In this blog post, we will discuss some of the challenges and techniques involved in designing a clock network for a very large scale integration (VLSI) circuit. A clock network is a special type of interconnect that distributes the clock signal to all the synchronous elements in the circuit, such as flip-flops, latches and registers. The clock signal is used to synchronize the operation of these elements and ensure correct functionality and timing of the circuit.
The main challenge: clock skew
One of the main challenges in designing a clock network is to minimize the clock skew, which is the difference in arrival time of the clock signal at different destinations. Clock skew can cause timing violations, such as setup and hold violations, and reduce the performance and reliability of the circuit. To minimize clock skew, the clock network should have low resistance, low capacitance and balanced routing. Additionally, the clock network should tolerate process variations, temperature variations and noise.
Why the clock network is special
The clock is the busiest net on the chip. It switches every cycle, it reaches every flip-flop, and it often spans the whole die. That makes it different from ordinary signal wiring in three ways:
- Timing: every setup and hold check in the design depends on when the clock arrives at both ends of the path, so errors in the clock network show up everywhere.
- Power: because it toggles every cycle and drives a huge load, the clock network is often one of the largest single consumers of dynamic power.
- Signal quality: slow, rounded clock edges make flip-flops slower and more sensitive to noise, so clock edges are kept sharp with strong buffers.
Besides skew, designers watch latency (insertion delay, the time from the clock source to the flops), transition time (how sharp the edges are at the flops) and jitter (edge-to-edge variation in time, mostly from the clock source and supply noise).
Topologies
Some of the techniques that can be used to design a clock network are:
- H-tree: This is a hierarchical structure that divides the chip area into smaller regions and distributes the clock signal using a binary tree topology. The H-tree can achieve low and balanced clock skew, but it requires a large area and power consumption.
- Mesh: This is a grid-like structure that connects all the destinations using horizontal and vertical wires. The mesh can achieve low resistance and high tolerance of variation, but it requires a large capacitance and routing complexity.
- Tree-mesh hybrid: This is a combination of H-tree and mesh structures that tries to balance the trade-offs between them. The hybrid structure can achieve moderate skew, resistance, capacitance and variation tolerance, but it requires careful optimization and tuning.
In everyday ASIC work there is a fourth option that most blocks actually use: a synthesized clock tree, built automatically by the clock tree synthesis (CTS) step of place and route. It inserts and sizes buffers so that the skew between flops stays within a target, without the strict geometric symmetry of an H-tree. Large high-performance processors are where the hand-planned H-tree, mesh and hybrid structures appear.
Choosing between them
| Topology | Skew | Power and wire | Variation tolerance | Typical use |
|---|---|---|---|---|
| Synthesized tree (CTS) | Moderate | Lowest | Moderate | Most ASIC and SoC blocks |
| H-tree | Low (by symmetry) | Moderate to high | Moderate | Global distribution, regular arrays |
| Mesh (grid) | Very low | High | High | High-frequency processor cores |
| Tree driving a mesh | Very low | High | High | Large microprocessors |
Common interview questions
- What is the difference between clock skew and clock jitter? (Skew is a spatial difference between two flops for the same edge; jitter is a variation in time of the edges at one point.)
- Why does positive skew help setup and hurt hold? (The capture edge arrives later, which gives the data more time but also lets new data overwrite the old value sooner relative to the capture edge.)
- Why does a mesh reduce skew? (The grid shorts together the outputs of many drivers, so local differences in buffer delay average out.)
- Why is clock gating done at the root of a subtree rather than at each flop? (One gate switches off the clock to the whole subtree, saving the power of all the buffers below it.)
These were the basic concepts and techniques involved in VLSI clock network design when this post was written in 2010. More advanced topics include clock gating, clock buffering and clock synchronization.
Deep dive: the numbers and details
This part holds the clock power equation, the H-tree arithmetic and published figures for a real tree-plus-grid network. Skip it if the overview is all you need.
Clock power
Dynamic power of a net is P = a x C x V2 x f, where a is the fraction of cycles in which the net makes a full charge and discharge. A clock does this every cycle, so a = 1, while typical logic nets have a well below 0.5. Illustrative numbers: a clock network with 1 nF of total capacitance (wires, buffer inputs and flop clock pins) at 0.8 V and 2 GHz burns 1 x 10-9 x 0.64 x 2 x 109 = 1.28 W, before counting the flops' internal clock power. This is why clock gating, which stops the clock to idle subtrees, is one of the most effective power techniques.
H-tree arithmetic
Each level of an H-tree splits one branch into four endpoints at the corners of an H, so k levels reach 4k endpoints. Six levels reach 4,096 regions; from each, a local tree or buffer drives the flops nearby. Every root-to-leaf path has the same length and the same number of buffers, so with identical wire and buffer delays the skew is zero by symmetry. In practice skew comes from what symmetry cannot control: uneven loads at the leaves, process variation between buffers, and temperature or supply differences across the die. Each level also halves the wire length of the level above (roughly), so the top levels carry the longest wires and need the strongest drivers.
A real tree-plus-grid network
Restle and colleagues at IBM described a global clock strategy used on several microprocessors in which buffered tunable trees drive a single common grid covering most of the chip, combining advantages of trees and grids. Tuning that network required a new method because it held up to 6 m of wire, modeled with 50,000 resistors, capacitors and inductors, and they reported measured global clock skew as low as 22 ps on large microprocessor chips (IBM Research summary; P. J. Restle et al., IEEE Journal of Solid-State Circuits, vol. 36, no. 5, pp. 792 to 799, May 2001, doi:10.1109/4.918917).
To put 22 ps in context: at 2 GHz the period is 500 ps, so that skew is about 4.4 percent of the cycle.
Skew budget worked example
For a 1 GHz block (1,000 ps period), suppose clock-to-Q is 80 ps, setup is 40 ps and the timing team reserves 50 ps for jitter. If the clock network has 60 ps of worst-case skew against the path, the logic between flops can use at most 1,000 - 80 - 40 - 50 - 60 = 770 ps. Cutting skew to 20 ps gives the logic 810 ps, a 5 percent gain in usable time, which can be traded for higher frequency or lower supply voltage. For hold, the same 60 ps of skew means the shortest path must be at least 60 ps longer than the hold time, often forcing delay buffers into short paths.
Further reading
- E. G. Friedman, "Clock distribution networks in synchronous digital integrated circuits", Proceedings of the IEEE, vol. 89, no. 5, pp. 665 to 692, May 2001, doi:10.1109/5.929649
- P. J. Restle et al., "A clock distribution network for microprocessors" (IBM Research)
- Useful skew: using clock skew as a tool to achieve optimal timing

Your comments will be moderated before it appears here.