Gate-Level Simulation Methodology Improving Gate-Level Simulation Performance

Objective
The purpose of this document is to present the best practices that can help
improve the performance of gate-level simulation (GLS). These best practices
have been collected from our experience in gate-level design, and also based
on the results of the Gate-Level Methodology Customer Survey carried out by
Cadence.
Starting GLS early is important because netlist modifications can continue
late into the design cycle, and are driven by timing-, area-, and power-closure
issues. It is also equally important to reduce the turn-around time for GLS
debug by setting up a flow that enables running meaningful GLS focused
on specific classes of bugs, so that those expensive simulation cycles are not
wasted toward re-verifying working circuits.
This application note describes new methodologies and simulator use models
that increase GLS productivity, focusing on two techniques for GLS to make
the verification process more effective:
• Extracting information from the tools applied to the gate netlist, such as
static timing analysis and linting. Then, passing the extracted information to
the GLS.
• Improving the performance of each GLS run by using recommended tool
settings and switches
These approaches can help designers focus on the verification of real gate-level
issues and not spend time on issues that can be caught more efficiently with
other tools. The new methodologies and simulator use models described in this
document can increase GLS productivity, which is important because produc-
tivity measures both the number of cycles executed and the throughput of the
simulator. Addressing the latter without addressing the former results in only
half of the productivity benefit.
Gate-Level Simulation Methodology
Improving Gate-Level Simulation Performance
Author: Gagandeep Singh
The increase in design sizes and the complexity of timing checks at 40nm technology nodes
and below is responsible for longer run times, high memory requirements, and the need for a
growing set of gate-level simulation (GLS) applications including design for test (DFT) and low-
power considerations. As a result, in order to complete the verification requirements on time,
it becomes extremely important for GLS to be started as early in the design cycle as possible,
and for the simulator to be run in high-performance mode. This application note describes
new methodologies and simulator use models that increase GLS productivity, focusing on two
techniques for GLS to make the verification process more effective.
Contents
Objective......................................1
Introduction..................................2
Gate-Level Simulation Flow
Overview......................................2
Why Gate-Level Simulation is
Required...................................2
Techniques to Improve Gate-Level
Performance.................................3
Improving Gate-Level Simulation
Performance with Incisive
Enterprise Simulator..................3
A Methodology for Improving
Gate-Level Simulation.............18
Summary....................................34
Contacts.....................................34

The first part of this document presents information on fine-tuning Cadence®
Incisive®
Enterprise Simulator to
maximize cycle speed and minimize memory consumption. The second part is dedicated to general steps designers
can take with the combination of any simulator, synthesis tool, DFT engine, and logic equivalence checking (LEC)
engine.
While only Incisive Enterprise Simulator users will find real benefits in the first section, all GLS users will find value
in this entire application note.
Introduction
Project teams are finding that they need more GLS cycles on larger designs as they transition to finer process
technologies. The transition from 65nm to 40nm, and the growth of 28nm and finer design starts, are driven by
the need to access low-power and mixed-signal processes (data from IBS, Inc.). These design starts do account for
some required increase in GLS cycles because they translate into larger designs. However, it is the low-power and
mixed-signal aspects, as well as the new timing rules below 40nm, of these designs that are creating the need to
run more GLS simulations. Given that the GLS jobs tend to require massive compute servers, and run for hours,
even days and weeks, they are creating a strain on verification closure cycles.
Simulation providers are continuing to improve GLS execution speed and reduce memory consumption to keep up
with demand at the finer process nodes. While faster engines are part of the solution, new methodologies for GLS
are also required.
A closer examination shows that many teams are using the same approaches in test development and simulation
settings to run GLS today as they did in the 1980s when Verilog-based GLS began. Given that the size of today’s
designs was almost inconceivable in the 1980s, and the dependencies on modern low-power, mixed-signal, and
DFT didn’t exist then, new methodologies are now warranted. The sections in this document describe these new
methodologies in detail.
Gate-Level Simulation Flow Overview
The typical RTL to gate-level netlist flow is shown in the following illustration.
Testbench Verification
RTL
Synthesis
Linting
ATPG Pattern Simulation
Gate-Level Netlist
STA
Logic
Equivalence
Check
Figure 1: Gate-Level Simulation Flow
Why Gate-Level Simulation is Required
GLS can catch issues that static timing analysis (STA) or logical equivalence tools are not able to report. The areas
where GLS is useful are in:
1. Overcoming the limitations of STA, such as:
–
– The inability of STA to identify asynchronous interfaces
–
– Static timing constraint requirements, such as those for false and multi-cycle paths
www.cadence.com 2

2. Verifying system initialization and that the reset sequence is correct
3. DFT verification, since scan-chains are inserted after RTL synthesis
4. Clock-tree synthesis
5. For switching factor to estimate power
6. Analyzing X state pessimism or an optimistic view, in RTL or GLS
Techniques to Improve Gate-Level Performance
This section gives you key techniques that can help reduce GLS run time and debug time. The section is presented
in two parts:
1. Improving Gate-Level Simulation Performance with Incisive Enterprise Simulator
2. A Methodology for Improving Gate-Level Simulation
Improving Gate-Level Simulation Performance with Incisive Enterprise Simulator
This section describes techniques that can help improve the performance of GLS by running Incisive Enterprise
Simulator in high-performance mode using specific tool features.
1. Applying More Zero-Delay Simulation
While timing simulations do provide a complete verification of the design, during the early stages of GLS, when
the design is still in the timing closure process, more zero-delay simulations can be applied to verify the design
functionality. Simulations in zero-delay mode run much faster than simulations using full timing.
Zero-delay mode can be enabled using the -NOSpecify switch. This option works for Verilog designs only and
disables timing information described in specify blocks, such as module paths and delays and timing checks. For
negative timing checks, delayed signals are processed to establish correct logic connections, with zero delays
between the connections, but the timing checks are ignored.
Since zero-delay mode can introduce race conditions into the design, and can also introduce zero delay loops,
Incisive Enterprise Simulator has many built-in delay mode control features that can help designers run zero-delay
simulations more effectively. These features are listed in the following sections.
1.1 Controlling Gate Delays
Incisive Enterprise Simulator provides delay mode control through command-line options and compiler directives to
allow you to alter the delay values. These delays can be replaced in selected portions of the model.
You can specify delay modes on a global basis, or on a module basis. If you assign a specific delay mode to a
module, then all instances of that module simulate in that mode. Moreover, the delay mode of each module is
determined at compile time and cannot be altered dynamically. There are delay mode options that use the plus (+)
option prefix, and the 12.2 and later releases of the tool include a minus (-) delay mode option.
The (+) options listed below are ordered from the highest to lowest precedence. When more than one plus option
is used on the command line, the compiler issues a warning and selects the mode with the highest precedence.
delay_mode_path
This option causes the design to simulate in path delay mode, except for modules with no module path delays. In
this mode, Incisive Enterprise Simulator derives its timing information from specify blocks. If a module contains a
specify block with one or more module path delays, all structural and continuous assignment delays within that
module, except trireg charge decay times, are set to zero (0). In path delay mode, trireg charge decay remains
active. The module simulates with black box timing, which means it uses module path delays only.
www.cadence.com 3

delay_mode_distributed
This option causes the design to simulate in distributed delay mode. Distributed delays are delays on nets, primi-
tives, or continuous assignments. In other words, delays other than those specified in procedural assignments and
specify blocks simulate in distributed delay mode. In distributed delay mode, Incisive Enterprise Simulator ignores all
module path delay information and uses all distributed delays and timing checks.
delay_mode_unit
This option causes the design to simulate in unit delay mode. In unit delay mode, the tool ignores all module path
delay information and timing checks, and converts all non-zero structural and continuous assignment delay expres-
sions to a unit delay of one (1) simulation time unit.
delay_mode_zero
This option causes modules to simulate in zero-delay mode. Zero-delay mode is similar to unit delay mode in that
all module path delay information, timing checks, and structural and continuous assignment delays are ignored.
Reasons for Selecting a Delay Mode
Replacing delay path, or distributed with global zero or unit delays, can reduce simulation time by an appreciable
amount. You can use delay modes during design debugging phases, when checking design functionality is more
important than timing correctness. You can also speed up simulation during debugging by selectively disabling
delays in sections of the model where timing is not currently a concern. If these are major portions of a large
design, the time saved can be significant.
The distributed and path delay modes allow you to develop or use modules that define both path and distributed
delays, and then to choose either the path delay or the distributed delay at compile time. This feature allows you to
use the same source description with all the Veritools and then select the appropriate delay mode when using the
sources with Incisive Enterprise Simulator. You can set the delay mode for the tool by placing a compiler directive
for the distributed or path mode in the module source description file, or by specifying a global delay mode at run
time.
The -default_delay_mode option has been added to enable command-line control of delay modes at the module
level and is available in release version 12.2 and later. An explicit delay mode can be applied to all the modules that
do not have a delay mode specified by using one of the following command-line options:
-default _ delay _ mode <arg>
-default _ delay _ mode full _ path[...]=delay _ mode
Typical Use Model
Typically, models are shipped with compiler directives that enable a specific delay _ mode and features. For
example:
ìfdef functional
`delay _ mode _ distributed
`timescale 1ps/1ps
èlse
`delay _ mode _ path
`timescale 1ns / 1ps
èndif
This means the default delay _ mode is the delay _ mode _ path, with path delays defined within a specify block
being overwritten by the values in the SDF.
www.cadence.com 4

