Clock Skew

Clock Skew 1
 Clock Skew

Probably the largest source of design

failures in FPGA.

Clock Skew 2
 Definitions
• Sequentially-Adjacent Register Pair
Two registers with only combinational logic or
interconnect between them

• Clock Skew
Given two sequentially-adjacent registers, Ri and Rj,
and an equipotential clock distribution network, the
clock skew between these two registers is defined as
Tskew-i,j = Tci - Tcj

where Tci and Tcj are the clock delays from the clock
source to the registers Ri and Rj, respectively.

Clock Skew 3
 Implications of Definitions
• Sequentially-Adjacent Register Pair
Skew is only meaningful between adjacent pairs of
registers, not between any pair of registers in a clock
domain.

- Clock skew is only important between the pairs (FF1, FF2) and (FF2, FF3).
- The skew between the pair (FF1, FF3) is meaningless.
- Calculations done quickly on sets is often unrealistic.

Clock Skew 4
 Implications of Definitions
• Definition of skew: “… from the clock source”
A circuit node that is common in time for both paths is
the origin of skew calculations.

- ‘A’ is the common point in time for this circuit.

- ‘B’ and ‘C’ are not common. Why? B1 and B2 have different transitors.
The loading on B and C are different. B and C can have different line
lengths. Parasitic capacitances can be different. Etc. Lines on a
schematic are not ideal wires.
- B1 and B2 will not “track” 100%. This is shown later for aging.
Clock Skew 5
 A First Skew Calculation

Two cases: 1) A → B is fast, A → C slow

2) A → B is slow, A → C fast

Just show the case 1 calculation (hold time)

Skew = tplh B1min - tplh B2max

Is this correct?
Is this realistic?

Clock Skew 6
 A First Skew Calculation

Skew = tplh B1min - tplh B2max

Is this correct?

No, it is wrong. It assumes that the signal ‘A’

arrives at the inputs to B1 and B2 at the same
time. In an FPGA in particular, they will not,
making this calculation too optimistic.

Clock Skew 7
 A First Skew Calculation

Skew = tplh B1min - tplh B2max

Is this realistic?

It is wrong if you use extreme value analysis without thinking, assuming that
this circuit is on one chip and would make thes calculation too conservative.

For example, temperature and voltage won’t have drastic differences on the
same chip at the same time. Life and radiation effects can not be assumed to
track. Note that calculations must be done over all corners.

Clock Skew 8
 A Zero th Skew Calculation

For this simpler case, we have to calculate the time from A → B (min) and from
A → C (max) to compute the worst-case skew.

The input to B1, A, is common to both, so that is our common timing point.

The paths from the output of B1 to FF1:CLK and to FF2:CLK are different.

But tplh (B1) is in common. Legally, that can be factored out, eliminating the
unrealistic conservatism of having a transistor be both ‘good’ and ‘bad’ at the
same time. Practically, however, some tools lump the gate and routing delays
together, so the calculation is needlessly conservative.
Clock Skew 9
 Tools: Your Friend?

Some tools try and calculate skews for you. However, you must
understand the model that it is using and apply it wisely.
Clock Skew 10
 Effects of Skew
Assume: Sequentially-Adjacent, Rising Edge-Triggered Flip-Flops
Routing delay bundled with gate delay
Case 1: Setup time

1 Clock Period

Ideal CLK

Clock Skew 11
 Effects of Skew
Assume: Sequentially-Adjacent, Rising Edge-Triggered Flip-Flops
Case 1: setup time

With an ideal, skew-free clock, we have:

Add delays: FF1:CLK→Q(max)
G1tP(max)
FF2:tSU
Margin = tperiod - total

Clock Skew 12
 Effects of Skew
Add Worst-Case Skew for tSU

1 Clock Period
Available Time
Decreased by tSKEW

Ideal CLK

FF1:CLK
FF2:CLK
Clock Skew 13
 Effects of Skew
Assume: Sequentially-Adjacent, Rising Edge-Triggered Flip-Flops
Routing delay bundled with gate delay
Case 2: Hold time

Ideal FF1:CLK
D
E
Ideal FF2:CLK
tH met
Clock Skew 14
 Effects of Skew

Early FF1:CLK

D
E

Late FF2:CLK

Note: used min, best case for prop delays and max, worst-case for
clock path to FF2.
Clock Skew 15
 Clock Skew, Our Friend

Data

Clock

Attempting to force the downstream flip-flops to be clocked

before the upstream flip-flops can change their data. This
improves hold time at the cost of setup time.
Clock Skew 16
 Clock Skew, Our Friend

FF2: Latest Sample FF1: Earliest Data Change

CLKB

The use of opposite edge clocking makes this circuit skew-

tolerant, at the cost of setup time, typically not an issue for shift
registers. Additionally, the clock can be run at 1/2 speed, saving
power.
Clock Skew 17
 Clock Skew, Our Friend

FF2: Latest Sample FF1: Earliest Data Change

PH1

PH2

The use of two-phase clocking makes this circuit skew-tolerant,

at the cost of setup time.
Clock Skew 18
 Clock Skew, Our Friend

L2: Latest Sample L1: Earliest Data Change

PH1

PH2

The use of two-phase, non-overlapping clocking makes this

latch-based circuit skew-tolerant.
Clock Skew 19
 Clock Skew, Our Friend
• For pipelined designs with unequal
combinational delays between latches or
registers, clock skew can provide a better “fit.”

