Domino Logic

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 1

 Tinkering with Logic Gates
Things to like about CMOS gates:
Š easy to translate logic to fets
Š rail-to-rail switching
Š good noise margins, no static power since fets are in cutoff
Š sizing not critical to correct operation

Things not to like about CMOS gates:

Š N inputs Ö 2N fets (i.e., one nfet and one pfet)
Š large circuit area, especially for pfets, “heavy” loading of inputs
Š pfets are either large or slow relative to nfets
Š series connections can get very slow

We can replace pfet pullup network with pseudo-NMOS load (pfet with
grounded gate) but
Š dissipate static power when output is low
Š have to make load fet small to ensure that
VOL is low enough to cut off nfets in next stage
Š reduces static power consumption (good!) A B
Š increases output rise time (bad!)

One alternative: dynamic CMOS gates

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 2
 Dynamic CMOS Gates
B When CLK is low
Š evaluate nfet is off and precharge pfet is on
Š output node is precharged to VDD, other nodes
may precharge to VDD - Vth,n depending on values
A of inputs
When CLK goes high
Š evaluate nfet is on and precharge pfet is off
Š output node may be discharged if inputs have
“precharge” configured a conducting path to GND, otherwise
switch output node stays charged high.
Š inputs must be stable before CLK goes high
because once output has been discharged it
A B won’t go high again until next cycle
Š for same reason, noise/glitches on inputs cannot
exceed nfet threshold, a much more stringent
CLK “evaluate” requirement than for static CMOS gates.
switch

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 3

 There’s good news and bad news
The good news:
Dynamic gates are faster than static gates despite the extra “evaluate”
fet in the pulldown path because of the reduction in self-loading and the
elimination of the pullup short-circuit current during the first part of the
output transition.
The bad news:
Dynamic gates cannot be cascaded. Because of finite pulldown
time for node , node
starts to discharge!

CLK
nfets nfets

CLK

Solution – develop techniques that avoid races:

CMOS Domino logic, CMOS NORA (no race) logic
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 4
 CMOS Domino Logic
precharge: high
evaluate: falls (maybe)

nfets nfets buffer might

be needed
in any case
CLK for high fan-out
precharge:low circuits.
evaluate: rises (maybe)

When CLK is low, dynamic node is precharged high and buffer inverter output
is low. Nfets in the next logic block will be off. When CLK goes high, dynamic
node is conditionally discharged and the buffer output will conditionally go
high. Since discharge can only happen once, buffer output can only make
one low-to-high transition.
When domino gates are cascaded, as each gate “evaluates”, if its output
rises, it will trigger the evaluation of the next stage, and so on… like a line
of dominos falling. Like dominos, once the internal node in a gate “falls”, it
stays “fallen” until it is “picked up” by the precharge phase of the next cycle.
Thus many gates may evaluate in one eval cycle.
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 5
 More Domino-style Circuits
weak pfet “keeper” keeps dynamic node pulled high
during evaluate phase if it’s not being pulled down
through nfets Ö gate is static in both clock
phases.
CLK

nfets

“latching” pfet acts like keeper above unless

dynamic node gets pulled down during evaluate
phase. When buffer output goes high it switches
keeper off saving static power. Good for leakage
current problems...
CLK Note that you can put an even
number of static gates after
nfets the inverter and before the
next domino gate.

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 6

 High-fanin Domino Circuits

! Be careful of capacitive coupling

to dynamic node (see later slide).

Use NOR gate instead of inverter

as the buffer to make a faster high
CLK
fan-in AND gate. Same trick works
for high fan-in OR or MUX
functions.

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 7

 Optimizing Domino Logic (I)

nfets nfets

CLK evaluate nfet

precharge: low not needed?

Since domino gate outputs are low during the precharge phase, gates which
have only domino output nodes as inputs don’t need the “evaluate” nfet
since all the nfets in the pulldown will be off anyway.

But remember: if evaluate nfet is removed, precharge will “ripple” through

cascaded gates just like evaluates do. Maybe only remove for gates where
nfet stack is tall (i.e. resistive) enough that pullup will start to “win” anyway
before ripple reaches gates and turns off pulldowns. To avoid short-circuit
current: delay arrival of precharge signal to downstream gate.
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 8
 Optimizing Domino Logic (II)
In domino logic circuits we want evaluate to happen as quickly as possible.
We can tweak fet sizes to optimize evaluate speed.

small large
Some designers also
“grade” the sizes of
the nfets, smallest large
at the top (increase small
nfets
in R offset by
decrease in C)
CLK

