Signoff ECO: Fixing the Last Violations Without Breaking Signoff

Chapter overview

Late in a project, the design is routed and almost clean. Signoff STA still shows a few hundred violations. Some are hold violations on short paths. Some are slew violations on long nets. Some paths pass in one corner and fail in another.

Fixing these by hand in the place-and-route (P&R) tool is slow. Each fix there is checked with a different timer. Thus, a fix can look good in P&R and still fail at signoff. A signoff ECO solves this. ECO stands for engineering change order – a small, controlled change to a nearly finished netlist.

In a signoff ECO, Tempus decides the fixes using signoff-accurate timing in every view at once. It writes the changes as ECO files. The P&R tool then applies them with minimal disturbance. This chapter teaches that flow, from the basic run to the large-design variants.

The chapter is long because the ECO stage has many tools. However, the core idea never changes. One timer decides, the other tool applies, and signoff checks the result. Every later section is a way to do the same job faster, on a bigger design, or with tighter limits.

Learning objectives

1. Describe the signoff ECO loop, its inputs, its outputs and the netlist changes it may make.
2. Build, save, load and merge an ECO timing database, or let the tool build it in batch.
3. Run eco_opt_design for DRV, setup, hold, area, leakage and power, including incremental steps.
4. Turn on glitch, SI slew, crosstalk, IR drop and cell EM fixing, with their prerequisites.
5. Use path-based retiming to recover power and then recover setup.
6. Set up a Paradime run and explain what the master and the clients each do.
7. Run the ECO from inside Innovus with signoffTimeDesign and signoffOptDesign.
8. Choose between hierarchical ECO, HC-ECO and Top plus boundary model ECO for a large design.
9. Restrict fixing to selected endpoints, buffer along routes, and skew clocks for setup.
10. Prepare and run a post-mask metal ECO with PFC or GA fillers.

Chapter roadmap

The chapter first builds the core flow and then scales it. Figure 1 shows the plan. Come back to it whenever you lose your place.

Signoff ECO on one page

Prepare, fix, recover and hand off; then the same flow for large designs

Key commands

Table 1 groups the main commands and settings by section. Each one is explained where it is first needed.

1  What a signoff ECO does

1.1  The concept

A signoff ECO is a short repair job on a finished layout. Its job is to fix the violations left after signoff STA, or to recover power where there is spare slack. It does not rebuild the design. It makes small changes and writes them down for the P&R tool.

A useful picture is a car after its final inspection. The inspector finds a worn tire and a loose mirror. Nobody rebuilds the car. A mechanic changes exactly those two parts, and the inspector checks the car again. In a signoff ECO, Tempus plays both the inspector and the planner. P&R is the mechanic.

The feature needs its own license and is started with tempus -⁠eco. The licensing rules are in Getting Started with Tempus.

1.2  Why it matters

Let us consider a block with 300 hold violations in the fast corner. P&R hold fixing used a simpler timer and thought the block was clean. Thus, fixing again in P&R repeats the same blind spot.

Tempus sees exactly the timing that signoff sees. Its fixes are judged by the same delay calculation, the same SI analysis and the same derates. Thus, a fix chosen by Tempus is much more likely to survive the next signoff run. The project closes in fewer loops, and each loop saved often means a day or more of schedule.

1.3  How it works

Figure 2 shows the ECO loop. Each pass through it is one ECO iteration.

The signoff ECO loop

Signoff, ECO, P&R ECO, extraction and signoff again

1. Signoff STA finds the remaining violations.
2. Tempus ECO chooses fixes and writes ECO files.
3. P&R ECO applies the files and runs an ECO route.
4. Re-extraction produces new parasitics for the changed nets.
5. Re-sign-off times every view again.

The inputs are the same as for signoff STA: libraries, the Verilog netlist, the MMMC view setup and SPEF. Optional inputs add detail – DEF and LEF for physical data, CPF or UPF for power intent, and ILM models. Physical data matters because a new buffer needs a legal location. Power intent matters because a cell must sit in the right supply domain.

The output is an ECO file that P&R applies incrementally. Typically, nothing in the existing layout moves unless a fix needs it.

Figure 3 gathers the inputs and outputs in one place. The required inputs on the left are the same data that signoff STA already reads.

What goes into and out of a signoff ECO

Signoff data in, two matching change files out

1.4  The changes an ECO may make

The tool has four basic transforms, shown in Figure 4. These are the netlist changes a normal timing ECO makes.

• Resize – replace a cell with a stronger or weaker version of the same function. A stronger cell drives its load faster, which helps setup and slew. A weaker cell saves area and power.
• Swap – replace a cell with a variant of the same size, for example another threshold voltage (Vt). A low-Vt cell is faster but leaks more. A high-Vt cell is slower but leaks less.
• Insert – add a buffer or an inverter pair on a net. This splits a long net, isolates a heavy load, or adds delay for hold.
• Delete – remove a buffer or an inverter pair that is no longer needed. This recovers area and power.

The four ECO transforms

Resize, swap, insert and delete

The tool respects dont_use cells and dont_touch objects. It also handles multi-supply-voltage (MSV) designs, hierarchical designs and master/clone modules.

2  The ECO timing database

2.1  The concept

Running full STA inside every ECO step would be slow. Thus, Tempus first writes an ECO timing database (ECO DB). It is a lightweight graph of the setup and hold timing. The ECO step reads this graph instead of re-timing the whole design.

Think of the DB as a map with the traffic already measured. The ECO plans its route changes on the map. It does not drive every road again before each decision.