Also, you can use the functional macro at compilation time to get a faster representation, with usually #1
distributed delays (often on the buffer of the output path).
However, in some cases, it is important to be able to override this delay _ mode, which is why the different modes
are available.
Delay Mode Summary Table
Delay Modes Zero Unit Distributed Path Default
Timing Ignored Ignored Ignored Used Used
#delays Set to 0 Set to 1 except
null
Used Ignored Used
Controlling Delays in Your Model
Keyword Specifies
-DELAY_MODE [<FULL|PATH>[...]=]<MODE> Specify a delay_mode for all/selected
-DEFAULT_DELAY_MODE DELAY_MODE Specify a delay_mode for all
1.2 Identifying Zero-Delay Loops
Apart from static tools, Incisive Enterprise Simulator also has a built-in feature that detects potential zero-delay
gate loops and issues a warning if any are detected.
This feature can be enabled by using -GAteloopwarn (Verilog only) on the command line.
This option can help identify zero-delay gate oscillations in gate-level designs. The option sets a counter limit on
continuous zero-delay loops. When the limit is reached, the simulation stops and a warning is generated stating
that a possible zero-delay gate oscillation was detected. You can also use a TCL command at the ncsim prompt to
detect the loop. The simulation will stop after the number of delta cycles specified hits a specified number.
ncsim>stop -delta <number> -timestep -delbreak 1
Once the loop is detected, you can then use the Tcl drivers -active command to identify the active signals and
trace these signals to the zero-delay loop.
The detected loop can be fixed by adding delays to the gates that are involved in the loop.
For example, the image below shows a zero-delay loop in the design. Once the loop is detected using Tcl stop
command or –gateloopwarn, you can add delay either on nand_4 (instance udly) or test _ buf _ noX (ub
instance) to fix the loop issue.
1.3 Handling Zero-Delay Race Conditions
Races in zero-delay mode can occur because the delay for all gates is zero by default. The following sections show
you how these races can be fixed.
1.3.1 Updating the Design by Adding # Delays
You can use # delays in designs to correct race conditions. For example, consider the following simple design.
www.cadence.com 5

FF1 Failing FF2
Data Out
Data In
Clock
D2
Race
Condition
Combinational
Logic
Figure 2: Design with a Race Condition
In the figure above, the simulation is running in zero-delay mode and there is some combinational logic in the clock
path. In such a case, a race condition can occur and the data at FF2 might get latched at the same clock edge.
In order to fix this race condition, you can add a unit delay #1 at the output of FF1. This will delay the output of
FF1. Consequently, the data out from FF1 will get latched at FF2 at the next clock edge.
The following line is an example of adding a buffer at the output of FF1 with a unit delay:
buf #1 buf _ prim _ delay1 (D2 _ FF2, Data _ OUT _ FF1); //Add buffer with unit delay.
1.3.2 Using SDF Annotation for Race Conditions
For this example, consider the same design shown in Figure 2. You can add a delay in SDF for FF1 instead of adding
it directly in the design code. When the design is simulated with SDF, the delays will remove the race conditions.
1.3.3 Sequential Circuits Have Unit Delays
For this case, consider the design in Figure 2. You can add unit delays by default for all sequential cells in the
design library. All the sequential cells will then have a unit delay, while combinational cells will have zero delay. For
example, the following listing shows adding a buffer at the output in the flip-flop:
module dff (clock ,din, q);
........... // Original Flip-Flop logic
buf #1 buf _ prim _ delay1 (q, q _ internal); // Add buffer with unit delay.
endmodule
1.3.4 Incisive Enterprise Simulator Race Condition Correcting Features
The tool has built-in features that can help designers fix race conditions in the design. The following sections
present these features.
1. The SEQ_UDP_DELAY delay_specification Argument
This feature applies a specified delay value to the input/output paths of all sequential UDPs in the design. The shift
register example in the following figure illustrates this.
www.cadence.com 6

CK
CK
D
D
CK
BCK
Q
QN
QN
Qout
data
Buf1
delay 0ns
flop1
delay 0ns
flop2
-delay_mode_zero makes the delays 0
Value from D can be clocked
through flop1and flop2 on
the same clock edge
When -delay_mode_zero is applied, the delays from
CK to BCK and from CK to Q are zero. This causes BCK and
data to transition at the same time, and potentially
allows the changed value of data to also be seen by
flop2 on the same clock edge.
Figure 3: Applying a Delay Value to I/O Paths of Sequential UPDs
Use the -seq _ udp _ delay option to set all of the delays to zero (as if you used the -delay _ mode _ zero
option) except for the sequential UDPs. The delay specified with -seq _ udp _ delay overrides any delay specified
for the sequential UDPs in the design that are in specify blocks, through SDF annotation, and so on. The option
also removes any timing checks associated with the sequential UDPs.
The delay _ specification argument can be a real number, or a real number followed by a time unit. The time
unit can be fs, ps, ns, or us. If no time unit is specified, ns is the default.
Examples
The following options assign a 10 ns delay to all sequential UDP paths.
-seq _ udp _ delay 10
-seq _ udp _ delay 10ns
The following option assigns a 0.7 ns delay to all sequential UDP paths.
-seq _ udp _ delay 0.7ns
The following option assigns a 5 ps delay to all sequential UDP paths.
-seq _ udp _ delay 5ps
The following figure illustrates the effect of using -seq_udp_delay for the shift register example.
www.cadence.com 7

CK
CK
D
D
CK
BCK
Q
QN
QN
Qout
data
Buf1
delay 0ns
flop1
delay 0ns
flop2
-seq_udp_delay 50ps adds a delay
of 50ps between CK and Q
Value from D is clocked
through flop1 on the first
CK edge and through flop2
on the second CK edge
When -seq_udp_delay is applied, it adds a delay from
CK to Q (50ps in this example}. This causes data to
transition 50ps after CK and BCK, and causes the change
in data to be seen by flop2 on the next clock edge.
Figure 4: Using –seq_udp_delay for a Shift Register
Note that seq _ udp _ delay also does the following:
• Adds a delay to the sequential UDPs
• Implies -DELAY _ MODE ZERO
• Implies -NOSPECIFY
2. ADD_SEQ_DELAY delay_specification
As mentioned above, seq _ udp _ delay also internally implies delay _ mode _ zero and nospecify mode. In
case you are not interested in this mode, add _ seq _ delay can be used instead. This will add a non-zero delay to
the UDPs. The add _ seq _ delay option can also be applied for a specific instance.
For example:
-add _ seq _ delay top.dut _ top.u3 40ps
3. DELTA_SEQUDP_DELAY delay_specification
If you are not interested in adding UDP delays, delta _ sequdp _ delay can also be used. In most cases, just
adding a delta delay to the UDP is enough to avoid the race condition, and delta _ sequdp _ delay also adds less
overhead than a non-zero delay.
4. TIMING FILE
Unit delays can also be controlled through the new tfile feature, which is available in release version 12.2 and
later. Using this feature, you can add unit delays in cells. This can be done at both the cell level and instance level.
To add unit delays at the cell level, the tfile syntax is:
CELLLIB L.C:V ADDUNIT //Add unit primitive delays
Example
CELLLIB worklib.sdffclrq _ f1 _ hd _ dh ADDUNIT //Adds unit delay to cell sdffclrq _ f1 _ hd _ dh
Similarly, to add a delay at the instance level, the tfile syntax is:
PATH <TOP>.<INST1>.<INST2> ADDUNIT //Adds unit delay to instance INST2
Example
PATH tb.top1.clock _ div1.divide0 ADDUNIT //Adds unit delay to instance divide0
www.cadence.com 8