If we make the nfet in the output inverter much smaller than the pfet then
Š the load on the internal node decreases, and
Š the switching threshold of the inverter increases
Both effects make the gate evaluate sooner. If large >> small, the gate delay
can be cut almost in half! However, the other edge is very slow, so ripple
precharge is a problem.
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 9
 “Flies in the ointment”

There are a few “little” difficulties:

Š “charge sharing” between nodes in the pulldown network
and the dynamic node can unintentionally reduce the
voltage of the dynamic node enough to switch output
buffer
Š the addition of the output inverter makes domino gates
non-inverting. One can often design around this limitation,
but some circuits cannot be implemented solely using
domino logic unless both polarities (true and complement)
of the inputs are available. If both polarities of inputs are
available then we can generate both polarities of internal
signals with two domino gates so subsequent stages will
have both polarities of their inputs available too.

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 10

 Charge Sharing
S’pose the dynamic node had been
discharged during the previous evaluate
cycle. Then during precharge, all the
intermediate nodes in the pulldown chain
3C
will remain discharged while the dynamic
C
node is precharged. When CLK goes high,
C
the voltage on the dynamic node goes to
3C V = .3V = 1.5V
C 3C + 6C DD DD

which is low enough to switch the output

C inverter.
C Fortunately this situation is easily detected by CAD
tools and can be resolved by (1) adding additional
C precharge devices to intermediate nodes or (2)
increasing size of output buffer which will increase
CLK capacitance of dynamic node (faster output buffer
may compensate for larger internal capacitance).

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 11

 Capacitive Coupling

OUT

OUT
t

CLK
When using multiple-input gates as the Domino buffer
changes in the “other” input during evaluate phase can
cause dynamic node voltage to sag due to capacitive
coupling, leading to unintended transition on OUT

Coupling
Couplingcan
canalso
alsooccur
occurbetween
betweenother
othersignal
signalwires
wiresand
andlong
longdynamic
dynamic
nodes
nodes(e.g.,
(e.g.,ones
onesthat
thatspan
spanmultiple
multiplebits
bitsininaadatapath).
datapath). Solutions:
Solutions: on
on
long
longroutes
routesaddadd“twists”
“twists”to
toavoid
avoidcontinguous
continguousroutes
routesor
orroute
routedynamic
dynamic
signals
signalsbetween
betweenmutually
mutuallyexclusive
exclusiveor
orcomplementary
complementarysignals.
signals.
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 12
 Domino Logic Design (I)
use Demorgan’s law
To convert to Domino-style design
A we need to create schematic that
B X
uses non-inverting gates:
C (1) look for CMOS gates followed
D by inverter
E Y (2) use Demorgan’s Law to create
F
G non-inv gates
H
Convert to Domino OR gate

Domino AND
A
B X
C
D
E Y
F
G
H Domino AND-OR
Domino OR
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 13
 X Y
Domino Logic
8/2 8/2
Design (II)
A E G

B D F

C nfet W/L = 4
pfet W/L = 8
CLK

s = static
d = domino (W/L = 4)
dd = domino (W/L = 8)

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 14

 Dual-rail Domino Logic
Domino circuits that generate both polarities of output

CLK

A CLK
A B
B

CLK
A A
A

B B
CLK

A CLK
A B
B

CLK

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 15

 Multiple-output Domino
Why stop at complementary outputs? There are interesting multiple-output
functions where there is a lot of sharing of nfets in the evaluate logic. For
example, in a carry-lookahead adder

CLK
Gi = AiBi C4
Pi =Ai + Bi P4 G4
C1 = G1 + P1C0 C3
C2 = G2 + P2G1 + P2P1Co P3 G3
C3 = G3 + P3G2 + P3P2G1 + C2
P3P2P1C0 P2 G2
C4 = G4 + P4G3 + P4P3G2 + C1
P4P3P2G1 + P4P3P2P1C0 P1 G1

C0

Domino version of the

Manchester carry chain

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 16

 Dual-rail “Keeper” Circuit

CLK
The cross-coupled pfets serve as
“keepers” for the output which is high
making the gate static rather than
dynamic! During precharge both
keepers are off; during the evaluate
A
phase, the output that goes low
A
switches on the keeper for the output
B
B that is staying high. Really solves
capacitive coupling problems with
dynamic logic in datapaths.
CLK

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 17

 p blocks n blocks n blocks p blocks
CMOS
NORA
pre eval pre
Logic
nfets pfets nfets

CLK eval CLK pre CLK eval

If we turn a dynamic gate “upside down” and use pfets to build the logic
block, we get a logic gate that “precharges” low and “discharges” high. By
using these gates in an alternating sequence with regular nfet dynamic gates
we can eliminate the race problem we had with nfet-only dynamic gate
sequences and hence we don’t need the buffer inverter present in domino
gates.
Removing the buffer is a mixed blessing since we may need it for drive reasons
and to keep compatibility with other domino gates. It also makes NORA logic
very susceptible to noise since during the evaluate phase all information is
stored dynamically.
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 18
 Dynamic Logic Summary
Advantages of dynamic logic:
This makes dynamic logic a
Š smaller area than fully static gates good choice for those parts
Š smaller parasitic capacitances hence of a circuit where the extra
higher speed engineering investment is
justified, e.g., along the
Š reliable operation if correctly designed. critical timing paths.

