Wire Delay: Elmore, the Pi Model and Effective Capacitance

Compute a wire delay by hand, see what a long wire does to its driver and receiver, and cut it with repeaters.

You Will Be Able To

  • Write the Elmore sum of an RC ladder or tree and estimate a wire delay by hand.
  • Explain how a wire slows both its own delay and the edge the next cell sees.
  • Find the effective capacitance a driver sees, and say how it depends on the driver and the edge.
  • Choose when repeaters make a long wire faster.
Lesson
T1.5
Time
40 min
Runs on
OpenROAD
Check
5 questions

A buf_2 drives 1 mm of met1 to a buf_1. Without wires the path arrives at 0.4776 ns, and with the wire's R and C from a SPEF file at 1.1580 ns. Three report lines take the difference: the driver, the wire, and the receiver, which now sees a slow 0.86 ns edge.

Timing path1 Mm of met1 Between Two BuffersFlop ff1 launches, drv (buf_2) drives the long net w, rcv (buf_1) receives it, and ff2 captures.

ff1 → ff2setup check, latest arrival3 cells

  • Clock latency
  • Cell delay
  • Wire delay
  • Setup time
1 mm of met1 between two buffers00.20.40.60.811.21.41.61.82time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1 dfxtp_1drv/A → drv/X (buf_2): 0.6003 ns2 buf_2drv/X → rcv/A: 0.1043 nsrcv/A → rcv/X (buf_1): 0.1405 ns3setup −0.0466 nsslack +0.7954 ns · metarrival 1.1580required 1.95341 mm of met1 between two buffers00.511.52time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1drv/A → drv/X (buf_2): 0.6003 ns2 buf_2drv/X → rcv/A: 0.1043 nsrcv/A → rcv/X (buf_1): 0.1405 ns3setup −0.0466 nsslack +0.7954 ns · metarrival 1.1580required 1.9534
  1. Driver

    drv takes 0.6003 ns, against 0.1010 ns with no wire.

  2. Wire

    The net arc from drv/X to rcv/A adds 0.1043 ns.

  3. Receiver

    rcv takes 0.1405 ns, against 0.0712 ns with no wire.

Read the path as a table

Launch: clk rise edge at 0.0000 ns, clock latency 0.0000 ns (ideal). Capture: clk rise edge at 2.0000 ns, latency 0.0000 ns, library setup time −0.0466 ns.

Arrival 1.1580 ns, required 1.9534 ns, slack 0.7954 ns (met).

CellArcThroughEdgeDelay (ns)Time (ns)Slew (ns)
1ff1/CLK → ff1/Qdfxtp_1rise0.31290.31290.0340
ff1/Q → drv/Awire q1rise0.00000.31290.0340
2drv/A → drv/Xbuf_2rise0.60030.91320.8405
drv/X → rcv/Awire wrise0.10431.01750.8609
3rcv/A → rcv/Xbuf_1rise0.14051.15800.0470
rcv/X → ff2/Dwire rrise0.00001.15800.0470
The report
Startpoint: ff1 (rising edge-triggered flip-flop clocked by clk)
Endpoint: ff2 (rising edge-triggered flip-flop clocked by clk)
Path Group: clk
Path Type: max

Fanout       Cap      Slew     Delay      Time   Description
-------------------------------------------------------------------------------------
                    0.1000    0.0000    0.0000   clock clk (rise edge)
                              0.0000    0.0000   clock network delay (ideal)
                    0.1000    0.0000    0.0000 ^ ff1/CLK (sky130_fd_sc_hd__dfxtp_1)
     1    0.0018    0.0340    0.3129    0.3129 ^ ff1/Q (sky130_fd_sc_hd__dfxtp_1)
                                                 q1 (net)
                    0.0340    0.0000    0.3129 ^ drv/A (sky130_fd_sc_hd__buf_2)
     1    0.1746    0.8405    0.6003    0.9132 ^ drv/X (sky130_fd_sc_hd__buf_2)
                                                 w (net)
                    0.8609    0.1043    1.0175 ^ rcv/A (sky130_fd_sc_hd__buf_1)
     1    0.0017    0.0470    0.1405    1.1580 ^ rcv/X (sky130_fd_sc_hd__buf_1)
                                                 r (net)
                    0.0470    0.0000    1.1580 ^ ff2/D (sky130_fd_sc_hd__dfxtp_1)
                                        1.1580   data arrival time

