routing priority. Higher layers are thicker, so their resistance is lower.
  • Signal integrity and shielding: because of their importance, clock nets have to be protected from becoming either aggressors or victims during signal integrity closure. They are generally shielded with VDD or ground wires, or given extra spacing, to prevent coupling with neighboring signals.

Related ideas you will meet

  1. Effective skew: the worst skew between two flops that are actually talking to each other. It is equal to or lower than the worst skew across the whole clock domain.
  2. Useful skew: skew that is introduced on purpose, delaying the clock at a capturing flop to give a slow path more time and fix setup violations.
  3. OCV (on-chip variation): the fact that identical buffers in different parts of the die do not have identical delays, which timing analysis must allow for.

The original post promised links with more detail on clock trees; the related posts at the end of this article now cover each of these topics.

Deep dive: the numbers and details

This part holds the equations, typical targets and a worked example. Skip it if the overview is all you need.

Where CTS sits in the flow

  1. Before CTS the clock is ideal: timing analysis assumes it arrives everywhere at once, with an estimated latency and a clock uncertainty value that covers expected skew and jitter.
  2. CTS builds and balances the tree, inserting buffers or inverters, and often clock gating cells are cloned or moved during this step.
  3. After CTS the clock is propagated: timing uses the real buffer and wire delays. Hold violations usually appear for the first time here, because real skew now exists, and are fixed by inserting delay on short data paths.
  4. Clock routing then follows, before signal routing, with the shielding and layer rules set for clocks.

What a CTS specification contains

ItemWhat it controls
Clock roots and sinksWhere each tree starts, which pins are endpoints, and which pins to ignore or treat as balance points
Target skewThe allowed arrival time spread between related sinks
Target or maximum latencyHow long the tree may be; shorter means less power and less OCV
Maximum transition and capacitanceEdge rate and load limits at each buffer output and sink
Buffer and inverter listWhich cells the tool may use in the tree
Routing rulesPreferred layers, double width or spacing, shielding

Why buffering helps: a quick delay estimate

An unbuffered wire has distributed resistance r and capacitance c per unit length. Its delay grows with the square of its length L: the 50% delay of a distributed RC line is about 0.38 · r · c · L2 for the wire alone (the Elmore estimate is 0.5 · r · c · L2). Splitting the wire into k equal segments with a buffer at each break turns the delay into k segments of length L/k, so the wire part becomes k · 0.38 · r · c · (L/k)2 = 0.38 · r · c · L2 / k, at the cost of k buffer delays. Delay then grows roughly linearly with length instead of quadratically. This is the main reason long clock routes are buffered, and why edge rates improve too.

Worked example: sizing the tree

Illustrative numbers. A block has 50,000 flops. Suppose the library and transition limit allow each clock buffer to drive about 20 loads.

  • Last level: 50,000 / 20 = 2,500 leaf buffers.
  • Next level up: 2,500 / 20 = 125 buffers.
  • Next: 125 / 20 = about 7 buffers, then 1 root driver.
  • That is 4 levels and about 2,630 clock buffers. At, say, 40 ps per level plus wire delay, the insertion delay lands in the hundreds of picoseconds.

Every one of those buffers switches twice per clock cycle, which is why the clock network is often one of the largest consumers of dynamic power on a chip, and why clock gating near the root saves so much.

How skew enters timing

With skew S defined as capture clock arrival minus launch clock arrival:

  • Setup: Tclk + S ≥ tcq + tlogic,max + tsetup
  • Hold: tcq + tlogic,min ≥ thold + S

Positive skew helps setup and hurts hold; negative skew does the reverse. This is why effective skew between communicating flops is the number that matters, and why useful skew can fix a setup path only when the hold margin on the same flops can absorb it.

OCV and common path pessimism

Under on-chip variation analysis, the tool times the launch clock path as slow and the capture clock path as fast for setup (and the reverse for hold). On the part of the tree that both paths share, one buffer cannot be fast and slow at once, so tools remove that false penalty with common path pessimism removal (CPPR, also called clock reconvergence pessimism removal). Trees where related flops share most of their clock path, and short trees in general, therefore lose less margin to OCV.

Tree topologies

  • H-tree: symmetric branches give equal path lengths by construction. Good for regular structures.
  • Balanced buffered tree: what most place-and-route tools build by default; buffers are inserted and sized to match delays to irregular sink positions.
  • Clock mesh or spine: a grid of wires shorted together and driven from many points, used in high-frequency processors to reduce skew and sensitivity to variation, at a power cost.

Common mistakes

  • Over-constraining skew to near zero, which adds buffers, power and latency for little gain.
  • Forgetting to define exceptions: pins that should not be balanced, such as the clock input of a divider whose output starts its own tree.
  • Leaving clock nets without shielding or spacing rules, then chasing crosstalk on clocks late in the flow.
  • Ignoring hold after CTS: the first propagated-clock timing run is when hold fixing starts in earnest.

Further reading on this blog

The Editor's Desk
About the Author
The Editor's Desk
Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!