There is one DB per active view. Each view writes two files: <top>_<view>.tgd and <top>_<view>.ckd. For example, a design top with views func_ss and func_ff writes four files. Figure 5 shows both ways to get a DB.

Two ways to get an ECO timing DB

Save it from signoff STA, or let the ECO build it in batch

2.2  Why it matters

The DB is the ECO's whole view of timing. If the DB was written with wrong settings, every fix is decided on wrong numbers. Thus, the DB must be written from a run that matches signoff – same SPEF, same derates, same SI settings. A DB from a careless run gives a careless ECO.

2.3  Saving and loading a DB

The DB folder is named with set_eco_opt_mode, and write_eco_opt_db writes it.

tempus> set_eco_opt_mode -save_eco_opt_db mySignOffTGDir   ;# name the DB folder
tempus> write_eco_opt_db                                   ;# one .tgd/.ckd pair per view

• -⁠save_eco_opt_db <dir> – names the folder that write_eco_opt_db fills. Without it, there is no saved DB to reuse later.
• write_eco_opt_db – writes the files for every active view. Run it after timing is fully up to date.

Later, in a new session, the ECO points to the same folder. In a concurrent MMMC session with fewer than about eight views, the load can happen in the same session.

tempus> set_eco_opt_mode -load_eco_opt_db mySignOffTGDir   ;# reuse the saved DB
tempus> eco_opt_design -hold                               ;# fix hold violations

• -⁠load_eco_opt_db <dir> – tells eco_opt_design to read this DB instead of building one. Set it before the ECO command runs.

If views were timed in separate sessions, copy all their files into one folder. The tool then sees one merged DB. For example, a func_ss run and a func_ff run can each write a DB, and one copy step joins them. The clients that build a DB can run locally or on LSF.

2.4  Letting the ECO build the DB

When -⁠load_eco_opt_db is not given, eco_opt_design builds the DB itself. It does this in batch with distributed MMMC (DMMMC) clients, one per view. The design must be fully loaded, with SPEF for every active RC corner.

read_lib $liberty                  ;# timing libraries
read_lib -lef $lef                 ;# physical libraries
read_verilog $netlist              ;# netlist
set_top_module my_top              ;# link
source viewDefinition.tcl          ;# MMMC views
read_def $def_file                 ;# placement and routing
source spef.tcl                    ;# SPEF for every active rc_corner
set_distribute_host -local         ;# clients on this machine
set_multi_cpu_usage -localCpu 8 -remoteHost 2 -cpuPerRemoteHost 4   ;# CPUs
eco_opt_design -hold               ;# builds the DB, then fixes hold

• set_distribute_host -⁠local – starts the DB-building clients on the local machine. LSF can be used instead.
• set_multi_cpu_usage – sets the CPUs for the master and the clients, as in Multi-Mode Multi-Corner Analysis.

3  Running the basic optimization

3.1  The command

The main ECO command is eco_opt_design. Each switch names one kind of fix or recovery:

tempus> eco_opt_design
    [-area]       ;# recover area
    [-drv]        ;# fix design rule violations (max transition, max capacitance)
    [-dynamic]    ;# dynamic power (confirm details with help eco_opt_design)
    [-hold]       ;# fix hold violations
    [-leakage]    ;# recover leakage power
    [-power]      ;# recover total power
    [-setup]      ;# fix setup violations

• -⁠drv – fixes design rule violations (DRV), such as slow transitions and overloaded nets. It is usually run first, because DRV fixes change the slews that setup and hold depend on.
• -⁠setup and -⁠hold – fix timing violations. Setup fixes speed paths up; hold fixes slow short paths down.
• -⁠area, -⁠leakage, -⁠power – recover area or power where slack allows. They spend only positive slack, so passing paths stay passing.
• -⁠dynamic – the name points to dynamic power. Confirm its exact behavior with help eco_opt_design before relying on it.

The run writes two files. eco_innovus.tcl holds the changes for Innovus. eco_tempus.tcl holds the same changes for Tempus, so you can check them in STA. A fix of one type never creates new violations of the other types. For example, a hold fix must not push a setup path into failure.

3.2  A hold fix, step by step

Let us consider the most common ECO – a hold fix. Figure 6 shows a path whose data arrives too soon after the clock edge.

• Assume the capture flop needs data stable for 0.08 ns after the edge.
• Before, the data changes 0.05 ns after the edge. Hold slack is 0.05 − 0.08 = −0.03 ns.
• After inserting DLYX1 (0.05 ns), the data changes at 0.10 ns. Hold slack is 0.10 − 0.08 = +0.02 ns.

The tool also checks the setup side of the same path. Adding 0.05 ns is safe only if setup slack was larger than 0.05 ns. Thus, a short path with plenty of setup margin is the easy case. A path tight on both sides needs a smaller step.

Fixing a hold violation with a delay cell

The new data must not arrive inside the hold window

3.3  Prerequisites

The ECO is only as good as the timing it sees. Thus, set up the timing exactly as for signoff before you run it:

• Read SPEF for every active RC corner, for example read_spef -⁠rc_corner rc_max max.spef.gz. A view without parasitics is timed wrongly.
• Include coupling capacitance in the SPEF, so SI is analyzed. Without it, crosstalk delays are missing.
• Set derates, AOCV, globals and multi-CPU usage before eco_opt_design. A setting added afterwards does not affect the fixes already chosen.