                    0.1000    2.0000    2.0000   clock clk (rise edge)
                              0.0000    2.0000   clock network delay (ideal)
                              0.0000    2.0000   clock reconvergence pessimism
                                        2.0000 ^ ff2/CLK (sky130_fd_sc_hd__dfxtp_1)
                             -0.0466    1.9534   library setup time
                                        1.9534   data required time
-------------------------------------------------------------------------------------
                                        1.9534   data required time
                                       -1.1580   data arrival time
-------------------------------------------------------------------------------------
                                        0.7954   slack (MET)

OpenSTA 3.1.0 on the sky130 tt library, with the authored SPEF lwire-1000. Ideal 2 ns clock, 0.1 ns slew.

A Wire Is a Resistor and a Capacitor

Every micron of wire has some resistance, and some capacitance to its neighbours and the substrate. The platform's met1 values give 1205.6 Ω and 172.4 fF for 1 mm, and both grow in proportion to the length.

Both are spread along the wire. A timer models that as a ladder: small resistors in series, with a capacitor at each joint. The SPEF file here cuts 1 mm into 10 segments of 120.6 Ω and 17.2 fF each.

Part of tools/pd/spef/lwire-1000.spef: the units, the net, its two pins, the capacitors and the resistors.
*T_UNIT 1 NS
*C_UNIT 1 PF
*R_UNIT 1 KOHM

*D_NET w 0.172375
*CONN
*I drv:X O
*I rcv:A I
*CAP
1 drv:X 0.00861875
2 w:1 0.01723750
...
*RES
1 drv:X w:1 0.12056500
...

SPEF (Standard Parasitic Exchange Format) is normally written by an extractor, as in L6.4; this one is hand-made teaching data. The header sets the units. The *D_NET total is the wire alone, because pin capacitance comes from the Liberty file, so the report shows 174.6 fF on drv.

From the Ladder to What the Driver SeesThe timer reduces the wire to a small model at the driver pin.
  • Driver
  • Interconnect
  • Load
  • Model
  • Data
  • Analysis
From the ladder to what the driver seesThe timer reduces the wire to a small model at the driver pin. driver drv (buf_2) Drives net w interconnect 10-segmentladder 1205.6 Ω, 172.4 fFin all load rcv/A (buf_1) 2.2 fF pin model Pi model C2 29.8 fF, R 590Ω, C1 144.8 fF load Effective C One capacitor for thedriver's arc chargeswire delayreducematch the driver
  1. driver drv (buf_2)Drives net w
    • charges → 10-segment ladder
  2. interconnect 10-segment ladder1205.6 Ω, 172.4 fF in all
    • wire delay → rcv/A (buf_1)
    • reduce → Pi model
  3. load rcv/A (buf_1)2.2 fF pin
  4. model Pi modelC2 29.8 fF, R 590 Ω, C1 144.8 fF
    • match the driver → Effective C
  5. load Effective COne capacitor for the driver's arc
  1. The real wire

    The wire delay to rcv/A comes from the full ladder.

  2. Reduce it

    For the driver, the ladder becomes a pi model.

  3. One capacitor

    The calculator finds one capacitor that loads the driver the same way.