Overriding Default Cell Delays
The timing file feature also provides a mechanism to override the default delays specified at the cell level. As
mentioned in section 1.3.3 Sequential Circuits Have Unit Delays, in case #0 and #1 delays have already been added
in the cell library, and you can override these values using the ALLUNIT, UNIT, or ZERO keywords. This override
works only with the `delay _ mode _ distributed compilation directive. The following lines show this.
CELLIB L.C:V ALLUNIT //all non-zero and zero primitive delays will be changed to unit delay
CELLLIB L.C:V UNIT //all non-zero primitive delays will be changed to unit delay and zero
delay (#0) will remain as it is
CELLLIB L.C:V ZERO //all non-zero primitive delays will be changed to zero delay
Make Zero Delay
CELLLIB worklib.sdffclrq _ f1 _ hd _ dh zero
This feature also works at an instance level, as follows:
PATH <TOP>.<INST1>.<INST2> zero //Adds zero delay to instance INST2
Similarly, you can use the unit or allunit keywords.
2. Improving the Performance of Gate-Level Simulation with Timing
2.1 Timing Checks
Timing checks have a significant impact on performance, so it is recommended that you disable timing checks if
they are not required. Timing checks can also be enabled partially, only for required blocks, by passing a timing file
to Incisive Enterprise Simulator. The following sections present features that can help you control timing checks.
2.1.1 Use of the –NOTImingchecks Option
The NOTImingchecks option prevents the execution of timing checks. This option turns off both Verilog and accel-
erated VITAL timing checks.
Please Note: The -notimingchecks option turns off all timing checks. Because the timing checks have been
turned off, any calculation of delays that would normally occur because of negative limits specified in the timing
checks is disabled. If your design requires that these delays be calculated in order for the design to simulate
correctly, use the -ntcnotchks option mentioned below instead.
2.1.2 Controlling Timing Through a tfile
You can turn off timing in specific parts of a Verilog design by using a timing file, which you specify on the
command line with the -tfile option. The timing file can help you disable timing for selected portions of a design.
For example:
% ncelab -tfile myfile.tfile [other _ options] worklib.top:module
% irun -tfile myfile.tfile [other _ options] source _ files
If you are annotating with an SDF file, the design is annotated using the information in the SDF file, and then the
timing constructs that you specify in the timing file are removed from the specified instances.
Using a timing file does not cause any new SDF warnings or remove any timing warnings that you would get
without a timing file. However, there is one exception, as follows:
The connectivity test for register-driven interconnect delays happens much later than the normal interconnect
delays. Any warning that may have existed for that form of the interconnect will not be generated if that
interconnect has been removed by a timing file.
The timing specifications you can use in the timing file are listed in the following table:
Timing Specification Description
-iopath Disable module path delays.
www.cadence.com 9

Timing Specification Description
+iopath Enable module path delays.
-prim Disable primitive delays. Sets any primitive delay within the specified
instance(s) to 0.
+prim Enable primitive delays within the specified instance(s).
-port Remove any port delays at the specified instance(s) or any interconnects
whose destination is contained by the instance. interconnect sources are not
affected by the -port construct.
+port Enable port delays at the specified instance(s) or any interconnects whose
destination is contained by the instance.
[list _ of _ tchecks]
-tcheck
Remove the listed timing checks from the instance(s).
[list _ of _ tchecks]
+tcheck
Enable the listed timing checks in the instance(s).
If specific timing checks are not listed, all timing checks will be disabled or
enabled.
Examples:
// Disable all timing checks in top.foo
PATH top.foo –tcheck
// Disable $setup timing check in top.foo
PATH top.foo $setup -tcheck
// Disable $setup and $hold timing checks in top.foo
PATH top.foo $setup $hold -tcheck
-timing
+timing
This is an alias for the four specifications shown above.
Example Timing File
The following listing shows an example timing file:
// Disable all timing checks in top.inst1
PATH top.inst1 -tcheck
/* Disable all timing checks in all scopes below top.inst1 except
for instance top.inst1.U3 */
PATH top.inst1... -tcheck
PATH top.inst1.U3 +tcheck
/* Disable $setup timing check for all instances under top.inst2,
except for top.inst2.U1. */
PATH top.inst2... $setup -tcheck
PATH top.inst2.U1 $setup +tcheck
// Disable $setup and $recrem timing checks for instance top.inst3
PATH top.inst3 $setup $recrem -tcheck
// Disable timing checks for all objects in the library mylib
CELLLIB mylib -tcheck
// Disable module path delays for all instances in the 2nd levels of hierarchy
PATH *.* -iopath
2.1.3 Use of the –NTCNOtchks Option
This option generates negative timing check (NTC) delays, but does not execute timing checks. The option is
available for Verilog designs only.
www.cadence.com 10

You can use the -notimingchecks option to turn off all timing checks in your design. However, if you have
negative timing checks in the design, this option also disables the generation of delayed internal signals, and you
may get the wrong simulation results if the design requires these delayed signals to function correctly. That is, if
you have negative timing checks, simulation results may be different when using -notimingchecks and without
-notimingchecks.
Use the -ntcnotchks option instead of the -notimingchecks option if you want the delayed signals to be
generated but want to turn off timing checks. This option removes the timing checks from the simulation after the
negative timing check (NTC) delays have been generated.
Controlling Timing Checks with the -NTCNOtchks Option Using a tfile
By default, all timing checks are disabled by the -ntcnotchks option. However, in case you are interested in the
timing violations for some portion of the design (typically the top block), the timing checks can be controlled using
the tfile timing file.
Please Note: This feature is available from release version 12.2 onwards.
The keyword ntcnotchks is available for use in the timing file. The keyword controls the -ntcnotchks option
at the instance level. Using the keyword enables the timing checks, and a delayed net will be generated for those
instances using the -ntcnotchks option.
Note: The – or + prefix is not required for the ntcnotchks option in the tfile. If ntcnotchks is present in the file,
it is always enabled.
Some example cases are presented here.
Case 1
Enable negative timing computation delays for all instances, with timing checks enabled for a particular instance
only.
Timing File Sample
//Enables Negative timing computation delays on the entire design but turns off timing
checks on all instances.
DEFAULT ntcnotchks
//Enables timing checks for a particular instance.
path top.u1 +tcheck
Case 2
Selectively enable negative timing check delays and timing checks.
Timing File Sample
//Disables timing checks & negative timing computation delays on the entire design
DEFAULT –tcheck
//Enables timing check & ntcnotchks for a particular instance.
path top.u1.u2 +tcheck
//Enables negative timing computations for a particular instance and disables tchecks
path top.u1.u3 ntcnotchks
path top.u1.u4 ntcnotchks
2.1.3 Controlling the Number of Timing Check Violations
The results of GLS are often not usable if timing violations are reported. Consequently, many designers do not
want to continue the simulation further if a timing violation is reported.
When timing violations are reported, you can cause the simulation to immediately exit by using the following
switch:
-max _ tchk _ errors <number>
www.cadence.com 11

When this option is specified, the value of the number argument specifies the maximum number of allowable
timing check violations before the simulation exits.
Please Note: This feature is available in release 12.2 and onwards.
2.1.4 Disabling Specific Timing Checks During Simulation
The tcheck command turns timing check messages and notifier updates on or off for a specified Verilog instance.
The specified Verilog instance can also be an instance of a Verilog module instantiated in VHDL.
Command syntax:
tcheck instance _ path -off | -on
2.1.5 Table - Disabling Timing Information
Keyword Specifies
-NOSPecify Disable timing information described in specify blocks and SDF annotation
-NOTImingchecks Do not execute timing checks.
-NONOtifier Ignore notifiers in timing checks
-NO _ TCHK _ msg Suppress the display of timing check messages, while allowing delays to be
calculated from the negative limits
-NTCNOtchks Generate negative timing check (NTC) delays, but do not execute timing
checks (in cases where you want the delayed signals to be generated but
want to turn off timing checks).
2.2 Removing Zero-Delay MIPDs, MITDs, and SITDs
You can improve performance by removing interconnect or port delays that have a value of 0 from the SDF file. The
SDF Annotator parses and interprets zero-delay timing information but does not annotate it. By removing the zero-
delay information from the SDF file, you eliminate unnecessary processing of this information.
Note: This performance improvement recommendation applies only to MIPDs, MITDs, and SITDs.
3. Improving Simulation Performance Using Incisive Enterprise Simulator Multi-Core Solution
The benefit of multiple CPU cores can be obtained in two ways. One way is to run separate simulations on all
available CPU cores. This is useful when running many tests in parallel, as in a regression environment. However,
running tests in parallel does not reduce the total simulation time for individual designs. To do this, different parts
of the design must be run on different cores in parallel.
If you are concerned about how to reduce a long-running simulation, you could use the Incisive Enterprise
Simulator multi-core capability, available in release 13.1 and onwards, to run the design under test (DUT) in parallel.
This approach is called design-level parallelism (DLP). DLP is implemented when multiple threads are run on
different cores to achieve parity.
To specify the parts of the DUT that are to run in parallel on separate threads, you write a partition file. This file
specifies which portions of the hierarchy belong to each partition. The partition file is included at elaboration time
using the -mcfile option (ncelab -mcfile or irun -mcfile). This is the only option required for running in multi-
threaded mode. The following section explains the flow of creating a partition file by using different sources of
information.
Multi-Core Use Model
The multi-core solution works well for scenarios where a simulation has enough parallel activities. In general, an
automatic test pattern generation (ATPG) pattern would activate almost all the logic in the DUT at every clock
cycle, which makes GLS DFT simulation (ATPG pattern) a good candidate for the Incisive Enterprise Simulator multi-
core solution.
The table below shows one of the designs with timing having good performance improvement.
www.cadence.com 12