Data

Clock

Clock Skew 20
 Design Impacts of Clock Skew
• Calculate skews carefully
– setup time
– hold time
• Use low-skew clock resources when
practical
• Understand skew-tolerant circuit when it is
not practical to use low-skew clock
resources
• Some ASICs require a minimum of 1 gate
between sequentially adjacent registers
Clock Skew 21
 Clock Skew
D Q D Q D Q D Q

Normal Routing Resource

Shift register is given as an example. Also seen

in counters and other logic structures.

Clock Skew 22
 Clock Skew
D Q D Q D Q D Q

D Q D Q D Q D Q

• Clock trees are made to increase fanout.

• Not placing buffers and flip-flops on the same row
– Can increase skew problem.
Clock Skew 23
 Clock Skew - Timing Model
TCQ TROUTE TH

D Q D Q

FF1 FF2

TSKEW
• Hold time at FF2 is the concern.
– Worst-case
– Low VIH FF1
– Hi VIH FF2
– Fast TCQ, TRoute
– High TSKEW
• TCQ + TROUTE + TH > TSKEW
Clock Skew 24
 Local Clock: Physical Realization
14

13 4 5 7

12 R

11 6 Note: Antifuse located

10 at each junction.
9

7
LVT
6

5 9 D

23 24 25 26 27 28 29 30 31 32 33 34 35

The net CMDREG/CLK3 driven at location XY = (27, 5) uses an LVT.

LVT data: column = 30, Y-span = (14, 6).
Clock Skew 25 Y-spread of inputs = (13, 5).
Net data: fanout = 13,
 Design Strategy (2)
Use of Local, High-Skew Clock
Global Clock converted to local clock
and fed to drive shift register.

Global clock
Shift register with
local clock and 8
flip-flop loads.1

1This project had a design rule of no more than 5 loads

on a local, high-skew clock. This was repeatedly violated.
Clock Skew 26
 Clock Skew - Timing Analysis
Most static timing analyzers give bounded numbers for min, max.

Just setting “MAX” or “MIN” does not account for variations as a result
of fabrication differences, anti-fuse resistance, changes as a result of
aging, etc, and will be too liberal.

A full MIN/MAX analysis is too conservative since elements near each

other on the same die cannot vary that widely. I.e., one part can’t be at
4.5VDC, the other at 5.5VDC.

For each environmental condition, it is fair to hold temperature, voltage,

fixed.

MIN/MAX will still be a bit conservative, since will range over all
manufacturing conditions, not limited to variation within a single die.

Clock Skew 27
 Antifuse Resistance Variation
ONO Antifuse Resistance Distribution
Programming Current = 5mA
[from Antifuse FPGAs, J. Greene, et. al.]
60

50
% Distribution

40

30

20

10

0
300 400 500 600 700 800 900

Resistance

Clock Skew 28
 Prop Delay Delta vs. Life

Change in delay time from a 1000 gate programmable

gate after 1000 hours dynamic burn-in at 125 °C and 5.75V.
Clock Skew 29
 Prop Delay Delta vs. Life
RH1280 Change in Propagation Delay
After 1000 Hour Life Test
Tested at 4.5 Volts, 125C
15
NOTES:

1. All delay values rounded to the

nearest integer number of ns.

2. The average propagation delay

Number of Samples

10 was 106.5ns before the life test.

0
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10
Change in Propagation Delay (ns)
Clock Skew 30
 Prop Delay Delta vs. Life
RH1280 Change in Propagation Delay
After 1000 Hour Life Test
Tested at 4.5 volts, 125C
25

20
Number of Samples

15

10

0
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
Change in Propagation Delay (percent)

Note: Over a long path, 16 modules + I/O, TP exceeding 100 ns.

Clock Skew 31
 Clock Skew - From VHDL
Coding Example
Library IEEE;
Use IEEE.Std_Logic_1164.All;

Entity Skew Is
Port ( Clk : In Std_Logic;
D : In Std_Logic;
Q : Out Std_Logic );
End Skew;

Library IEEE;
Use IEEE.Std_Logic_1164.All;

Architecture Skew of Skew Is

Signal ShiftReg : Std_Logic_Vector (31 DownTo 0);
Begin
P: Process ( Clk )
Begin
If Rising_Edge (Clk)
Then Q <= ShiftReg(0);
ShiftReg (30 DownTo 0) <= ShiftReg (31 DownTo 1);
ShiftReg (31) <= D;
End If;
End Process P;
End Skew;

Clock Skew 32
 Clock Skew - From VHDL
Synthesized Results

Results will depend on coding, directives and attributes, synthesizer, and

synthesizer revision.

Here we see that the logic synthesizer generated a poor circuit.

Clock Skew 33
 Clock Skew Correction
No “PRESERVE”

High-skew Clock

Clock Skew 34
 Clock Skew - Chip-to-Chip

D Q D Q

FF1 FF2

Analysis may show problems. Some architectures are

designed with 0 ns tH; others incorporate delay elements
(configurable) on the data inputs to ensure reliable clocking.
Clock Skew 35