Read the diagram as text
  • drv (buf_2): Drives net w
  • 10-segment ladder: 1205.6 Ω, 172.4 fF in all
  • rcv/A (buf_1): 2.2 fF pin
  • Pi model: C2 29.8 fF, R 590 Ω, C1 144.8 fF
  • Effective C: One capacitor for the driver's arc
  • drv (buf_2) → 10-segment ladder: charges
  • 10-segment ladder → rcv/A (buf_1): wire delay
  • 10-segment ladder → Pi model: reduce
  • Pi model → Effective C: match the driver

Arrows show how the timer reduces the wire.

Lumped or Distributed

The simplest model puts the whole wire capacitance on the driver as one capacitor, which in SDC is set_load on the net. It has no resistance, so it says nothing about the time the signal takes to cross the wire.

Model of net wDriver delay (ns)Wire delay (ns)Arrival at ff2 (ns)
No parasitics0.10100.00000.4776
Lumped: set_load 0.1723750.65520.00001.1083
Distributed: SPEF, 1 mm0.60030.10431.1580

Three runs; only the model of net w changes.

The lumped model makes the driver too slow, 0.6552 against 0.6003 ns, and it drops the wire delay completely. Here the errors partly cancel and the lumped arrival is only 49.7 ps early, but other nets differ.

Model50 % delay after a sharp stepNote
One lumped R and C0.69·RCThe time constant times ln 2
Uniform distributed wireabout 0.38·RCThe wire charges from the near end first
Elmore delay of the same wire0.5·RCAn upper bound on the 50 % delay

Textbook step responses, with R and C of the whole wire. A real driver gives a ramp, not a step.

The no-wire and lumped paths
Timing pathNo Parasitics

ff1 → ff2setup check, latest arrival3 cells

  • Clock latency
  • Cell delay
  • Setup time
No parasitics00.20.40.60.811.21.41.61.82time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3053 ns1 dfxtp_1drv/A → drv/X (buf_2): 0.1010 ns2rcv/A → rcv/X (buf_1): 0.0712 ns3setup −0.0874 nsslack +1.4350 ns · metarrival 0.4776required 1.9126No parasitics00.511.52time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3053 ns1drv/A → drv/X (buf_2): 0.1010 ns2rcv/A → rcv/X (buf_1): 0.0712 nssetup −0.0874 nsslack +1.4350 ns · metarrival 0.4776required 1.9126
Read the path as a table

Launch: clk rise edge at 0.0000 ns, clock latency 0.0000 ns (ideal). Capture: clk rise edge at 2.0000 ns, latency 0.0000 ns, library setup time −0.0874 ns.

Arrival 0.4776 ns, required 1.9126 ns, slack 1.4350 ns (met).

CellArcThroughEdgeDelay (ns)Time (ns)Slew (ns)
1ff1/CLK → ff1/Qdfxtp_1fall0.30530.30530.0224
ff1/Q → drv/Awire q1fall0.00000.30530.0224
2drv/A → drv/Xbuf_2fall0.10100.40630.0221
drv/X → rcv/Awire wfall0.00000.40630.0221
3rcv/A → rcv/Xbuf_1fall0.07120.47760.0215
rcv/X → ff2/Dwire rfall0.00000.47760.0215
The report
Startpoint: ff1 (rising edge-triggered flip-flop clocked by clk)
Endpoint: ff2 (rising edge-triggered flip-flop clocked by clk)
Path Group: clk
Path Type: max

Fanout       Cap      Slew     Delay      Time   Description
-------------------------------------------------------------------------------------
                    0.1000    0.0000    0.0000   clock clk (rise edge)
                              0.0000    0.0000   clock network delay (ideal)
                    0.1000    0.0000    0.0000 ^ ff1/CLK (sky130_fd_sc_hd__dfxtp_1)
     1    0.0016    0.0224    0.3053    0.3053 v ff1/Q (sky130_fd_sc_hd__dfxtp_1)
                                                 q1 (net)
                    0.0224    0.0000    0.3053 v drv/A (sky130_fd_sc_hd__buf_2)
     1    0.0020    0.0221    0.1010    0.4063 v drv/X (sky130_fd_sc_hd__buf_2)
                                                 w (net)
                    0.0221    0.0000    0.4063 v rcv/A (sky130_fd_sc_hd__buf_1)
     1    0.0017    0.0215    0.0712    0.4776 v rcv/X (sky130_fd_sc_hd__buf_1)
                                                 r (net)
                    0.0215    0.0000    0.4776 v ff2/D (sky130_fd_sc_hd__dfxtp_1)
                                        0.4776   data arrival time