3.4  Fillers, flops and buffers

• Fillers. The ECO deletes filler cells automatically to make room for new cells. However, it must know which cells are fillers. Declare them with set_filler_mode -⁠core {{cells1} {cells2}}. Undeclared fillers look like logic, so the tool may find no free space.
• Flops. Flop sizing is enabled with set_eco_opt_mode -⁠optimize_sequential_cells true. Turn it on when flops drive weak or slow outputs.
• Buffer removal. set_eco_opt_mode -⁠delete_inst true lets area recovery delete buffers, not only resize cells.
• Inverter pairs. A setup fix may insert a pair of inverters instead of a buffer. The function is unchanged, because two inversions cancel.

3.5  Incremental steps

You can split the ECO into several steps, for example DRV first and then hold. Turn this on with -⁠allow_multiple_incremental true. Give each step a unique -⁠eco_file_prefix, so its files do not overwrite the previous ones.

tempus> set_eco_opt_mode -load_eco_opt_db ECOdb                ;# saved DB
tempus> set_eco_opt_mode -allow_multiple_incremental true      ;# allow steps
tempus> set_eco_opt_mode -eco_file_prefix DRV_FIX              ;# files for step 1
tempus> eco_opt_design -drv                                    ;# step 1: DRV
tempus> set_eco_opt_mode -eco_file_prefix HOLD_FIX             ;# files for step 2
tempus> eco_opt_design -hold                                   ;# step 2: hold

• -⁠allow_multiple_incremental true – lets a later step build on the changes of an earlier one in the same session.
• -⁠eco_file_prefix <name> – names the files of the next step. P&R then applies the files in order.

If the fixes need room, recover area first. However, keep the number of steps small. Each extra step adds risk that P&R and Tempus timing drift apart. As a rule of thumb, two or three steps are normal; ten steps are a warning sign.

4  Fixing SI, IR drop and cell EM

4.1  The concept

Some violations are not plain slew or slack problems. Crosstalk can add delay or cause glitches. IR drop can slow a cell. Electromigration (EM) can wear out a cell's metal over years of use. All of these are fixed through eco_opt_design -⁠drv, with extra switches. Figure 7 shows the switches.

Set these switches during DB generation too. The DB must contain the data each fix needs. For example, a DB written without SI slew data cannot drive an SI slew fix.

What eco_opt_design -⁠drv can fix

Each extra fix has its own switch

4.2  Glitch

A glitch is a short noise pulse on a quiet net, caused by a switching neighbor. If it is large enough, a receiving flop can capture a wrong value. Glitch, Overshoot and Undershoot explains the analysis.

Turn glitch fixing on with -⁠fix_glitch true; the default is false. It runs before max transition and max capacitance fixing. To analyze glitches first, use set_si_mode -⁠enable_glitch_report true and report_noise.

To fix only glitches, switch max capacitance and max transition fixing off. The switch settings are:

-⁠fix_max_cap false -⁠fix_max_tran false -⁠fix_glitch true

Then run eco_opt_design -⁠drv. Glitches that remain are usually fixed with SI routing rules in P&R, such as extra spacing or shielding.

4.3  SI slew

Crosstalk also makes transitions slower. This is the SI slew. A slow transition can break a max transition limit, and it slows every cell it drives. Turn its fixing on with -⁠fix_si_slew true; the default is false.

Only DRV fixing honors SI slew. Thus, run the SI slew fix first, before setup and hold. The DB must be generated with set_si_mode -⁠enable_drv_with_delta_slew true.

To report SI slew violations, set timing_use_incremental_si_transition and use report_constraint:

tempus> set_global timing_use_incremental_si_transition true   ;# include SI slew
tempus> report_constraint -drv_violation_type max_transition \
    -all_violators                                              ;# list them

4.4  Crosstalk delta delay

A delta delay is the extra delay crosstalk adds to a net. A large crosstalk delta delay can be fixed directly with -⁠fix_xtalk true; the default is false. It needs these SI settings:

tempus> set_si_mode -separate_delta_delay_on_data true \
    -delta_delay_annotation_mode lumpedOnNet   ;# delta lumped per net

Then turn on -⁠fix_xtalk true. The SI settings make the delta visible per net, so the tool can pick nets to fix. Two thresholds decide which nets qualify. Figure 8 shows them on one net.

When a crosstalk delta gets fixed

A net qualifies when its delta is large and its slack is small enough

• -⁠setup_xtalk_delta_threshold and -⁠hold_xtalk_delta_threshold – fix when the delta is at least this value. The default is 0.3 ns.
• -⁠setup_xtalk_slack_threshold and -⁠hold_xtalk_slack_threshold – fix when the slack is at most this value. The default is 1000 ns, which in practice means any slack.

For example, a net with a 0.35 ns delta on a path with +0.4 ns slack qualifies by default. Setting the slack threshold to 0 ns would skip it, because the path already passes.

4.5  IR drop

IR drop is the supply voltage lost in the power grid. A cell with less voltage switches more slowly. IR drop fixing reads voltage-drop results from Voltus. It downsizes attackers that have positive slack when they hurt a violating victim. A smaller attacker draws less current, so the victim sees less drop.

Figure 9 shows the idea on two cells that share a supply rail. The attacker gives up some of its spare slack, and the victim gains back the voltage it needs.

Fixing IR drop by downsizing an attacker

A smaller attacker draws less current from the shared rail

tempus> set_eco_opt_mode -fix_ir_drop true \
    -load_irdrop_db voltus_ir_dir     ;# illustrative folder of Voltus files
tempus> eco_opt_design -drv           ;# fix

• -⁠fix_ir_drop true – turns IR-aware fixing on.
• -⁠load_irdrop_db <dir> – points to the Voltus result files. Without them, the tool has no IR data.

