lecture08 delay
TRANSCRIPT
-
8/11/2019 Lecture08 Delay
1/40
Lecture 8 ECE 425
Lecture 8 -- Gate-level Delay Estimation
-
8/11/2019 Lecture08 Delay
2/40
Lecture 8 ECE 425
Outline
Estimating delay of single gates and multi-gate circuits
-
8/11/2019 Lecture08 Delay
3/40
-
8/11/2019 Lecture08 Delay
4/40
Lecture 8 ECE 425
Pictorial View
-
8/11/2019 Lecture08 Delay
5/40
Lecture 8 ECE 425
Why do we care?
Its pretty obvious why we care about gate delay, in thatwe like circuits to go fast
Rise/fall times matter for a number of reasons
Theyre a component of total gate delay
While the inputs to a gate are rising or falling, aconductive path exists between power and ground
Power dissipation
Can potentially harm the chip if too much current
flows
For signals that have high inductance, overly short
rise/fall times can lead to di/dt-induced swings
Mostly relevant on chip I/O pins
-
8/11/2019 Lecture08 Delay
6/40
Lecture 8 ECE 425
Estimating Delay
Gate delays are determined by how quickly the drivinggate can charge/discharge its load capacitance
-
8/11/2019 Lecture08 Delay
7/40
Lecture 8 ECE 425
Estimating Delay
Gate delay may vary depending on which inputs arechanging -- generally use the worst case
-
8/11/2019 Lecture08 Delay
8/40
Lecture 8 ECE 425
General Approach
Divide circuit into DC-connected components, solve foreach component
-
8/11/2019 Lecture08 Delay
9/40
Lecture 8 ECE 425
Fall Time Analysis
During the fall time one or more nMOS transistorsdischarge the energy stored in the output capacitance
-
8/11/2019 Lecture08 Delay
10/40
Lecture 8 ECE 425
Fall Time Analysis
During the fall time, the nMOS transistor starts in thesaturated region and passes into the linear region
-
8/11/2019 Lecture08 Delay
11/40
Lecture 8 ECE 425
Fall Time Analysis
Divide fall time into two components: tf,satand tf,linear
In saturation, current through the transistor is constant
-
8/11/2019 Lecture08 Delay
12/40
Lecture 8 ECE 425
Fall Time
This becomes
Define t1, t2such that Vo(t1) = 0.9Vddand Vo(t2) = Vdd- Vt.
Then,
And tf, satis:
-
8/11/2019 Lecture08 Delay
13/40
Lecture 8 ECE 425
Fall Time
In the linear region, current through the transistordepends on Vo
And tf, linearbecomes
-
8/11/2019 Lecture08 Delay
14/40
Lecture 8 ECE 425
Fall Time
Integrating, we get tf, linear =
And
For many processes, Vt~= 0.2Vdd, allowing us toapproximate
-
8/11/2019 Lecture08 Delay
15/40
Lecture 8 ECE 425
Rise Time
Redoing the same analysis for the pMOS transistor in pull-up gives
Note that beta for pMOS tends to be about 1/2 beta for
nMOS given equivalent size devices, so typically want
pMOS about twice as wide as nMOS to get equivalent
rise and fall times
-
8/11/2019 Lecture08 Delay
16/40
Lecture 8 ECE 425
Gate Delay Estimation
Somewhat more complicated -- depends on rise and falltimes of the input signals.
Assuming (unrealistically) that the input rises or falls inzero time, then the gate delay can be approximated ashalf of the rise or fall time for the gate, and averaged to
Since gates are generally driven by other gates, we needa better approximation for real circuits
Could just simulate the design, but that wouldnt givemuch insight
Simulation is good for verifying that something worksthe way you want but not for designing it to do so
-
8/11/2019 Lecture08 Delay
17/40
Lecture 8 ECE 425
Circuit Delay Estimation
1. Divide circuit into DC-connected blocks as we talkedabout at the start of lecture.
2. Compute a simple delay model for each block
3. Add the delays for each block to get overall delay.
In CMOS, a DC-connected block (stage) will be either:
1. A single logic gate
2. A transmission-gate network and the gates driving it
In the simple case, a block switches in response to a single
signal changing at one of its inputs, called the trigger
signal
-
8/11/2019 Lecture08 Delay
18/40
Lecture 8 ECE 425
Making the System Tractable
As weve seen, solving even simple transistor networksanalytically requires solving differential equations that
change at region boundaries
For example, the rise and fall time derivations we
presented really only apply for inverters, and
computing them for larger gates is more complex
Now think about doing that for a 100-transistor
network, much less anything bigger
Solution: linearizethe transistor equations by replacingeach transistor with an equivalent resistor, Reff
-
8/11/2019 Lecture08 Delay
19/40
Lecture 8 ECE 425
Computing Reff
Refffor a transistor depends on what region its operatingin
Example: Reff= dVds/dIds = infinity in saturation
In linear region,
During switching, a transistor travels through different
regions of operation
We want a single Reffthat approximates the behavior of atransistor during the entire rise/fall time
Reffwill be different depending on whether the
transistor is transmitting a zero or a one.