                    0.1000    2.0000    2.0000   clock clk (rise edge)
                              0.0000    2.0000   clock network delay (ideal)
                              0.0000    2.0000   clock reconvergence pessimism
                                        2.0000 ^ ff2/CLK (sky130_fd_sc_hd__dfxtp_1)
                             -0.0874    1.9126   library setup time
                                        1.9126   data required time
-------------------------------------------------------------------------------------
                                        1.9126   data required time
                                       -0.4776   data arrival time
-------------------------------------------------------------------------------------
                                        1.4350   slack (MET)

The driver sees only the receiver pin.

Timing pathLumped: set_load 0.172375 [get_nets w]

ff1 → ff2setup check, latest arrival3 cells

  • Clock latency
  • Cell delay
  • Setup time
Lumped: set_load 0.172375 [get_nets w]00.20.40.60.811.21.41.61.82time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1 dfxtp_1drv/A → drv/X (buf_2): 0.6552 ns2 buf_2rcv/A → rcv/X (buf_1): 0.1403 ns3setup −0.0465 nsslack +0.8452 ns · metarrival 1.1083required 1.9535Lumped: set_load 0.172375 [get_nets w]00.511.52time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1drv/A → drv/X (buf_2): 0.6552 ns2 buf_2rcv/A → rcv/X (buf_1): 0.1403 ns3setup −0.0465 nsslack +0.8452 ns · metarrival 1.1083required 1.9535
Read the path as a table

Launch: clk rise edge at 0.0000 ns, clock latency 0.0000 ns (ideal). Capture: clk rise edge at 2.0000 ns, latency 0.0000 ns, library setup time −0.0465 ns.

Arrival 1.1083 ns, required 1.9535 ns, slack 0.8452 ns (met).

CellArcThroughEdgeDelay (ns)Time (ns)Slew (ns)
1ff1/CLK → ff1/Qdfxtp_1rise0.31290.31290.0340
ff1/Q → drv/Awire q1rise0.00000.31290.0340
2drv/A → drv/Xbuf_2rise0.65520.96800.8374
drv/X → rcv/Awire wrise0.00000.96800.8374
3rcv/A → rcv/Xbuf_1rise0.14031.10830.0466
rcv/X → ff2/Dwire rrise0.00001.10830.0466
The report
Startpoint: ff1 (rising edge-triggered flip-flop clocked by clk)
Endpoint: ff2 (rising edge-triggered flip-flop clocked by clk)
Path Group: clk
Path Type: max

Fanout       Cap      Slew     Delay      Time   Description
-------------------------------------------------------------------------------------
                    0.1000    0.0000    0.0000   clock clk (rise edge)
                              0.0000    0.0000   clock network delay (ideal)
                    0.1000    0.0000    0.0000 ^ ff1/CLK (sky130_fd_sc_hd__dfxtp_1)
     1    0.0018    0.0340    0.3129    0.3129 ^ ff1/Q (sky130_fd_sc_hd__dfxtp_1)
                                                 q1 (net)
                    0.0340    0.0000    0.3129 ^ drv/A (sky130_fd_sc_hd__buf_2)
     1    0.1746    0.8374    0.6552    0.9680 ^ drv/X (sky130_fd_sc_hd__buf_2)
                                                 w (net)
                    0.8374    0.0000    0.9680 ^ rcv/A (sky130_fd_sc_hd__buf_1)
     1    0.0017    0.0466    0.1403    1.1083 ^ rcv/X (sky130_fd_sc_hd__buf_1)
                                                 r (net)
                    0.0466    0.0000    1.1083 ^ ff2/D (sky130_fd_sc_hd__dfxtp_1)
                                        1.1083   data arrival time