Machine/SW Intel E5-4650 0@2.70GHz, 20M/IES13.1s09
Ncvlog ncelab ncsim SimTime
Gain
Total Time Overall
Time Gain
Single Core 30.5s 1278.4s 2816.8s 1x 4125.7 1X
Memory usage 328.5MB 8492.7MB 7347MB
2 Cores 30.6 1349.7 1702.6 1.65X 3082.9 1.34X
Memory usage 328.5MB 8493.1MB 7509.2MB
4 Cores 30.5 1340.7 1318.4 2.13X 2689.6 1.53X
STIL File
A STIL file is a standard way to describe the scan data. Some of the ATPG tools, such as Cadence Encounter®
Test,
could generate the STIL with scan-chain information. The scan-chain information would list the scan flip-flops on
each scan chain. Partial STIL fill contents are shown below. You can extract the scan cells information to find out
the flip-flop numbers in each instance.
STIL.FULLSCAN.full_all.signals.stil
ScanStructures {
ScanChain “Control_Observe_Reg_1_FULLSCAN” {
ScanLength 741;
ScanIn “port_pad_data_in[0]”;
ScanCells
“COMPACTOR.compressor.mreg2_reg_0_.__i0.dff_primitive”
…
ScanOut “port_pad_data_out[0]”;
ScanInversion 0; }
ScanChain “Control_Observe_Reg_2_FULLSCAN” {
ScanLength 740;
ScanIn “port_pad_data_in[1]”;
ScanCells
…
“DTMF_REG1_digit_out_reg_14_.__i2.dff_primitive”
“DTMF_REG1_digit_out_reg_15_.__i2.dff_primitive” ;
ScanOut “port_pad_data_in[3]”;
ScanInversion 0; }
}
www.cadence.com 13

The Limitation of Using STIL
Not every tool provides the capability to write out the scan chain information.
Parallel Scan Testbench
In order to reduce the runtime of scan patterns, the ATPG tools usually generate parallel scan patterns which
will directly load the scan data to scan flip-flops by using Verilog force statement. It is expected to have the scan
flip-flop information in the Verilog testbench. It is possible to extract the flip-flop numbers in each instance from
the testbench.
VER.FULLSCAN.full_all.mainsim.v
assign ( supply0, supply1 ) // CR = 1
dtmf_chip_inst.port_pad_data_in[0] = part_SLs_1[0741]!==1’bZ ? part_SLs_1[0741] : 1’bZ
, // pinName = port_pad_data_in[0]: CR = 1: pos = 1: index = 21: nodeID = 22: latchEntry
= 34: :
dtmf_chip_inst.COMPACTOR.compressor.mreg2_reg_1_.D = part_SLs_1[0740]!==1’bZ ?
part_SLs_1[0740] : 1’bZ , // pinName = COMPACTOR.compressor.mreg2_reg_1_.D: CR = 1:
pos = 2: index = 2061: nodeID = 376: latchEntry = 36: :
…..;
Limitation
Different ATPG tools have different testbench coding styles. You need to understand the format of these
testbenches in order to extract the information. For details on the multi-core solution, please refer to the
Multi-Core Simulation in the Incisive Enterprise Simulator document in Cadence Help.
4. Improving Performance in Debugging Mode
www.cadence.com 14

4.1 Wave Dumping
Waveform dumping negatively impacts simulation performance, so it should be used only if required. If the
waveform dumping activity is high, you can use the parallel waveform mcdump feature of ncsim. Waveform dumps
inside cells/primitives should be turned off (using probe … -to_cells).
Enabling Multi-Process Waveform Dumping
If you are running the simulator on a machine with multiple CPUs, you can improve the performance of waveform
dumping, and decrease peak memory consumption, by using the -mcdump option (for example, ncsim -mcdump
or irun -mcdump). In multi-process mode, the simulator forks off a separate executable called ncdump, which
performs some of the processing for waveform dumping. The ncdump process runs in parallel on a separate CPU
until the main process exits. You can see a 5% to 2X performance improvement depending upon the number of
objects probed.
4.2 Access
By default, the simulator runs in a fast mode with minimal debugging capability. To access design objects and lines
of code during simulation, you must use ncelab command-line debug options. These options provide visibility into
different parts of a design, but disable several optimizations performed inside the simulator.
For example, Incisive Enterprise Simulator includes an optimization that prevents code that does not contribute in
any way to the output of the model from running. Because this dead code does not run, any runtime errors, such
as constraint errors or null access de-references, that would be generated by the code are not generated. Other
simulation differences (for example, with delta cycle-counts and active time points) can also occur. This dead
code optimization is turned off if you use the ncvlog -linedebug option, or if read access is turned on with the
ncelab -access +r option.
The following line shows the ncelab –access debug command and command-line options that provide additional
information and access to objects, but reduce simulator speed. Because these options slow down performance, you
should try to apply them selectively, rather than globally.
ncelab –access +[r][w][c]
r provides read access to objects
w gives write access to objects
c enables access to connectivity (load and driver) information
To apply global access to the design, the following command can be used for read, write, and connectivity access:
% ncelab -access +rwc options top _ level _ module
Try to specify only the type(s) of access required for your debugging purposes. In general:
• Read (r) access is required for waveform dumping and for code coverage.
• Write (w) access is required for modifying values through PLI or Tcl.
• Connectivity (c) access is required for querying drivers and loads in C or Tcl, and for features like tracing signals
in the SimVision Trace Signals sidebar. In regression mode, connectivity access must be always turned off.
For example, if the only reason you need access is to save waveform data, use ncelab -access +r.
To maximize performance, consider using an access file with the ncelab -afile option instead of speci-
fying global access using the -access option. Any performance improvement is proportional to the amount of
access that is reduced. The maximum improvement from specifying -access rwc to using no access can be >2X.
Even removing connectivity (c) access can result in a 20-30% improvement. The typical gain of fine-tuning access in
large environments is 15-40%
You can also write an afile to increase performance at the gate level.
Sample afile:
DEFAULT +rw
CELLINST –rwc
www.cadence.com 15

Note: A new switch –nocellaccess is also available in release 12.1 and later to turn off cell access.
Multiple Snapshots for Access
You can also maintain two snapshots―one with more or global access for debug mode, and the other with limited
or no access, for regressions. For performance, you can always use the snapshot with limited or no access and, in
the case of debugging, use the global access snapshot.
5. Other Useful Incisive Enterprise Simulator Gate-Level Simulation Features
5.1 Initialization of Variables
The -NCInitialize option provides you with a convenient way to initialize Verilog variables in the design when
you invoke the simulator, instead of writing code in an initial block, using Tcl deposit commands at time zero,
or writing a VPI application to do the initialization. This option works for Verilog designs only, and you can enable
the initialization of all Verilog variables to a specified value. To initialize Verilog variables, you use -ncinitialize
with both the ncelab and the ncsim commands as follows:
• Use ncelab -ncinitialize to enable the initialization of Verilog variables.
• Use ncsim –ncinitialize to set the value to which the variables are to be initialized when you invoke the
simulator.
For example:
% ncvlog test.v
% ncelab -ncinitialize worklib.top // Enable initialization
% ncsim -ncinitialize 0 worklib top // Initialize all variables to 0
When you invoke the simulator, all variables can be initialized to 0, 1, x, or z. For example, the following line
initializes all variables to 0:
ncsim -ncinitialize 0
Different variables can be initialized to 0, 1, x, or z randomly using rand:n, where n is a 32-bit integer used as a
randomization seed. For example:
ncsim -ncinitialize rand:56
Different variables can be initialized to 0 or 1 randomly using rand _ 2state:n. For example:
ncsim -ncinitialize rand _ 2state:56
See ncsim -ncinitialize in Cadence Help for details.
Note: You can also enable initialization by specifying global write access to all simulation objects with the -access
+w option. However, this option provides both read and write access to all simulation objects in the design, which
can negatively affect performance. The -ncinitialize option provides read and write access only to Verilog
variables.
6. Improving Elaboration Performance Using Multi-Snapshot Incremental Elaboration (MSIE)
Incremental elaboration technology can be used in any case where turn-around time (time for newer versions to
simulate) is an overriding concern. In an SoC environment, there can be multiple IP used and newer revisions might
come in at different intervals in the verification process. MSIE technology can help the designer save build time and
start gate-level simulations sooner.
www.cadence.com 16