-
8/11/2019 Lecture08 Delay
20/40
Lecture 8 ECE 425
Case One -- nMOS Transistor, Logic 0
nMOS transistor, discharging capacitor to ground
-
8/11/2019 Lecture08 Delay
21/40
Lecture 8 ECE 425
nMOS Transistor, Logic 0
Want to model as a single network with Reff, C
Choose Reffso that both circuits have the same fall time
-
8/11/2019 Lecture08 Delay
22/40
Lecture 8 ECE 425
nMOS Transistor, Logic 1
Back to the I-V curve
-
8/11/2019 Lecture08 Delay
23/40
Lecture 8 ECE 425
nMOS Transistor, Logic 1
Problem: VOnever makes it to 0.9 Vddin this case,assuming Vt= 0.2Vdd, so cant try to match the full rise
time
If we try to match the time it takes to hit 0.5Vdd, we get
Note that this is about 3x larger than the dischargingcase. Why does this make sense?
-
8/11/2019 Lecture08 Delay
24/40
Lecture 8 ECE 425
pMOS Transistors
-
8/11/2019 Lecture08 Delay
25/40
Lecture 8 ECE 425
Estimating Stage Delay
Need to decide whether each transistor is off, transmittinga logic 0, or transmitting a logic 1.
Then, can replace the stage with an RC network
-
8/11/2019 Lecture08 Delay
26/40
Lecture 8 ECE 425
Estimating Stage Delay
This is stilltoo complex -- solving these networks exactlyrequires solving several simultaneous differential
equations.
Simplify further by reducing to an equivalent single-RC
network with time constant
-
8/11/2019 Lecture08 Delay
27/40
Lecture 8 ECE 425
Example of the Process
-
8/11/2019 Lecture08 Delay
28/40
Lecture 8 ECE 425
Estimating Stage Delay
At most one node should be the input node to the stage,corresponding to either Vddor Vss
Generally, were interested in the delay at a specific
output node
Question: how to compute Reqand Ceq? When the RC-network is a tree (no loops), there is a good
estimate that is also easy to compute.
This estimate of is called the Elmore Time
Constant
-
8/11/2019 Lecture08 Delay
29/40
Lecture 8 ECE 425
Estimating the Stage delay
1. Identify an input node and an output node, and theunique path between them. Call the path P
2. For every node n in the tree other than the input node
1. Identify the unique path between the input node and n
2. Identify the sub-path that is the intersection of thispath and P
3. Sum the resistances along the sub-path, call it Rn4. If Cn is the capacitance at n, create a !n= RnCn
3. Sum the !n to get
!
eqfor the network
4. Given !eq, can select Reqand Ceqvalues to match, but
generally dont need to, !eqis enough
-
8/11/2019 Lecture08 Delay
30/40
Lecture 8 ECE 425
Example
-
8/11/2019 Lecture08 Delay
31/40
Lecture 8 ECE 425
Example
-
8/11/2019 Lecture08 Delay
32/40
Lecture 8 ECE 425
One Note
If a node n is already at the voltage it will have after thetransition, set !n= 0, since the node will not be charging
or discharging
Example:
-
8/11/2019 Lecture08 Delay
33/40
Lecture 8 ECE 425
Cascade of Stages
Suppose stage 1 feeds stage 2
If the time constant of stage 1 is !1, when is stage 2
triggered? (I.e., when does it start switching)
Time for yet another approximation
-
8/11/2019 Lecture08 Delay
34/40
Lecture 8 ECE 425
Cascade of Stages
By convention, assume that stage 2 is triggered when theoutput of stage 1 has completed 90% of its logic swing
and is thus well within Vtof one of the rails
Thus, the delay through stage 1 is
Thus, if stage 1 triggers at time 0, stage 2 triggers at time2.3 !1, stage 3 triggers at time 2.3 !1 + 2.3 !2, etc.
-
8/11/2019 Lecture08 Delay
35/40
Lecture 8 ECE 425
Interconnect Delay
Interconnect delay can be a big issue, particularly forsystems larger than a few gates
Interconnect delay is a function of interconnect length
To be fully accurate, youd treat interconnects as
transmission lines and analyze them that way Were not going to do that, surprise surprise
-
8/11/2019 Lecture08 Delay
36/40
Lecture 8 ECE 425
Interconnect Delay
Instead, break long lines into lumpedRC-segments
-
8/11/2019 Lecture08 Delay
37/40
Lecture 8 ECE 425
Interconnect Delay
-
8/11/2019 Lecture08 Delay
38/40
Lecture 8 ECE 425
Interconnect Delay
-
8/11/2019 Lecture08 Delay
39/40
Lecture 8 ECE 425
Rules of Thumb
Formally, lowest delay of an interconnect occurs when thedelay of each wire segment equals the delay of the buffer
With this model, interconnect delays become linear, not
quadratic
-
8/11/2019 Lecture08 Delay
40/40
Lecture 8 ECE 425
Next Time
Logical effort: a way of making circuits fast, not justknowing how slow they are.