                    0.1000    2.0000    2.0000   clock clk (rise edge)
                              0.0000    2.0000   clock network delay (ideal)
                              0.0000    2.0000   clock reconvergence pessimism
                                        2.0000 ^ ff2/CLK (sky130_fd_sc_hd__dfxtp_1)
                             -0.0465    1.9535   library setup time
                                        1.9535   data required time
-------------------------------------------------------------------------------------
                                        1.9535   data required time
                                       -1.1083   data arrival time
-------------------------------------------------------------------------------------
                                        0.8452   slack (MET)

The whole wire capacitance on the driver, with no resistance.

Elmore Delay by Hand

The Elmore delay to a node is a sum over the resistors on the way there. Each resistor is multiplied by all the capacitance beyond it, as seen from the driver, including the receiver pins.

  1. Walk from the driver

    Each 120.6 Ω resistor charges everything past it, including the 2.2 fF pin of rcv.

  2. Close the sum

    For a uniform ladder with half a segment at each end, the sum is exactly R·(C/2 + Cpin). R and C are for the whole wire.

  3. Put in the numbers

    1205.6 Ω × (86.2 fF + 2.2 fF) = 106.6 ps, and the tool prints 104.3 ps.

DC sweepWire Delay Against Length, Hand Elmore and OpenSTAFour authored SPEFs: 250 µm, 500 µm, 1 mm and 2 mm of met1, each as a 10-segment ladder.
  • OpenSTA wire delay
  • Hand Elmore
0100 ps200 ps300 ps400 ps500 psWire delay0.4 mm0.6 mm0.8 mm1.0 mm1.2 mm1.4 mm1.6 mm1.8 mm2.0 mmWire length0100 ps200 ps300 ps400 ps500 psWire delay0.5 mm1.0 mm1.5 mm2.0 mmWire length
Read the plot as a table
Wire lengthOpenSTA wire delayHand Elmore
400 µm19.1 ps19.24 ps
600 µm42.54 ps43.15 ps
800 µm73.42 ps74.85 ps
1 mm104.3 ps106.6 ps
1.2 mm162.8 ps169.4 ps
1.4 mm221.3 ps232.3 ps
1.6 mm279.8 ps295.2 ps
1.8 mm338.3 ps358.1 ps
2 mm396.8 ps420.9 ps

OpenSTA points from four pd-route runs; Elmore points from the formula above.

Doubling the length doubles R and C, so the delay grows about fourfold: 7.1 ps, 27.1 ps, 104.3 ps, 396.8 ps. The hand Elmore number tracks it closely up to 1 mm, and at 2 mm it is 24.1 ps higher.

Elmore is the limit for a slow input ramp, and a proven upper bound for RC trees. Under a sharp step the real delay is lower, about 0.38·RC. OpenSTA uses the driver's real ramp, 0.84 ns at 1 mm. That is far slower than the wire's 106.6 ps, so the tool lands close to Elmore. At 2 mm the wire is a larger share of the ramp, and the tool falls further below.

The other three lengths
Timing path250 µm

ff1 → ff2setup check, latest arrival3 cells

  • Clock latency
  • Cell delay
  • Wire delay
  • Setup time
250 µm00.20.40.60.811.21.41.61.82time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1 dfxtp_1drv/A → drv/X (buf_2): 0.2270 ns2 buf_2drv/X → rcv/A: 0.0071 nsrcv/A → rcv/X (buf_1): 0.1065 ns3setup −0.0440 nsslack +1.3024 ns · metarrival 0.6536required 1.9560250 µm00.511.52time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1drv/A → drv/X (buf_2): 0.2270 ns2drv/X → rcv/A: 0.0071 nsrcv/A → rcv/X (buf_1): 0.1065 ns3setup −0.0440 nsslack +1.3024 ns · metarrival 0.6536required 1.9560
Read the path as a table

Launch: clk rise edge at 0.0000 ns, clock latency 0.0000 ns (ideal). Capture: clk rise edge at 2.0000 ns, latency 0.0000 ns, library setup time −0.0440 ns.