4.6  Cell EM

Cell EM is electromigration inside a cell's own metal. High current slowly moves metal atoms until a wire thins or breaks. Cell EM fixing is turned on with -⁠fix_cell_em true; the default is false. -⁠em_type selects the current type: all, avg, peak or rms. The default is all.

• -⁠fix_clock_drv true fixes only cell EM. Max transition, max capacitance, glitch, crosstalk and IR fixing are all off.
• From Innovus, the same fix is setSignoffOptMode -⁠fixCellEm true with its -⁠emType option, followed by signoffOptDesign -⁠drv.
• Cell EM depends on switching activity. Apply set_switching_activity or read_activity_file before eco_opt_design.
• In Paradime, the activity comes from the dynamic power view, or from eco_dynamic_view or eco_leakage_view. Set these before distribute_read_design.

5  Path-based ECO and power recovery

5.1  The concept

Graph-based analysis (GBA) takes the worst slew at each pin, for every path through it. This is safe but pessimistic. Path-based analysis (PBA) retimes each path with its own real slews. PBA slack is usually better. Path-Based and Robustness Analysis explains why.

For an ECO, this difference is valuable. Let us consider a path with −0.02 ns GBA slack and +0.01 ns PBA slack. A GBA-driven ECO fixes it. A PBA-driven ECO leaves it alone, because it never really failed. Thus, PBA fixes fewer false violations and finds more spare slack for power recovery.

5.2  PBA settings

These set_eco_opt_mode options control PBA inside the ECO:

• -⁠max_slack – the default 0 retimes only failing paths. With -⁠max_slack 10, paths with positive GBA slack are retimed too. Their PBA slack is larger, which gives more margin for power recovery.
• -⁠max_paths and -⁠nworst – limit the work. Large limits give better results but longer runtime.
• -⁠pba_effort – trades runtime for accuracy. Use high for the final ECO.

Figure 10 compares the two -⁠max_slack settings on seven paths. Filled dots are retimed with PBA; hollow dots keep their GBA slack.

Which paths -⁠max_slack sends to PBA

The limit is a GBA slack value; paths at or below it are retimed

5.3  Power recovery

eco_opt_design -⁠power recovers leakage, switching and internal power together. Run report_power first, so the tool has the power numbers. It uses default toggle rates, or a VCD or SAIF activity file if one is read. Flops do not need a Fixed attribute for this step.

Power recovery spends positive slack. A path with +0.3 ns slack may get slower, higher-Vt cells and end at +0.02 ns. Figure 11 shows the typical effect on the slack distribution.

Slack before and after power recovery

Positive slack is turned into lower power

5.4  Setup recovery

After power recovery, some paths may sit close to failing. -⁠setup_recovery true with eco_opt_design -⁠setup recovers setup by Vt swaps only. There is no re-routing, so P&R impact stays small.

Note that a power ECO can touch 30 to 80 percent of the netlist. That is far more than a timing ECO. Thus, plan P&R time and a full re-signoff for it.

5.5  The best total-power flow

The following script gives the best total-power result. It runs PBA power recovery and then setup recovery.

read_design -physical_data postroute.enc.dat top        ;# restore the routed design
# Assuming views, timing settings and parasitics are loaded here
set_multi_cpu_usage -localCpu 16                         ;# CPUs
set_eco_opt_mode -allow_multiple_incremental true        ;# two steps follow
set_eco_opt_mode -retime path_slew_propagation -check_type both -pba_effort high \
  -max_slack 10 -max_paths 10000000 -nworst 50 -delete_inst true \
  -save_eco_opt_db ECO-DB-PBA                            ;# PBA DB with margin
write_eco_opt_db                                         ;# write it
report_power                                             ;# power numbers first
eco_opt_design -power                                    ;# or -leakage
set_eco_opt_mode -max_slack 0 -max_paths 10000000 -nworst 50 \
  -save_eco_opt_db ECO-DB-PBA2                           ;# fresh DB after power
write_eco_opt_db                                         ;# write it
set_eco_opt_mode -load_eco_opt_db ECO-DB-PBA2 -setup_recovery true   ;# recover setup
eco_opt_design -setup                                    ;# Vt swaps only
update_timing -full                                      ;# re-time to check

Let us read it in three parts:

1. The first DB uses -⁠max_slack 10, so many passing paths are retimed with PBA. This exposes the most margin for power.
2. Power recovery spends that margin. -⁠delete_inst true also lets it delete buffers.
3. The second DB uses -⁠max_slack 0 and drives setup recovery on the changed design. A final update_timing -⁠full confirms the result.

If you run -⁠leakage instead of -⁠power, load ECO-DB-PBA with -⁠load_eco_opt_db.

6  Paradime: interactive multi-view ECO

6.1  The concept

A design with many views needs many timers. Paradime splits the job. A master session holds the netlist and makes the changes. One client per view holds the full timing for that view. The master prompt is dmmmc>. Figure 12 shows the setup.

A kitchen is a fair picture. The head chef decides each change to the dish. Each taster checks it for one customer's taste. The chef never tastes; the tasters never cook.

Paradime needs the -⁠eco license plus DMMMC licenses. Turn it on with set_distribute_mmmc_mode -⁠eco true, and set remoteHost to the number of views. For example, eight views need eight remote hosts.

6.2  Why it matters

A single session with many views may run out of memory or take too long. Paradime spreads the views over machines, yet keeps one place where changes are made. You can also work interactively: make a change, ask every view, and decide the next step.