IP1
Comp/Elab Simulate
Analyze / Debug / Fix
Comp/Elab Simulate
IP2
Comp/Elab Simulate
IPn
Comp/Elab Simulate
SoC Timeline
…
IP1
IP2
IPn
SoC
MSIE can be applied as shown in the figure below, where there are multiple IP cores in an SoC top module. The
different primary snapshots can be created for each IP and top, and the testbench environment can be in the incre-
mental portion. Even in the case where a complete elaboration is required at once, all the primaries can be built
in parallel and, finally, the incremental portion can be built, including all primaries. This saves a lot of elaboration
time.
Top Level
SoC Top Verification Env.
Test-Specific
VE
Shared
VE
All primaries can be
built in parallel
irun -ip1.v
-mkprimsnap
-snapshot ip1
irun -ip2.v
-mkprimsnap
-snapshot ip2
irun top.v -primname
-ip2 -primname ip2
-primname ip3
irun -ip3.v
-mkprimsnap
-snapshot ip3
Incremental
Snampshot
IP1 (Inst 1)
IP1 Hierarchy
IP2 (Inst 2)
IP2 Hierarchy
IP3 (Inst 3)
IP3 Hierarchy
Primary
Snapshot
Primary
Snapshot
Primary
Snapshot
MSIE Use Model
1. Apply to GLS without timing (without SDF)—The technology can be directly applied and design partitioning can
be done as shown in the figure above.
2. Apply to GLS with timing (SDF)—Requires a separate SDF for each partition and IP.
MSIE can also be applied for DFT-style tests (serial or parallel scan). If the tests are set up so that each test has
a separate top-level HDL file containing the patterns (which is somewhat common), then building the DUT as a
primary snapshot with the SDF file, and then making the incremental snapshot the test+DUT, works very well.
www.cadence.com 17

7. Accelerating Gate Level Simulations
The Cadence Palladium®
XP platform accepts design descriptions in both synthesizable RTL and gate-level repre-
sentations or mixed abstractions equally well. Gate-level designs are compiled using the UXE compiler targeting all
verification modes, including simulation acceleration and in-circuit emulation with minimal to no compromise on
capacity or speed. Generally speaking, as design sizes increase (or the number of objects to simulate increase), the
performance of a simulator subsides whilst the performance of the Palladium XP platform remains fairly constant
thereby achieving relatively higher verification speeds. For simulation acceleration, speed-ups of 100x or higher are
typical, whereas for in-circuit emulation, speed-ups of 10,000x or higher are achievable, reaching MHz speeds. For
large designs where the simulation speeds for gate-level designs are in the range of 1 to 10Hz, simulation accel-
eration can be as high as 10,000x, and in-circuit emulation could be as high as 100,000x over the GLS.
With the Palladium XP platform, gate level designs can be used for both pre-silicon verification and post-silicon
validation. Users can validate gate-level netlists synthesized by standard synthesis tools targeting silicon vendor
libraries, including DFT and scan overlay structures. Gate level netlists can also be used to identify and analyze peak
and average power at sub-system and system level.
A Methodology for Improving Gate-Level Simulation
This section shows some methodologies that can help you reduce overall GLS verification time and make the
process more effective.
1. Effectively Use Static Tools Before Starting Gate-Level Simulation
Using static tools like linting and static timing analysis (STA) tools can effectively reduce the gate-level verification
time.
1.1 Linting Tool
It is recommended that you use static linting tools such as the Cadence HDL Analysis and Lint (HAL) tool before
starting zero-delay simulations. This can help you identify the issues or potential areas that can lead to unnecessary
issues in gate-level simulation.
Since simulation runs at gate level can take a lot of time, and based on the issues reported by the linting tools,
updates can be done in the gate-level environment up front, rather than waiting for long gate-level simulations
to complete and then fixing the issues. Some of the potential issues that can be identified by linting tools are as
follows:
• Detecting zero-delay loops
• Identifying possible design areas that can lead to race conditions
If race issues are detected, you can use the techniques listed in section 1.3 Handling Zero-Delay Race Conditions to
fix them.
1.2 Static Timing Analysis Tool
Static timing analysis tool information and reports can be used to start gate-level simulations with timing early in
the design cycle. The information from STA reports can help you run meaningful gate-level simulations with timing,
by focusing on the design areas that have met the timing requirements. Also, you should temporarily fix timing for
other portions of the design that require changes or updates for timing closure.
Some techniques that can be used for this are presented in this section.
1.2.1 Reducing Gate-Level Timing Simulation Errors and Debug Effort Based on STA Tool Reports, Even If Timing
Closure Is Incomplete for Some Parts of the Design
Since GLS runs much slower than simulations without timing, starting GLS timing verification early can be helpful.
However, this can lead to a lot of unnecessary effort in debugging the issues that are already reported by STA tools
like Cadence Encounter Timing System, as the design is still in the process of timing closure. As the timing issues
that have been reported by STA tools require fixes to the design, using the SDF from this phase directly in the
simulation will show failures in the simulation as well.
www.cadence.com 18

The designer can temporarily fix these violations in SDF through STA tools and start GLS in parallel. And at the
same time, timing issues in the design can be fixed by the timing closure team. This technique can help you:
• Reduce GLS errors based on STA (Encounter Timing System reports)
• Improve the design cycle/debug time during GLS:
–
– You can focus on gate-level issues rather than known STA (Encounter Timing System) errors
–
– Debugging and fixing issues in parallel during design phase:
»
» STA (Encounter Timing System) timing violations
»
» GLS (other than Encounter Timing System errors)
–
– Running functionally correct GLS by ignoring Encounter Timing System errors
The following section gives an overview of the typical STA GLS flow and also describes techniques to temporarily
fix setup and hold violations.
Static Timing Analysis and Gate-Level Simulation Flow
The following figure illustrates an overview of a typical STA tool, such as Cadence Encounter Timing System and a
gate-level simulator, such as Incisive Enterprise Simulator.
Gate-Level Netlist
STA Tool
GLS Timing
SIM
SDF
Generates
Violation Report
Designers Work
on Real
Design Fix
Generates
SDF ERRORS
SIM PASS
Design Tape Out
Fix the design
issue and validate
it in STA again
Simulation can start
only once complete
STA is clean
STA Tool generates
timing report and SDF
Check for
errors in
report
SDF with no
timing issue
No No
Yes
Yes
STA FLOW SIMULATOR FLOW
Figure 5: Typical Static Timing Analysis Tool and Gate-Level Simulator
Handling STA Setup Violations
This technique can help designers temporarily fix setup timing violations reported by STA tools so that they do not
appear in the GLS. Consider the following simple example for a setup violation:
www.cadence.com 19

FF1 Failing FF2
Data In
D2
Data Out
Clock
Comb. Logic
G2
C
B
FF1Data Out
G1
G3
G4
G6
InputtoFF2.D2
G5
Comb Logic Example
Setup Violation Example
End Point Slack (ns) Cause
Top/core/ctrl/FF2/D2 -0.56 Violated
Figure 6: Setup Violation
In the example illustrated above, a setup violation is reported at FF2 and the slack is -0.56. The sample combi-
national logic is also shown in the figure. If the SDF is generated by an STA tool (in this case, Encounter Timing
System) directly, then it will not only report timing violations during GLS, but there will also potentially be a
functionality mismatch, as the data for D2 might not get latched at FF2 at the next clock cycle. This might impact
the functionality of other parts of the design as well.
Since this is a known issue and requires a fix in the combinational logic from FF1 to FF2, the GLS can temporarily
ignore this path, or the timing needs to be temporarily fixed. This can be done in the STA tool itself by setting a
few gate delays to zero and compensating for the slack. Once the slack is compensated for, then the SDF can be
generated and used by the GLS.
To help you understand this approach in detail, the following table shows the delays for each gate.
Gate Delay (ns) Cumulative Delay (ns)
G1 0.32 0.32
G2 0.25 0.57
G3 0.4 0.97
G4 0.35 1.32
G5 0.28 1.60
G6 0.33 1.99
Referring to Figure 6, within Encounter Timing System, starting from node FF2.D2, a traversal is done and the slack
at each node is checked. Looking at the table above, it can be seen that setting delays for G1 and G2 to zero will
compensate the slack of -0.56.
Similarly, all the setup violations can be temporarily fixed. STA can be run again on the modified timing to check if
there are any new violations added by assigning zero delays to the gates.
This approach can be helpful only when the violations reported by STA tools are limited in number and are present
only in some small part of the design. Using this approach will not work in cases where there are a lot of viola-
tions, and where they affect almost the complete design, as it would change the timing information in the SDF file
completely and make the simulation very optimistic.
www.cadence.com 20