Arrival 0.6536 ns, required 1.9560 ns, slack 1.3024 ns (met).

CellArcThroughEdgeDelay (ns)Time (ns)Slew (ns)
1ff1/CLK → ff1/Qdfxtp_1rise0.31290.31290.0340
ff1/Q → drv/Awire q1rise0.00000.31290.0340
2drv/A → drv/Xbuf_2rise0.22700.53990.2222
drv/X → rcv/Awire wrise0.00710.54700.2226
3rcv/A → rcv/Xbuf_1rise0.10650.65360.0357
rcv/X → ff2/Dwire rrise0.00000.65360.0357
The report
Startpoint: ff1 (rising edge-triggered flip-flop clocked by clk)
Endpoint: ff2 (rising edge-triggered flip-flop clocked by clk)
Path Group: clk
Path Type: max

Fanout       Cap      Slew     Delay      Time   Description
-------------------------------------------------------------------------------------
                    0.1000    0.0000    0.0000   clock clk (rise edge)
                              0.0000    0.0000   clock network delay (ideal)
                    0.1000    0.0000    0.0000 ^ ff1/CLK (sky130_fd_sc_hd__dfxtp_1)
     1    0.0018    0.0340    0.3129    0.3129 ^ ff1/Q (sky130_fd_sc_hd__dfxtp_1)
                                                 q1 (net)
                    0.0340    0.0000    0.3129 ^ drv/A (sky130_fd_sc_hd__buf_2)
     1    0.0453    0.2222    0.2270    0.5399 ^ drv/X (sky130_fd_sc_hd__buf_2)
                                                 w (net)
                    0.2226    0.0071    0.5470 ^ rcv/A (sky130_fd_sc_hd__buf_1)
     1    0.0017    0.0357    0.1065    0.6536 ^ rcv/X (sky130_fd_sc_hd__buf_1)
                                                 r (net)
                    0.0357    0.0000    0.6536 ^ ff2/D (sky130_fd_sc_hd__dfxtp_1)
                                        0.6536   data arrival time

                    0.1000    2.0000    2.0000   clock clk (rise edge)
                              0.0000    2.0000   clock network delay (ideal)
                              0.0000    2.0000   clock reconvergence pessimism
                                        2.0000 ^ ff2/CLK (sky130_fd_sc_hd__dfxtp_1)
                             -0.0440    1.9560   library setup time
                                        1.9560   data required time
-------------------------------------------------------------------------------------
                                        1.9560   data required time
                                       -0.6536   data arrival time
-------------------------------------------------------------------------------------
                                        1.3024   slack (MET)

Authored SPEF lwire-250.

Timing path500 µm

ff1 → ff2setup check, latest arrival3 cells

  • Clock latency
  • Cell delay
  • Wire delay
  • Setup time
500 µm00.20.40.60.811.21.41.61.82time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1 dfxtp_1drv/A → drv/X (buf_2): 0.3583 ns2 buf_2drv/X → rcv/A: 0.0271 nsrcv/A → rcv/X (buf_1): 0.1245 ns3setup −0.0449 nsslack +1.1324 ns · metarrival 0.8228required 1.9551500 µm00.511.52time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1drv/A → drv/X (buf_2): 0.3583 ns2 buf_2drv/X → rcv/A: 0.0271 nsrcv/A → rcv/X (buf_1): 0.1245 ns3setup −0.0449 nsslack +1.1324 ns · metarrival 0.8228required 1.9551
Read the path as a table

Launch: clk rise edge at 0.0000 ns, clock latency 0.0000 ns (ideal). Capture: clk rise edge at 2.0000 ns, latency 0.0000 ns, library setup time −0.0449 ns.