Paradime master and clients

One master for changes, one client per view for timing

6.3  Two ways to start

Full load. The clients load and time the design from scripts:

set_distribute_mmmc_mode -eco true                     ;# ECO master
distribute_read_design -design_script design_loading.tcl \
    -outdir Output                                     ;# load design
distribute_views -script report.tcl -views {v1 v2}     ;# time views
eco_opt_design -hold                                   ;# fix hold

• -⁠design_script – the design-loading script each client runs.
• -⁠outdir – where client output goes.
• distribute_views -⁠script report.tcl – here report.tcl reads SPEF for all corners and runs update_timing -⁠full.

Restored sessions. Each client restores a saved single-view (SMSC) session, one per view:

distribute_read_design -restore_design $Views \
    -restore_dir SavedDir -outdir Output         ;# by folder
# OR, by explicit session paths:
distribute_read_design -outdir Output \
    -restore_design { {v1 path/session.enc} {v2 path2/session.enc} }   ;# by path

• Libraries are read from the paths in the view definition. They are not saved with the session, so those paths must still be valid.
• If the sessions do not share one common view-definition file, use set_distribute_mmmc_mode -⁠merge_view_definition.

Restoring typically saves time, because each client skips the design import.

6.4  Paradime variables

Set these variables before distribute_read_design:

For example, limiting eco_hold_view_list to the fast views keeps slow views out of hold fixing. Setup is still checked in the setup views, so a hold fix cannot quietly break setup.

6.5  Rules that catch beginners

• Manual ECOs on the master only. Use change_cell or add_repeater on the master. Never send them through distribute_command – that errors out.
• The master has no timing. Query timing with distribute_command [-views {..}] {cmds}, or switch to interactive mode with set_distribute_mmmc_mode -⁠interactive true -⁠views {..|all}.
• Settings go to both sides. Apply timing settings on the master and on the clients, for example distribute_command {set_timing_derate -⁠late -⁠cell_delay 1.01} (an illustrative derate). Push variables with set_distribute_variables "vars" [-view v].
• DB options differ. -⁠load_eco_opt_db is not accepted, and -⁠save_eco_opt_db is ignored. The clients hold live timing instead.

7  Running the ECO from Innovus

7.1  The concept

Many teams prefer to drive the ECO from the P&R session. The design, the extraction and the ECO route are already there. Innovus provides two commands for this. signoffTimeDesign runs Tempus signoff timing. signoffOptDesign runs the Tempus ECO. If no DB exists, signoffOptDesign runs signoffTimeDesign first by itself.

Figure 13 lays out a typical hold ECO driven from Innovus as eight steps. The top row measures and fixes; the bottom row routes, extracts and measures again. The script below follows the same steps.

A hold ECO driven from Innovus

The ECO loop of Figure 2, run inside one P&R session

7.2  Command breakdown

• -⁠reportOnly – reuse the existing parasitics instead of extracting again. Use it when the extraction is already current.
• -⁠noEcoRoute – apply the fixes but do not run the ECO route. You can then inspect the changes and route them yourself.
• setSignoffOptMode -⁠preStaTcl FILE – source a Tcl file before the STA step, for example extra timing settings.

setMultiCpuUsage -localCpu 16 -remoteHost 4 -cpuPerRemoteHost 4   ;# CPUs and hosts
setExtractRCMode -effortLevel high                    ;# signoff-quality extraction
extractRC                                             ;# extract
signoffTimeDesign -reportOnly -outDir RPT_initial     ;# timing before
signoffOptDesign -hold -noEcoRoute                    ;# Tempus hold ECO
ecoRoute                                              ;# route the changes
extractRC                                             ;# extract again
signoffTimeDesign -reportOnly -outDir RPT_final       ;# timing after

The script follows the ECO loop of Figure 2 inside one tool. High-effort extraction gives signoff-quality parasitics. The two -⁠outDir folders keep the before and after reports apart, so you can compare them.

8  Hierarchical and master/clone ECO

8.1  The concept

Large chips are split into partitions, each built as its own block by its own team. Some violating paths cross a partition boundary – from the top into a partition, or from one partition to another. A flat ECO cannot fix these cleanly, because each partition is implemented separately.

Hierarchical ECO fixes them while keeping each partition's ports unchanged. Unchanged ports mean the other teams' work stays valid. Figure 14 shows the idea.

• dont_touch a module to freeze it. The ECO then fixes the path elsewhere.
• The output is one ECO file for the top and one per partition.

Hierarchical ECO

A path from one partition to another, fixed without changing the ports

8.2  The script

read_lib $liberty                                     ;# libraries
read_view_definition viewDefinition.tcl               ;# views
read_verilog "my_top.v pnt1.v pnt2.v"                 ;# top and partitions
set_top_module my_top                                 ;# link
source timing_settings_and_derating.tcl               ;# signoff settings
merge_hierarchical_def "my_top.def pnt1.def pnt2.def" ;# one physical view
read_spef -rc_corner rc_max "my_top_max.spef ptn1_max.spef ptn2_max.spef"   ;# per corner
set_eco_opt_mode -buffer_cell_list "BUFX1 BUFX2 BUFX8 DLYX1 DLYX4"   ;# allowed buffers
set_eco_opt_mode -load_eco_opt_db myECODB             ;# saved DB (mandatory)
set_eco_opt_mode -partition_list_file partition.txt   ;# list the partitions
eco_opt_design -hold                                  ;# fix hold

• merge_hierarchical_def – joins the top and partition DEF files into one physical picture.
• read_spef -⁠rc_corner with several files – reads the top and partition parasitics for one corner. Repeat it for every corner.
• -⁠buffer_cell_list – the buffers and delay cells the ECO may insert.
• -⁠partition_list_file – names the partitions, so the tool knows where each fix belongs.