Adding Negative Delays in SDF (Limited Cases)
The effects of the above case can also be mitigated by adding negative delays in the STA environment at gate G1,
if the negative delay value in the above example is less than or equal to the G1 delay (which is 0.32ns). This feature
of adjusting negative delays by the simulator is available only in the Incisive 13.2 release and onwards. This new
feature in Incisive Enterprise Simulator can only adjust delays for one level. If the delays cannot be adjusted for
more than level 1, the remaining negative values are set to zero. By default, the elaborator zeros out all negative
delays and issues a warning. To enable negative interconnect and module path delay adjustments, run the elabo-
rator with the -negdelay option (ncelab -negdelay or irun -negdelay). For example:
% ncvlog -messages test.v
% ncelab -messages tb -negdelay
or
% irun test.v -negdelay
Use the -neg _ verbose option (ncelab -neg _ verbose or irun -neg _ verbose) to print out the negative
delay adjustments and save them to the default log file, as shown:
% irun test.v -negdelay -neg _ verbose
...
file: test.v
...
Elaborating the design hierarchy:
Top level design units:
tb
...
Building instance-specific data structures.
Processing Negative Interconnect Delays...
Adjusting driver of net tb.out1...
Driver delay adjusted successfully...
Adjusting driver of net tb.out3...
Driver delay adjusted successfully...
Number of negative interconnect delays adjusted successfully: 2
2 Negative interconnect delays have been adjusted successfully.
...
Writing initial simulation snapshot: worklib.tb:v
Adjustment Rules for Negative Delays
Incisive Enterprise Simulator now supports negative delays on single output gates without the need to zero their
values. When enabling negative delay adjustments, the adjustments are applied in the following sequence:
1. Other interconnect delays at the port
2. Driver delays at the input
3. Load delays at the output
Note: When comparing the negative delay values to the total delays in a design, adjustments to negative delay
values will introduce inconsistencies with the delays specified in the SDF file and may affect elaborator perfor-
mance.
Please Note: For this approach Cadence is evaluating different designs to test corner scenarios and would like to
partner with design teams to review and deploy it. Contact your field AE if you are interested.
Handling STA Hold Violations
Similar to set-up violations, for hold violations consider the example below:
www.cadence.com 21

FF1 Failing FF2
Data In
D2
Hold
Violations
Data Out
Clock
Comb. Logic
End Point Slack (ns) Cause
Top/core/ctrl/FF2/D2 -0.50 Violated
Figure 7: Hold Violation
A hold violation is reported at FF2 and the slack is -0.50. In order to temporarily fix the issue, an extra interconnect
delay of 0.5ns can be added at D2.
The following are the advantages, assumptions, and limitations for this flow.
Advantages
• Encounter Timing System violations will not appear in the GLS
• Improvements in design cycle time, as the designer can focus on GLS issues, other than issues reported by
Encounter Timing System
• Adding 0 delays for gates can improve GLS performance
Assumptions
• Most of the design is STA clean and has limited known violations
• Adding 0 delays will not add other STA timing errors
• Needs to be checked by re-running STA
Limitations
• New errors added by the above updates might not get corrected.
Generating and Using Smart SDF with Timing Abstractions Running Timing Simulations
STA tools have a capability for doing hierarchical timing analysis and generating SDF hierarchically. This feature is
included in STA tools because flat STA for a complete chip takes a lot of time.
Typically, there are IP cores in, or in portions of, the design that are re-used across different SoCs and are already
silicon-proven. Turning on complete timing for an SoC enables timing of internal cells of all the blocks and IP cores,
leading to a lot of redundant simulation overhead. Currently, the designer can turn off the timing of the complete
IP/module using the tool features, but this adds optimism at the SoC level since the delay at output ports of the IP
become zero. Also, the timing checks at the input pins of the IP are ignored.
IP that is timing-clean and integrated at the SoC level, if run with full timing enabled, requires timing for all flops
and combination logic for accurate timing results. Since the IP is already timing-clean, the only set of timing issues
that can come are at the SoC level, or at the integration of the IP as shown in the figure below.
www.cadence.com 22

Comb. Logic 1
IN1 OUT1
IN2
CLK
OUT2
IN3 OUT3
FF1 Comb.
Logic 2
FF1
Comb.
Logic 1
FF2 Comb.
Logic 2
FF4
Comb.
Logic 4
FF3
Comb.
Logic 3
FF2
FF5
Comb.
Logic 3
FF3 FF4
Comb.
Logic 4
Comb. Logic 6
SoC TOP
Potential timing issues at SoC integration
IP
TOPIN1 TOPOUT1
TOPIN2 TOPOUT2
TOPIN3
CLK
TOPOUT3
Comb. Logic 5
The following types of timing models can be used to generate SDF for gate-level simulation: Interface logic models
• Extracted timing models
Please Note: For this approach Cadence is evaluating different designs to test corner scenarios and would like to
partner with design teams to review and deploy it. Contact your field AE if you are interested.
2. Interface Logic Models
An interface logic timing model (ILM) has partial timing of the block that includes the timing at boundary logic
only, but hides most of the internal register-to-register logic.
Consider the following simple example.
Comb. Logic 1
IN1 OUT1
IN2
CLK
OUT2
IN3 OUT3
FF1 Comb.
Logic 2
FF2
FF5
Comb.
Logic 3
FF3 FF4
Comb.
Logic 4
Comb. Logic 6
Comb. Logic 5
Figure 8: Full Timing Model Example Design
Flop-to-flop internal paths are not required, as they do not affect the interface path timing and we are interested in
the interface path timing only. In the figure below, the timing of internal gates is set to zero (marked in green)
www.cadence.com 23

Comb. Logic 1
IN1 OUT1
IN2
CLK
OUT2
IN3 OUT3
FF1 Comb.
Logic 2
FF2
FF5
Comb.
Logic 3
FF3 FF4
Comb.
Logic 4
Comb. Logic 6
Comb. Logic 5
Remove timing for these gates
Figure 9: Interface Logic Model
2.1 Advantages of Using Interface Logic Models
The advantages of using interface logic models include:
• Good performance improvements, both in STA and GLS, since only a partial design is used
• High accuracy as interface logic and interface cells are preserved
Please Note: This is a patented flow (beta version) and is available in Encounter Timing System (ETS) to dump
abstracted interface timing models for running effective GLS. Cadence would like to partner with design teams to
review and deploy this flow. Contact your field AE if you are interested.
The figure below shows the results for an IP core with full timing versus interface logic timing. As the waveforms
show, both behave exactly the same. The performance gains are expected to be in the range of 1.5X – 2X,
depending upon the number of IP cores considered for timing abstraction or the overall level of abstraction.
Sub Block waveform with full timing
Sub Block waveform with abstracted models (Timing accurate)
Figure 10: Functionality and Timing of an IP Core with Full Timing and with Interface Logic Timing
www.cadence.com 24

2.2 Extracted Timing Models
An extracted timing model (ETM) creates a timing representation for the interface paths, which are the timing arcs
created for each input-to-flop/gate, input-to-output, and flop/clock-to-output paths.
If there are multiple clocks capturing data from an input port, then an arc, with respect to each input port, is
extracted.
IN1 OUT1
IN2
CLK
OUT2
IN3 OUT3
Path Delay Cominational
Path Delay Sequential
Path Delay Sequential
Setup/Hold (CLK & IN3)
Setup/Hold (CLK & IN2)
Extracted Timing Model
Figure 11: Timing Model Details
In the example, the timing model that is extracted is context-independent and does not contain timing details for
the logical gates and flops inside it. It just contains timing information for path delays from input to output, and for
setup and hold violations at inputs.
There is also no need to re-extract the model if some of the context gets changed at a later stage of development.
This is because none of the boundary conditions or constraints (input_transitions, output loads, input_delays, and
neither output_delays nor clock periods) are taken into consideration for extracting the model.
However, the model depends upon the operating conditions, wireload models, annotated delays/loads, and RC
data present on internal nets defined in the original design. If these elements change at a development stage of
design, then you need to re-extract the model for correlation with the changed scenario.
Advantages of Using Extracted Timing Models
The advantages of using extracted timing models include:
• Huge performance improvements both in STA and GLS because detailed timing is removed
• IP reuse and interchange of timing models
• IP protection in black boxing the design
3. Controlling or Handling Timing Checks Based on STA Reports
Since STA does the complete timing analysis, and in cases where timing for the complete or a portion of the design
is already met, the timing checks during simulation might not be required for this portion of the design, specifically
the internal flops.
As mentioned above in section 4, “Timing Checks” of 1.3.4 Incisive Enterprise Simulator Race Condition Correcting
Features, timing checks have a significant impact on performance, so they can be turned off either completely or
selectively, based on the requirement.
The STA and simulator flow is illustrated in the following figure.
www.cadence.com 25

STA Flow
Generates
Violation Report
Fix Violation
Errors
Generate Tfile and
Romove Timing
for these Gates
Identify Internal
Flops and Gate
ERRORS
Simulator Flow
Tfile
GLS Timing SIM
Gate-Level
Netlist
Figure 12: Timing File Example
Consider the example in Figure 8 above (titled Original Design). Based on STA, timing information for internal gates
and flops is not required. In the figure below, an Incisive Enterprise Simulator timing file can be generated without
timing information for gates marked in green.
Comb. Logic 1
IN1 OUT1
IN2
CLK
OUT2
IN3 OUT3
FF1 Comb.
Logic 2
FF2
FF5
Comb.
Logic 3
FF3 FF4
Comb.
Logic 4
Comb. Logic 6
Comb. Logic 5
Remove timing for these gates
Figure 13: A Timing File Can Be Generated Without Timing Information for Gates
4. Focusing on Limitations of Static Timing Analysis and Logical Equivalence Tools
STA, or even logical equivalence tools, are not able to catch all the issues that can only be seen in a GLS run. You
can define coverage tests around these limitations so that any functionality that is not handled by the static tools
can be tested.
Limitations of STA include:
1. Inability to identify asynchronous interfaces
www.cadence.com 26