Arrival 0.8228 ns, required 1.9551 ns, slack 1.1324 ns (met).

CellArcThroughEdgeDelay (ns)Time (ns)Slew (ns)
1ff1/CLK → ff1/Qdfxtp_1rise0.31290.31290.0340
ff1/Q → drv/Awire q1rise0.00000.31290.0340
2drv/A → drv/Xbuf_2rise0.35830.67120.4273
drv/X → rcv/Awire wrise0.02710.69830.4299
3rcv/A → rcv/Xbuf_1rise0.12450.82280.0393
rcv/X → ff2/Dwire rrise0.00000.82280.0393
The report
Startpoint: ff1 (rising edge-triggered flip-flop clocked by clk)
Endpoint: ff2 (rising edge-triggered flip-flop clocked by clk)
Path Group: clk
Path Type: max

Fanout       Cap      Slew     Delay      Time   Description
-------------------------------------------------------------------------------------
                    0.1000    0.0000    0.0000   clock clk (rise edge)
                              0.0000    0.0000   clock network delay (ideal)
                    0.1000    0.0000    0.0000 ^ ff1/CLK (sky130_fd_sc_hd__dfxtp_1)
     1    0.0018    0.0340    0.3129    0.3129 ^ ff1/Q (sky130_fd_sc_hd__dfxtp_1)
                                                 q1 (net)
                    0.0340    0.0000    0.3129 ^ drv/A (sky130_fd_sc_hd__buf_2)
     1    0.0884    0.4273    0.3583    0.6712 ^ drv/X (sky130_fd_sc_hd__buf_2)
                                                 w (net)
                    0.4299    0.0271    0.6983 ^ rcv/A (sky130_fd_sc_hd__buf_1)
     1    0.0017    0.0393    0.1245    0.8228 ^ rcv/X (sky130_fd_sc_hd__buf_1)
                                                 r (net)
                    0.0393    0.0000    0.8228 ^ ff2/D (sky130_fd_sc_hd__dfxtp_1)
                                        0.8228   data arrival time

                    0.1000    2.0000    2.0000   clock clk (rise edge)
                              0.0000    2.0000   clock network delay (ideal)
                              0.0000    2.0000   clock reconvergence pessimism
                                        2.0000 ^ ff2/CLK (sky130_fd_sc_hd__dfxtp_1)
                             -0.0449    1.9551   library setup time
                                        1.9551   data required time
-------------------------------------------------------------------------------------
                                        1.9551   data required time
                                       -0.8228   data arrival time
-------------------------------------------------------------------------------------
                                        1.1324   slack (MET)

Authored SPEF lwire-500.

Timing path2 Mm

ff1 → ff2setup check, latest arrival3 cells

  • Clock latency
  • Cell delay
  • Wire delay
  • Setup time
2 mm00.20.40.60.811.21.41.61.82time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1 dfxtp_1drv/A → drv/X (buf_2): 0.9942 ns2 buf_2drv/X → rcv/A: 0.3968 nswirercv/A → rcv/X (buf_1): 0.1489 ns3setup −0.0503 nsslack +0.0969 ns · metarrival 1.8528required 1.94972 mm00.511.52time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1drv/A → drv/X (buf_2): 0.9942 ns2 buf_2drv/X → rcv/A: 0.3968 nswirercv/A → rcv/X (buf_1): 0.1489 ns3setup −0.0503 nsslack +0.0969 ns · metarrival 1.8528required 1.9497
Read the path as a table

Launch: clk rise edge at 0.0000 ns, clock latency 0.0000 ns (ideal). Capture: clk rise edge at 2.0000 ns, latency 0.0000 ns, library setup time −0.0503 ns.

