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

All About Clock Skew and Short Paths: Hold Violations Explained

The Editor's Desk
by
2
Share Facebook X LinkedIn WhatsApp Email

Clock skew is the difference in arrival time of the same clock edge at two different registers. A short path is a register-to-register data path so fast that, when the capture clock arrives late, new data races through and overwrites the value the capture flop should still be holding. That race is a hold-time violation, and unlike a setup violation it cannot be fixed by slowing the clock. The cure is to measure skew with static timing analysis at the fast corner, keep it small with low-skew clock networks, and add delay to any path that is still too short.

Key takeaways

  • Skew = capture clock arrival minus launch clock arrival. Positive skew helps setup and hurts hold; negative skew does the opposite.
  • A short path fails when the new data reaches the capture flop before that flop has finished holding the old data.
  • Hold checks do not depend on the clock period, so a hold failure on silicon fails at every frequency.
  • Check setup (long paths) at the slow corner and hold (short paths) at the fast corner, with clock constraints defined.
  • In FPGAs, put clocks on the global clock network. In ASICs, clock tree synthesis balances skew and the tools insert delay buffers to fix hold.

What clock skew is

Differences in clock signal arrival times across the chip are called clock skew. Every register-to-register transfer has to satisfy the setup and hold requirements of the capturing register, and both the data propagation delay and the clock skew are part of that calculation. Clocking sequentially adjacent registers on the same edge of a high-skew clock can cause timing violations or even functional failures.

Skew comes from unequal wire lengths and buffer counts in the clock network, uneven loading, and on-chip variation in voltage, temperature and process. Clock jitter is a related but separate effect: skew is a spatial difference between two points for the same edge, while jitter is the variation of an edge in time at one point.

Why short paths are the danger

The problem of short data paths in the presence of clock skew is the same as a hold-time violation in flip-flops. It arises when the data path between two adjacent flip-flops is too fast compared with the clock skew. Put simply, it happens when the fastest possible arrival of new data (the launch flop's minimum clock-to-Q delay plus the minimum logic and wire delay) is shorter than the clock skew plus the hold time of the capture flop.

A common example is a shift register or a direct flop-to-flop connection: there is no logic in between, so the only delay is clock-to-Q plus a short wire. If the clock reaches the second flop later than the first, the first flop's new output can arrive at the second flop before its hold window closes.

The key difference from a setup failure is that the clock period does not appear in the hold check. A setup failure can be worked around in the lab by lowering the frequency; a hold failure cannot. The Harvey Mudd CMOS VLSI Design lecture on sequential circuit design puts it bluntly: if hold times are violated, the chip fails at any speed.

How to find and fix short paths

  • Measure. Run static timing analysis after place and route, when the real clock network delays are known, and look at the hold (minimum delay) section of the report.
  • Constrain. The report is only meaningful if every clock is constrained.
  • Use the right corner. Hold problems appear where data paths are fastest: fast process, high voltage and the temperature that makes cells fastest.
  • Reduce skew. Put clocks on dedicated low-skew clock networks.
  • Add delay. Where a path is still too short, add buffers or routing delay, then check that the same path still meets setup at the slow corner.

Deep dive: the numbers and details

This part holds the equations, a worked example and the details of timing reports. Skip it if the overview is all you need.

Sign convention and the two checks

For a path from a launch flop FF1 to a capture flop FF2:

skew = t_clk(FF2) - t_clk(FF1)

Setup: tcq(max) + tlogic(max) + tsetup  <=  Tclk + skew
Hold:  tcq(min) + tlogic(min)           >=  thold + skew

Positive skew (the capture clock arrives later) gives the data more time to meet setup, but it also holds the capture window open longer, so the next data has to arrive later to avoid corrupting it. The same relationship, written as tcd >= thold - tccq + tskew, is derived in the Harvey Mudd sequential circuit lecture cited above.

Worked example (illustrative numbers)

ParameterValue
tcq(min) of FF10.10 ns
tlogic(min), wire only0.05 ns
thold of FF20.05 ns
Skew, FF2 clock later than FF10.25 ns

Arrival = 0.10 + 0.05 = 0.15 ns. Required = 0.05 + 0.25 = 0.30 ns. Hold slack = 0.15 - 0.30 = -0.15 ns, a violation. Halving the clock frequency changes nothing in this sum. Two fixes work: reduce the skew to below 0.10 ns (then slack is 0.15 - 0.15 = 0, just met, so aim lower still), or add at least 0.15 ns of delay to the data path with buffers or routing, while checking that the longest path through the same logic still meets setup.