8.3  Master/clone modules

A master module may be placed several times as clones. A fix in the master changes every clone. Thus, the fix must be safe in every clone's context.

For example, a master with three clones may see a hold violation in only one. Adding a delay cell there also slows the other two. If one of them has tight setup, the fix is rejected.

Figure 15 draws this example. The same delay cell lands in all three clones, and each clone judges it in its own context.

One master fix, three clone contexts

A fix in the master is accepted only if every clone stays clean

• Turn this on with set_eco_opt_mode -⁠optimize_replicated_modules true. It also enables leakage optimization.
• Each fix is checked against every clone context. The output is one ECO file per clone set.
• partition.txt lists the master cell.
• Load the design with Verilog plus merge_hierarchical_def, or with read_design -⁠physical_data top.enc.dat my_top.
• Clones with different orientations must be assembled in a Tempus or Innovus DB. One flat DEF or Verilog cannot describe them.

9  High-capacity ECO (HC-ECO)

9.1  The concept

On a very large design, even the ECO DB is big. Most of the design, however, has no violations and needs no fixes. HC-ECO keeps only the violating zone – the part of the design the fixes can touch. Figure 16 shows the two-session flow.

The reduction can use three inputs:

• STA violations – keep only the violating paths and their neighborhood.
• A partition list – keep inter-partition or intra-partition paths.
• Path groups plus a DRV pin list – keep chosen groups and chosen pins.

The HC-ECO flow

Session 1 writes a reduced DB; session 2 loads it and fixes

Set the reduction before DB generation. In Paradime, set it before distribute_read_design. A reduction set later has no effect, because the DB is already written.

9.2  Session 1: STA and the reduced DB

Session 1 can be any STA style: single view (SMSC), concurrent MMMC, DMMMC or DSTA.

# Assuming the design is loaded in logical mode with SPEF
set_eco_opt_mode -enable_high_capacity_eco true                 ;# reduced DB
set_eco_opt_mode -retime path_slew_propagation -max_path 5000000 \
  -nworst 50 -max_slack 0 -pba_effort high                      ;# PBA settings
set_power_analysis_mode -dynamic_power_view setup_ss_1p170v     ;# power view
set_eco_opt_mode -save_eco_opt_db HC-ECO-DB                     ;# DB folder
write_eco_opt_db                                                ;# write it

9.3  Session 2: the physical ECO

Session 2 loads the full physical design and fixes inside the zone. The DB pointer comes first.

set_multi_cpu_usage -localCpu 32                      ;# CPUs
set_eco_opt_mode -load_eco_opt_db HC-ECO-DB           ;# DB pointer BEFORE the load
read_lib -lef {tech.lef macro.lef}                    ;# physical libraries
read_view_definition viewDefinition.tcl               ;# views
read_verilog "top.v blockA.v blockB.v"                ;# netlists
set_top_module top                                    ;# link
set_filler_mode -core {{FILL1 FILL2 FILL4}}           ;# illustrative filler names
merge_hierarchical_def top.def blockA.def blockB.def  ;# physical data
# Assuming SPEF for every corner is read here
set_eco_opt_mode -pba_effort high                     ;# PBA effort
set_eco_opt_mode -partition_list_file partition.txt   ;# partitions
set_eco_opt_mode -optimize_replicated_modules true    ;# masters and clones
eco_opt_design -hold                                  ;# fix hold

The ECO is then implemented separately at the top and in each block in P&R. A final full-flat STA confirms the result. The flat run matters, because no reduced view can prove the whole chip is clean.

9.4  Variants

Set these in both sessions:

• -⁠inter_partition {blockA blockB} -⁠intra_partition {} – select inter-partition paths. This also enables inter-partition power recovery.
• -⁠include_setup_path_groups {pg} and -⁠include_hold_path_groups {pg} – select path groups.
• -⁠select_drv_pin_file {f} – select DRV pins from a file.

Then run eco_opt_design with -⁠drv, -⁠setup or -⁠hold.

9.5  Paradime HC-ECO

In Paradime, the master is prepared in this order:

1. set_distribute_mmmc_mode -⁠eco true and the merge setting -⁠merge_view_definitions.
2. The high-capacity setting -⁠enable_high_capacity_eco, and the filler setting -⁠set_filler_mode with the value {-corePrefix FILL}.
3. The eco_* variables.
4. distribute_read_design -⁠restore_design all -⁠restore_dir DIR -⁠outdir rpt.
5. The PBA, partition and replicated-module settings, then eco_opt_design -⁠hold.

For the path-group variant, give -⁠path_group_file_name {script} before the restore.

10  Top plus boundary model ECO

10.1  The concept

Another way to shrink a large ECO is to replace each child block with a boundary model (BM). A BM keeps only the block's interface timing – the logic near its ports. The top-level ECO then runs on the top plus the BMs. The same reductions and approaches as HC-ECO apply. Boundary models are taught in detail in Timing Models.

The difference from HC-ECO is where the saving comes from. HC-ECO keeps the full design but cuts the DB to a zone. Top plus BM cuts the blocks themselves down to their interfaces.

10.2  Generating the boundary models

timing_enable_boundary_model_eco must be set before the BMs are generated, and again before Top plus BM is loaded. The BM generation run uses these settings:

set timing_full_context_flow 1                  ;# full context timing
set timing_sum_cc_for_all_small_aggressors 1    ;# sum small aggressor coupling
set timing_write_aggressors_to_disk 1           ;# keep aggressor data
set timing_enable_model_accurate_nef 1          ;# accurate model data
set timing_enable_filter_fanin_cone 1           ;# filter the fanin cone
timing_enable_boundary_model_eco                ;# BM ready for ECO
create_boundary_model -overwrite -no_extended_fanin -latch_effort_low -skip_constants \
  -exclude_no_delay_port -dir BM_${BlockName}_${view}   ;# one BM per block and view

The comments describe each setting by its name. Confirm the exact meaning of each global and option with help before you change them. The -⁠dir value builds one folder per block and per view.