2. Static timing constraints like false paths and multi-cycle paths
Based on constraints, GLS can define the coverage points. Simulation must focus on these areas as they constitute
the majority of issues in GLS.
5. Using DFT Verification
Gate-level DFT simulation is performed for the verification of test structures inserted by specialized DFT tools,
such as Encounter Test. As designs have exploded in size and complexity, great advancements have taken place in
automated test tools over the last decade in the area of scan chain insertion, compression logic to save I/Os and
speed up testing, removal of hotspots, Built-in Self Test (BIST) logic, and so on.
However, these techniques and technologies are not the focus of this document. This section summarizes the
motivations for DFT simulations, and offers you best practices that could optimize the running and debugging of
long DFT simulations.
Simulation with Netlists
The motivations behind simulating netlists with test structures include:
• Functional equivalency checks
• Meeting timing requirements
• Fault coverage in the design
• BIST
Let’s examine them one by one for better understanding.
Functional Equivalency Checks
Functional equivalency checks answer the very basic question: Did I change something in my design functionally
by inserting additional scan or BIST logic? The following figure shows a simplified view of a scan chain (scan-in to
scan-out) highlighted by the red path.
Gates
Gates
Gates
Gates
Primary
Inputs
Primary
Outputs
Scan-in (SI)
Scan-out (SO)
Scan Flip-Flop
Figure 14: Simplified Scan Chain Representation in a Design
D Flip-Flop Scan Flip-Flop
D
CK
Q D
CK
Q
0
1
SI
SE
Figure 15: Basic Logical Representation of a Scan Flip-Flop
www.cadence.com 27

Each scan flip-flop, as shown in the figure above, introduces additional controls such as scan enable, scan input,
and scan mux into the design.
Inserting these scan elements into a complex design with asynchronous I/Os, gated clocks, and latches is
automated by test tools, and may introduce functional errors. The primary concerns here are to verify that all nets
are connected with scan flip-flops, and that no unwanted functional effects are introduced.
Static equivalency checking is the primary method for catching functional errors introduced in the scan insertion
phase. However, designers find GLS to be an additional insurance, and so they run simulations to find functional
errors, as well as to verify logic inserted by the test tool. We call this effort Functional Integrity Simulation Testing
(FIST).
Meeting Timing Requirements
“Have I met the timing requirements?” This is often the most challenging question because meeting timing require-
ments on SoCs with multiple asynchronous I/O blocks, multi-cycle paths/false paths, and clock speeds pushing
into multi-GHz ranges makes timing closure a tremendous challenge for teams. And, including scan multiplexers
adds additional gates in the critical path for functional operation, which increases flip-flop fan-out as shown in the
figure above.
This step is done with a post-layout netlist, since routing and placement effects must be used to validate the
timing of the chip. Design teams use constraint-based static timing delay checkers to validate that timing closure is
met with test logic. However, multi-cycle paths and false paths are excluded from such checks. Therefore, timing
annotated GLS is required to verify timing on multi-cycle paths. Often, simulation-based timing checkers are used
as additional insurance against static timing tools to make sure human errors were not introduced in constraints
that were specified statically.
Fault Coverage in the Design
Test tools provide input vectors and expected outputs, which are targeted to be run on automated test equipment
to catch manufacturing defects. Some design teams run these vectors in simulation with toggle coverage enabled
to measure how many wires in the design toggle. Using these vectors can give an independent gauge of testability.
BIST
For high-speed memory tests, a BIST controller is placed in the design to generate pseudo-random patterns that
write to different memory banks. The memory is read back and compared against expected values to detect
defects in memory arrays. Typically, BIST controllers are verified at an IP level and basic tests are performed to make
sure a subset of all addressable memories can be accessed by the BIST controller.
Initialize DUT
Enable
Scan
Mode
Update PI
Enable
Scan
Mode
Apply
Functional
Mode
N*Clock
SI
Clock 1 or
2 cycles
Enable
Scan
Mode
Apply
Functional
Mode
Clock 1 or
2 cycles
Enable
Scan
Mode
Apply
Functional
Mode
Clock 1 or
2 cycles
N*Clock
Sample and
Compare SO
Clock in next
pattern
N*Clock
Sample and
Compare SO
N*Clock
Sample and
Compare SO
Clock in next
pattern
Pattern 1
Pattern 2
More Patterns...
Figure 16: Typical ATPG Simulation Flow Using Serial Mode
Tests that run in seconds of real time at GHz speeds can take several hours to several days in simulation, depending
upon a few important factors. The figure above shows a typical ATPG test flow.
www.cadence.com 28

The red boxes highlighted in the graphic indicate qualitatively where the majority of simulation time is spent. These
time-consuming scenarios include:
a. Scan patterns clocked in serially. These patterns take “n” clock cycles to scan in and “n” cycles to scan out.
Even with optimized scan chains, this could be thousands of clock cycles per pattern. Parallel load and unload
techniques can be used to drastically reduce the simulation time in these cases.
b. SDF annotated timing simulations with advanced node libraries. These could take 4X-5X more wall-clock
time and compute farm memory resources to simulate. Several verification concerns could be handled with a
pre-layout unit delay netlist.
c. Design initialization time versus pattern simulation time. The simulation flow in Figure 15 shows a typical case.
Let’s look at each of the above scenarios to examine trade-offs for different methods for the task at hand.
Functional Integrity Simulation Testing (FIST)
FIST should be done in pre-layout mode, without timing, to catch any errors introduced in the scan-chain insertion
step. Two kinds of tests are required for this verification. First, a few functional tests in zero-delay mode should be
simulated to build confidence in the test insertion. Second, a single ATPG pattern should be run in serial model that
exercises all scan chains to verify scan chain integrity.
Depending upon the amount of compute resources available for the job at the time, a few additional, top-ranking
patterns for simulation should be selected, based on coverage grading produced by the test tool. A hardware
emulation solution, such as the Cadence Palladium®
XP platform, can be used very effectively to verify the
functional integrity between RTL and pre-layout netlists.
Timing Verification
Typically, this activity takes the bulk of the DFT simulation effort since simulations with timing annotated on large
netlists are very slow, usually in the range of a few cycles per second to a few seconds per cycle, on the latest
workstations. Simulating a single pattern could take several hours in serial mode. If 10-12 patterns are chained
together in a single simulation run, a simulation run could take days.
A few techniques can be applied to improve simulation runtime and, therefore, the debug time associated with
these simulations:
a. Use a compute-farm to break down long serial runs into several shorter simulation runs.
This involves trading-off compute efficiency versus turn-around time for validating late design changes and debug
time. A calculation can be made to make the right trade-off. To do so, measure the DUT initialization time and call
it TinitDUT . Now measure simulation times associated with a few patterns and take an average to compute average
time per pattern, and call this Tpattern_av.
Note that pattern simulation time will vary due to different event densities produced by different patterns.
Calculate the amount of time it takes to start a simulation on the farm and call it TSimStart. Total simulation time for
running “m” patterns serially should be:
Total simulation time on a single machine = TSimStart + TinitDUT + m x Tpattern_av
By segmenting patterns into multiple parallel simulation runs, designers can reduce turn-around time significantly.
We can apply constraint solving to identify the optimal solution for the minimum number of machines required to
achieve a target regression time, as shown in the table below (all numbers are in minutes).
TsimStart TDutInit Tpattern_av Number of
patterns
Desired
regression
time
1 sim
run-time
Number of
machines
Average
single test
time
Patterns
per
machine
30 60 50 80 500 4090 10 490 8
30 60 100 80 500 8090 20 490 4
30 60 150 80 500 12090 40 390 2
30 60 200 80 500 16090 40 490 2
30 60 250 80 500 20090 80 340 1
www.cadence.com 29