Sizing the fix against setup

Suppose the same FF1 to FF2 path runs at 500 MHz (Tclk = 2.0 ns) and its slow-corner setup arrival is 1.40 ns with a 0.08 ns setup time. With the 0.25 ns of positive skew, setup slack is 2.0 + 0.25 - 1.40 - 0.08 = 0.77 ns. A delay cell that adds 0.20 ns at the fast corner may add about 0.35 ns at the slow corner (slow corners stretch every delay), which leaves 0.42 ns of setup slack. These numbers are illustrative, but the method is the general one: size the hold fix at the fast corner, then confirm setup at the slow corner.

Reading the timing report

The timing report is only valid if one or more clock constraints are specified. If the design clocks are not constrained, the report will be empty. In the Actel (now Microchip) Designer Timer flow that this post was written against, the report had four sections; other vendors use different names but report the same things:

  • Header: software version, design name, operating condition, device type, speed grade and Timer preferences.
  • Clock constraint violation: the critical paths limiting any clock frequency constraint set in the General tab.
  • Max delay constraint violation: the critical paths limiting any Max Delay constraint set in the Timer Path tab.
  • Min delay constraint violation: short data paths susceptible to hold-time violations.

In the timing report, the skew of the clock network is taken into account when calculating slack. Each section is sorted by slack, and a negative slack indicates a violation. The report reflects the operating conditions set in the Timer preferences. To examine long data paths against clock or Max Delay constraints, export the report with the preferences set to worst case (long paths). To find all possible hold-time violations, create it with the preferences set to best case (short paths). After each change of operating conditions in the Timer window, select "calculate delays" before exporting the report.

The same idea in current tools

CheckCornerTypical command or report
Setup (long paths)Slow process, low voltage, worst temperaturereport_timing -delay_type max, report_timing_summary (Vivado), report_timing -setup (Quartus Timing Analyzer)
Hold (short paths)Fast process, high voltage, best temperaturereport_timing -delay_type min, report_timing -hold
Clock skew itselfBothClock skew and clock network reports, or the clock path section of each timing path

Modern tools analyze several corners in one run and apply on-chip variation derates, so best case and worst case are no longer separate exports. The principle is unchanged: hold is a fast-corner problem and setup is a slow-corner problem.

Minimizing the clock skew

The short-path problem is created by unacceptably large clock skew, so minimizing the skew is the best way to reduce the risk. Many FPGA devices offer global routing resources (dedicated clock networks) that keep skew low.

If free global resources are available on the device, assign the clock signals to them. Keeping the clock skew below the smallest register-to-register delay in the design, by using low-skew global resources, makes the design tolerant of short-path problems.

Other practical steps:

  • Do not gate or divide clocks in general logic. Use the clock enables, PLLs and clock buffers the device provides.
  • In ASICs, let clock tree synthesis balance the tree, then run hold fixing after routing. Tools insert delay cells on the short paths automatically.
  • Where two registers must be clocked by unbalanced branches, check the hold slack explicitly and add delay to the data path if needed.
  • Remember that deliberate (useful) skew can rescue a setup-critical path, but every picosecond borrowed is taken from the hold margin of the same flop.

Common mistakes

  • Running only setup analysis, or only at the slow corner.
  • Leaving clocks unconstrained, which produces an empty or misleading report.
  • Assuming a hold violation can be fixed in the lab by lowering the clock frequency.
  • Fixing hold with so much delay that the same path fails setup at the slow corner.

FAQ

Is clock skew always bad? No. It is only harmful when it is unplanned or too large. Useful skew is a deliberate technique for balancing setup slack between pipeline stages.

Why does slowing the clock fix setup but not hold? The clock period appears in the setup equation but not in the hold equation. Hold compares the launch and capture of the same edge.

Which corner shows hold problems? The fast corner, where data paths are quickest. Skew can also vary between corners, so check hold in every corner the tool supports.

Further reading

Comments are most invited.

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

Post a Comment

2 Comments

Your comments will be moderated before it appears here.

  1. You are doing a wonderful job! Lot of useful info. Keep it up! Please change the background color to simple white.

    thanks a lot,
    ak

    ReplyDelete
  2. wonderful posts.....I Super Liked it

    ReplyDelete
Post a Comment

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

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