10.3  The two sessions

Figure 17 puts the block run and the two top sessions side by side. Note the two red steps.

The Top plus boundary model ECO flow

Generate models per block, then run the two HC-ECO sessions at the top

• Session 1: run timing_enable_boundary_model_eco, load Top plus BM, set -⁠enable_high_capacity_eco true and -⁠save_eco_opt_db HC-ECO-DB, then write_eco_opt_db.
• Session 2: the same as HC-ECO session 2.
• Paradime: the STA run ends with save_design view1.enc, one saved session per view.

11  Other ECO controls

The controls in this section are ECO-mode switches given by name. Confirm the command that takes each one with help set_eco_opt_mode before scripting.

11.1  Buffering along the route

A new buffer placed off the existing route forces a detour. The extra wire adds delay and may need new routing tracks. Along-route buffering places the buffer on the route instead. Figure 18 shows the difference. Turn it on with -⁠along_route_buffering true.

Buffering along the route

Node coordinates let the tool place the buffer on the existing wire

The SPEF must carry *N node coordinates – the location of each point on the wire. The Quantus command output_db -⁠type spef -⁠subtype standard -⁠disable_subnodes false writes such a SPEF. From Innovus, put that command in the Quantus command file given to setExtractRCMode -⁠qrcCmdFile. Then read the SPEF with read_spef -⁠starN.

11.2  Choosing which endpoints to fix

Sometimes only some endpoints should be fixed. For example, a debug block may be excluded, or one view may need extra margin. These controls apply to setup and hold only.

• -⁠select_hold_endpoints <file> and -⁠select_setup_endpoints <file> – a file of include and exclude lines.
• Each line has the form <view|V*> include|exclude <pins | minSlack maxSlack | E*>. The default is V* include E*, meaning all views and all endpoints.
• Pin wildcards are not allowed. An excluded pin cannot be included again later.

V* include E*
func_ff exclude u_core/u_dbg/reg_0/D

The first line keeps everything. The second line drops one debug pin in the func_ff view only. Because an excluded pin cannot return, put exclusions after the broad includes.

Figure 19 splits a selection line and a margin line into their parts.

Anatomy of endpoint control lines

Each line names a view, an action or margin, and a target

Margins make chosen endpoints tighter:

• -⁠specify_hold_endpoints_margin and -⁠specify_setup_endpoints_margin – lines of the form <view|V*> <margin_ns> <eps|range|E*>.
• A positive margin makes the target tighter. For example, a 0.02 ns margin in effect asks for at least +0.02 ns of slack.
• -⁠optimize_core_only true fixes register-to-register paths only. Other path-group settings are then ignored.

11.3  SmartMMMC

SmartMMMC uses a separate DB that STA writes for the ECO to load next to the normal DB. As its name suggests, it is aimed at runs with many views; confirm its scope with help set_eco_opt_mode.

1. In STA, set -⁠create_smart_mmmc_db true before write_eco_opt_db.
2. In the ECO, set -⁠load_smart_mmmc_db <dir> before the design load, together with -⁠load_eco_opt_db.
3. In Paradime, use set_distribute_mmmc_mode -⁠enable_smart_mmmc with -⁠eco true. Then restore with distribute_read_design -⁠restore_design -⁠read_db all -⁠restore_dir SavedDir.

The order matters, as with HC-ECO. A DB pointer set after the load is too late.

11.4  Clock skewing for setup

A setup violation can also be fixed by delaying the capture clock. The data then has more time to arrive. Let us consider Figure 20, with a 0.9 ns period and a 0.06 ns setup time.

• Before: required time 0.9 − 0.06 = 0.84 ns. Arrival 0.92 ns. Slack 0.84 − 0.92 = −0.08 ns.
• After a 0.10 ns capture delay: required time 1.0 − 0.06 = 0.94 ns. Slack 0.94 − 0.92 = +0.02 ns.

The switch settings are:

-⁠clock_cell_list {CKBUFX1 CKBUFX4} -⁠allow_skewing true   (both mandatory)
-⁠clock_max_level 2                                      (illustrative limit)

Then run eco_opt_design -⁠setup. The clock cell list and -⁠allow_skewing true are both mandatory. -⁠clock_max_level <n> limits how deep in the clock tree the tool may change. Skewing never creates new hold violations. That matters, because a later capture edge also makes hold harder on the same flop.

Fixing setup by skewing the capture clock

A later capture edge gives the data more time

12  Metal (post-mask) ECO

12.1  The concept

A chip is built in layers. The base layers hold the transistors; the metal layers wire them together. After the masks for the base layers are made, a change that touches transistors is very costly. A metal ECO changes only the metal layers. It builds new logic from spare cells that are already on the chip.

Figure 21 shows the idea with gate-array (GA) fillers. A GA filler is a spare cell whose transistors can be wired into a real gate by metal alone.

A metal-only fix with gate-array fillers

Spare filler sites become a working buffer by changing only metal

Turn it on with set_eco_opt_mode -⁠post_mask true; the default is false. It fixes DRV, setup and hold, using GA fillers or PFC fillers.