The table shows that partitioning long single-pattern simulations into multiple shorter simulations can achieve
faster regression turn-around time. This step must be planned up-front so this step does not become the critical
constraint right before tape-out.
b Use the parallel load and unload technique
Modern test tools, such as Encounter Test, provide an option to create testbenches and test patterns that use a
simulator API, such as the Verilog Programming Interface (VPI) to back-door load and unload scan flip-flops. This
technique eliminates the time-consuming activity of consuming processor cycles to shift data in and out, and to
clock data in and out.
This technique should be used in combination with a few patterns tested in serial mode to ensure there are no
timing issues in shifting data in and out of scan flops. Once that is done, the majority of the timing verification task
can be achieved through the parallel load and unload technique. This technique checks various timing scenarios
that could be exposed by different patterns, which can cause transitions to propagate by way of different logic
paths in the circuit.
Initialize DUT
Enable
Scan
Mode
Update PI
Enable
Scan
Mode
Apply
Functional
Mode
Parallel
Load all
FFs in 0
Sim Time
Parallel
Load all
FFs in 0
Sim Time
Clock 1/2
cycles
Enable
Scan
Mode
Update PI
Enable
Scan
Mode
Apply
Functional
Mode
Parallel
Load all
FFs in 0
Sim Time
Parallel
Load all
FFs in 0
Sim Time
Clock 1/2
cycles
Enable
Scan
Mode
Update PI
Enable
Scan
Mode
Apply
Functional
Mode
Parallel
Load all
FFs in 0
Sim Time
Parallel
Load all
FFs in 0
Sim Time
Clock 1/2
cycles
Pattern 1
Pattern 2
More Patterns...
Pattern n
Figure 17: Typical ATPG Simulation Flow Using Parallel Mode
Fault Coverage
New DFT tools use sophisticated algorithms to insert test structures and report accurate fault coverage results.
There is no need to use simulation tools to cross-check those DFT tools. If a cross-check must be done, simulations
can be run without timing, and in parallel mode, as a confidence-building verification.
BIST Testing
The BIST technique is used to test memory faults in the device without using expensive ATE time. A stimulus
generator and a response collector are placed in the design and hooked up to memory structures with a multi-
plexer to select between BIST-generated stimulus and system stimulus, as illustrated in the following figure.
Circuit Under Test
MUX
Pattern
Generator
Response
Analyzer
BIST
Controller
from System to System
good/fail
biston bistdone
Figure 18: Typical BIST Architecture for Memory Verification
www.cadence.com 30

Functionally, the BIST controller and the response analyzer are exhaustively tested at the IP level. Therefore, the
objective at the gate-level is to ensure connectivity and timing. A subset of BIST tests should be simulated with a
post-layout netlist and with timing turned on selectively for components involved in the BIST path. Then, black-box
test the rest of the DUT.
6. Catching Gate-Level Simulation X Mismatches at RTL Using Incisive Enterprise Simulator
X-Propagation Solution
Standard Verilog semantics define X as an unknown value. This value appears when simulations cannot acutely
resolve signals to 1 or 0. Unfortunately, RTL simulations often mask X values by converting unknown values into
known values, while GLS show these hidden Xs. These X issues can be costly, adding complications as you famil-
iarize yourself with abstract synthesized logic, while consuming time as you exhaustively simulate and synthesize
designs for verification purposes.
The Incisive Enterprise Simulator x-propagation feature enables the same behavior as that of hardware at the RTL
level to identify the ‘X’ issues much earlier in the design phase, saving costly GLS and debugging effort.
X-Optimism Problem at RTL
To understand the ‘X’ optimism issue, let’s consider a simple “if” block in the figure below. In the case of RTL, if
the select signal is ‘X’ then, according to LRM schematics, else block of the “if” statement is executed. In case of
real hardware depending upon the values of input signals ‘A’ and ‘B’, the output will be computed. In this case,
if both ‘A’ and ‘B’ have the same value, then the output will be the same. But if the values of input signals are
different, then the output will be non-deterministic.
The truth table below shows the difference between the two.
This can also be seen in for case statement and if the select goes to ‘X’, there is a difference between the behaviors
of RTL vs hardware.
RTL Hardware
A
B
sel
OUT
Figure 19: X-optimism
www.cadence.com 31

Similarly, the ‘X’ optimism problem can be seen in a sequential-edge triggered block where the edge goes from ‘0’
or ‘1’ to ‘X’ or ‘X’ to 1 or ‘0’. Support is available in Incisive Enterprise Simulator to handle such edge-triggered ‘X’
issues at the RTL level.
To learn how to enable the x-propagation function, please refer to the user’s guide.
7. Black-Boxing Modules Based on the Test Activity
Black-boxing some of the modules can be a good approach to improving GLS performance, as typically each
test verifies only a certain portion of the design. Also, the IP cores that are already verified can be black-boxed
completely and only their interface-level details are required for verifying the other portions of the design.
Black-boxing can have a positive performance impact on:
1. Reducing elaboration and simulation memory footprint because the details of the module are missing
2. Reducing elaboration time
8. Saving and Restarting Simulations
Typically, much GLS time is spent in the initialization phase. This time can sometimes be very significant. As a
result, a single simulation should be saved and all other (n-1) simulations should be run from the saved checkpoint
snapshot. The figure below describes this flow.
Test A
Test B
Test C
Test D
Test A
Initial Setup Run
Test B
Test C
Test D
Save
State
Restore
State
Significant reduction
in simulation time
Original Regression Start of interesting traffic
Re-Start/Re-Use Initial Setup
Figure 20: Saving the State After Initialization Saves Simulation Time
Library Modeling Recommendation
Complex timing checks in 40nm technology nodes and below are responsible for longer run times and high
memory requirements. Writing library models efficiently can help improve the performance of GLS with and
without timing.
Let’s consider a simple example of a finer technology node, where the timing checks in specific blocks not only
have conditional logic (marked in red), but also the conditions require extra logic (marked in blue). Similarly, the I/O
path delays have conditional paths (marked in green).
module FF1 ( Q, TE, CPN, E );
input TE, CPN, E;
output Q;
………..
……….
www.cadence.com 32

not (nTE, TE);
and (E _ TE, E, TE);
mybuf (E _ TE _ SDFCHK, E _ TE, 1’b1);
mybuf (nTE _ SDFCHK, nTE, 1’b1);
…………
…..…….
specify
if (E == 1’b0 && TE == 1’b0)
(posedge CPN => (Q+:1’b1)) = (delay1, delay2);
if (E == 1’b1 && TE == 1’b1)
(CPN => Q) = (delay3, delay4);
$width (posedge CPN &&& E _ TE _ SDFCHK, 0, 0, notifier);
$setuphold (negedge CPN &&& nTE _ SDFCHK, posedge E , 0, 0, notifier);
endspecify
Due to these additional conditional statements and additional logic, the events on the simulator increase signifi-
cantly, as shown in Figure 21 below.
Actual
Expected
Log
(events)
90 65 40 28 20 14
Figure 21: How Additional Conditional Statements and Logic Impact Simulation Events
In order to reduce the number of events in the simulation, consider the following while writing a library:
1. Since conditional timing checks require extra logic and are more complex, check the following:
a. Whether the conditions can be reduced by removing unnecessary, insignificant conditions
b. Whether the conditional logic can re-used or merged for multiple cases
2. Use of # delays in libraries should be avoided if the simulation is running with timing mode as using the absolute
# delays turns off simulator optimizations. These delays can be used only for functional mode, i.e. zero-delay
GLS. The example below shows the library models having functional mode for zero-delay simulations and pure
timing mode.
module FF (clk , d, q ) ;
input clk;
input d;
output q;
……………..
`ifdef FUNCTIONAL
//For pure zero delay simulations, models have #1 delays in sequential cells to avoid race
conditions.
not #1 NOT1( q, q1) ;
`else
www.cadence.com 33

//For timing simulations, models do not have # delays in sequential cells.
not NOT1( q, q1) ;
`endif
1. Use the minimum number of gates to represent the functionality. A smaller number of gates results in a smaller
number of events during the simulation.For example, use a three-input and gate instead of two two-input and
gates.
// Use of 3 input and gate
and and1(q, a,b, c);
// Instead of two 2 input and gates
and and1(out1, a, b);
and and2(q, out1, c);
2. Logic on the CLK signal should be minimized; also, bear in mind that a heavier load on the CLK signal slows
simulation.
Summary
We have shown that in order to improve GLS performance, both the simulator runtime performance as well as the
simulation methodology need to be improved.
The runtime performance of Incisive Enterprise Simulator has been continuously improved for
GLS. In addition, some recent enhancements in the simulator can now deliver from 1.2X to 8X performance
improvements for a particular design.
There are also continuous improvement efforts going on in the simulator development area to improve
GLS performance even further. However, in order to match the verification requirements for newer, larger designs
with increased complexities, a combined simulation and methodology approach has to be taken in order to achieve
an effective and efficient verification process.
Contacts
Gagandeep Singh gagans@cadence.com
Amit Dua adua@cadence.com
Cadence Design Systems enables global electronic design innovation and plays an essential role in the
creation of today’s electronics. Customers use Cadence software, hardware, IP, and expertise to design
and verify today’s mobile, cloud, and connectivity applications. www.cadence.com
©2014 Cadence Design Systems, Inc. All rights reserved. Cadence, the Cadence logo, Encounter, Incisive, and Palladium are registered trademarks of
Cadence Design Systems, Inc., All rights reserved. 2088 03/14 CY/DM/PDF

Gate-Level Simulation Methodology Improving Gate-Level Simulation Performance

More Related Content

What's hot (20)

Similar to Gate-Level Simulation Methodology Improving Gate-Level Simulation Performance (20)

Recently uploaded (20)

Gate-Level Simulation Methodology Improving Gate-Level Simulation Performance