Concerns: Engineers who like this sort

of design will find this the sort
capacitive coupling to dynamic nodes of design they like!
charge sharing with dynamic nodes
subthreshold leakage currents in eval logic
minority carrier injection and latchup
alpha particle immunity
vdd/gnd noise and resistance

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 19

 Domino
1
Actively precharging

Life
2
Cycle
Waiting for precharge Waiting for data
(holding output value) (holding precharge)

4 3 4
3 CLK
Actively evaluating
1 2
The “9 O’clock” state is very interesting: once a Domino gate has finished
evaluating, the gate’s immediate predecessors can start to precharge
(forcing the gate’s inputs low) without affecting the value of the gate’s
output. The gate is acting as latch so long as its predecessors don’t start
another evaluate cycle. Perhaps we can build a pipeline of domino stages
where each stage serves as both logic and latch depending on where it is in its
cycle. Need to have each stage supply its own precharge/evaluate timing
dependent on what its neighbors are doing...
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 20
 0 = precharged
1 = evaluation done
Self-timed
Pipelines

P/E done? P/E done? P/E done?

F1 F2 F3

Simplest correctness rules:

Š a stage only precharges when both Sdone = 1
(a) its successor has finished evaluating
(it’s done with our values) Pdone = 0
So, what logic goes in
(b) its predecessor has finished precharging the clouds? And how
(old values are gone so we can’t use ‘em twice!) do we build the
Sdone = 0
Š a stage only evaluates when both “done?” boxes?
(a) its successor has finished precharging
(our new output won’t affect its stored value)
(b) its predecessor has finished evaulating
(there are new inputs for us to consider) Pdone = 1
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 21
 Muller C-Element
The Muller C-Element is the “AND” gate for self-timed logic because it
changes its output only after both inputs have changed. As shown here, it’s
an elegant implementation for both sets of rules on the previous slide.

Add weak feedback

inverter if we’re worried
about dynamic storage
for precharge/eval signal

P/E
Pdone Hey! Hasn’t the Muller C-
Element been around since
Sdone the days of Petri Nets?

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 22

 Completion Detectors
Self-timed logic
use dual-rail signaling (i.e., two wires) to encode
reset (not yet evaluated) 00
ready with value 0 01
ready with value 1 10
and then build handshake logic that starts next stage when
current stage is done and next stage has completed its
previous computation and delivered its values...

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 23

 1 = precharged
0 = evaluation done
Self-timed
Pipeline
C C C Latency
P/E done? P/E done? P/E done?

F1 F2 F3

Propagation through self-timed pipelines is constrained in both directions:

In the forward direction by how long it takes for the evaluate edge in
one stage to trigger the evaluate edge in the next stage:
LF = tFÇ + tDÈ + tCÇ

In the reverse direction by how long it takes for the precharge in one
stage to trigger a new evaluate in the stage after first evaluating the
previous stage (remember not double count!):
LR = 0.5*(tCÈ + tFÈ + tDÇ + tCÇ + tFÇ + tDÈ)
6.371 – Fall 2002 10/9/02 L11 – Domino Logic 24
 Further Improvements
We don’t have to delay evaluation until successor has finished its precharge
(signaling that it’s finished with our values). We can just check that
successor has started precharging… Even with this improvement, the
correct sequencing will still happen for any combination of precharge and
evaluate times for all the gates. We can modify the control element like so:

S P/E

P/E Eliminate the “extra” inverter for good

measure and use dynamic storage as
Pdone control element memory

Sdone We’re going to stop here, but there are

other improvements that can be made.
Hint: do we have to wait until the
predecessor is done computing new values
before starting our eval? etc., etc., etc.

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 25

 Quiz #1 Info
• Friday, 10/11, 80 mins: 11:05 – 12:25
• Two rooms: 31-161 [A – Li], 37-212 [Lu – Z]
• Open book (suggestion: bring lecture handouts)
• Calculator may come in handy

6.371 – Fall 2002 10/9/02 L11 – Domino Logic 26