12.2  Prerequisites

• Every cell is placed, and the rows are at 100 percent density. The spare fillers fill every gap.
• Fillers are fine-grained enough to make room for new cells. Coarse fillers leave no small spaces to use.
• No tie cells in tie mode, if tie insertion is not allowed.
• GA cells are the library cells that share the SITE of the GA fillers. Only they can be built from GA filler sites.

12.3  PFC and GA scripts

With PFC fillers, list each filler kind separately – plain fillers, decap fillers and regular logic cells:

tempus> set_eco_opt_mode -post_mask true          ;# metal-only ECO
tempus> set_eco_opt_mode -use_pfc_filler_list \
    {ln10_ecofiller40x ln10_ecofiller4x}          ;# PFC fillers
tempus> set_eco_opt_mode -use_pfc_decap_list \
    {ln10_ecofillsgcap40x}                        ;# PFC decaps
tempus> set_eco_opt_mode -use_pfc_regular_list \
    {ln10_ecodlybt01 ln10_ecobuft04}              ;# PFC logic cells
tempus> set_tie_hi_lo_mode -cell \
    {{ln108_ecotielo01 ln10_ecotiehi01}}          ;# tie cells
tempus> eco_opt_design -hold                      ;# fix hold

With GA fillers, the list is shorter:

tempus> set_eco_opt_mode -post_mask true          ;# metal-only ECO
tempus> set_eco_opt_mode -use_ga_filler_list \
    {GFILL1BWP GFILL2BWP GFILL4BWP GFILL10BWP}    ;# GA fillers
tempus> set_tie_hi_lo_mode -cell {{GATIELO GATIEHI}}   ;# GA tie cells
tempus> eco_opt_design -hold                      ;# fix hold

set_tie_hi_lo_mode -⁠cell names the tie-low and tie-high cells the metal ECO may use. The cell names in both scripts belong to example libraries; use the names from your own library.

13  Worked example: closing hold and slew in one loop

The task. A routed block has three views: func_ss, func_ff and test_ss. Signoff shows 210 hold violations in func_ff and 35 max transition violations, some caused by SI slew. The goal is to close both in one ECO loop. All numbers are illustrative.

Figure 22 shows a typical order of ECO steps. This example needs only the first two: DRV with SI slew, then hold.

Step 1 – write the DB at signoff

SI slew fixing needs SI slew in the DB. Thus, the SI setting goes in before the DB is written.

# Assuming the design, views and SPEF for every RC corner are loaded
set_si_mode -enable_drv_with_delta_slew true       ;# keep SI slew for the DRV fix
# Turn on -fix_si_slew true in the ECO mode here (confirm the form with help)
set_eco_opt_mode -save_eco_opt_db blk_ecodb        ;# DB folder
write_eco_opt_db                                   ;# one DB per view

What to check: the DB folder holds six files – a .tgd and a .ckd for each of the three views.

A typical order of ECO steps

Each step builds on the one before it

Step 2 – fix DRV first, then hold

SI slew is honored only by DRV fixing. Thus, DRV runs first, and hold runs second as an incremental step.

set_eco_opt_mode -load_eco_opt_db blk_ecodb             ;# reuse signoff timing
set_filler_mode -core {{FILL1 FILL2 FILL4 FILL8}}       ;# illustrative filler names
set_eco_opt_mode -allow_multiple_incremental true       ;# two steps
# Keep -fix_si_slew true on for this step (ECO-mode switch)
set_eco_opt_mode -eco_file_prefix DRV_FIX               ;# step 1 files
eco_opt_design -drv                                     ;# fix slew first
set_eco_opt_mode -eco_file_prefix HOLD_FIX              ;# step 2 files
eco_opt_design -hold                                    ;# then hold

What to check: two sets of ECO files, one per prefix. Load each eco_tempus.tcl in a Tempus session and re-time. The slew and hold violations should be gone, and no new setup violation should appear.

Step 3 – apply in P&R and sign off again

1. Source the eco_innovus.tcl files from both steps in Innovus, in order.
2. Run the ECO route and re-extract every RC corner.
3. Rerun signoff STA in all three views.
4. Confirm that no new setup violations appeared in func_ss.

If a few hold violations remain after P&R, compare the Tempus result with the post-route timing. A gap often points to routing changes rather than to the ECO choice.

14  Common mistakes and how to catch them

Table 4 lists the mistakes from this chapter. Each row starts with the symptom you would see first.

15  Interview preparation

Chapter recap

Key takeaways

• A signoff ECO fixes the last violations with small transforms, decided by signoff timing and applied by P&R.
• The ECO reads a timing DB per view. Save it at signoff, or let the ECO build it in batch.
• Read SPEF for every RC corner, declare fillers, and keep incremental steps few with unique prefixes.
• SI, IR and EM fixes all run through -⁠drv. Set them at DB time and run SI slew first.
• PBA retiming finds margin for power recovery; setup recovery then repairs paths with Vt swaps.
• Paradime splits changes (master) from timing (clients). Hierarchical, HC-ECO and Top plus BM scale the flow.
• Endpoint files, along-route buffering, clock skewing and metal ECO give finer control late in the project.

Command recap

Compact glossary

Practical perspective

In a real tapeout, the ECO stage is where schedules often slip. Every extra loop costs a P&R ECO, an extraction and a full signoff run. Thus, the habits in this chapter pay off directly. Save the DB at signoff, read every corner, fix SI slew first, and keep the steps few. Teams that follow these habits often close in one or two loops instead of several.

What comes next

The large-design ECO flows used boundary models without explaining them in depth. In the next chapter, Timing Models, we see how ETM, boundary models, PTM and ILM capture a block's timing. We also see when to use each one.