Arrival 1.8528 ns, required 1.9497 ns, slack 0.0969 ns (met).

CellArcThroughEdgeDelay (ns)Time (ns)Slew (ns)
1ff1/CLK → ff1/Qdfxtp_1rise0.31290.31290.0340
ff1/Q → drv/Awire q1rise0.00000.31290.0340
2drv/A → drv/Xbuf_2rise0.99421.30711.6552
drv/X → rcv/Awire wrise0.39681.70391.8166
3rcv/A → rcv/Xbuf_1rise0.14891.85280.0633
rcv/X → ff2/Dwire rrise0.00001.85280.0633
The report
Startpoint: ff1 (rising edge-triggered flip-flop clocked by clk)
Endpoint: ff2 (rising edge-triggered flip-flop clocked by clk)
Path Group: clk
Path Type: max

Fanout       Cap      Slew     Delay      Time   Description
-------------------------------------------------------------------------------------
                    0.1000    0.0000    0.0000   clock clk (rise edge)
                              0.0000    0.0000   clock network delay (ideal)
                    0.1000    0.0000    0.0000 ^ ff1/CLK (sky130_fd_sc_hd__dfxtp_1)
     1    0.0018    0.0340    0.3129    0.3129 ^ ff1/Q (sky130_fd_sc_hd__dfxtp_1)
                                                 q1 (net)
                    0.0340    0.0000    0.3129 ^ drv/A (sky130_fd_sc_hd__buf_2)
     1    0.3469    1.6552    0.9942    1.3071 ^ drv/X (sky130_fd_sc_hd__buf_2)
                                                 w (net)
                    1.8166    0.3968    1.7039 ^ rcv/A (sky130_fd_sc_hd__buf_1)
     1    0.0017    0.0633    0.1489    1.8528 ^ rcv/X (sky130_fd_sc_hd__buf_1)
                                                 r (net)
                    0.0633    0.0000    1.8528 ^ ff2/D (sky130_fd_sc_hd__dfxtp_1)
                                        1.8528   data arrival time

                    0.1000    2.0000    2.0000   clock clk (rise edge)
                              0.0000    2.0000   clock network delay (ideal)
                              0.0000    2.0000   clock reconvergence pessimism
                                        2.0000 ^ ff2/CLK (sky130_fd_sc_hd__dfxtp_1)
                             -0.0503    1.9497   library setup time
                                        1.9497   data required time
-------------------------------------------------------------------------------------
                                        1.9497   data required time
                                       -1.8528   data arrival time
-------------------------------------------------------------------------------------
                                        0.0969   slack (MET)

Authored SPEF lwire-2000. The slack is nearly gone.

The Edge Slows Down Along the Wire

A wire does not only delay an edge; it also stretches it. The driver's output edge is already slow, because it charges a big load. The far end of the wire turns slower still, and the receiving cell pays for that slow input edge.

WireSlew at drv/X (ns)Slew at rcv/A (ns)rcv delay (ns)
None0.02210.02210.0712
250 µm0.22220.22260.1065
500 µm0.42730.42990.1245
1000 µm0.84050.86090.1405
2000 µm1.65521.81660.1489

The same runs as the plot; slew is the 20-80 % transition time.

At 2 mm the receiver sees a 1.82 ns edge, and its own delay is double what it was with no wire. A slow edge also burns short-circuit current, and it can break the library's max_transition limit. Both are reasons to cut long wires, as the repeater section shows.

Effective Capacitance: What the Driver Sees

The driver delay with the SPEF, 0.6003 ns, is less than with the lumped load because of resistive shielding. While the driver's output switches, the far end of the wire is still charging through the wire's resistance. The driver does not feel all of the far capacitance during its own transition.

To measure what it feels, drop the SPEF and raise a plain set_load on net w until the driver delay matches. At set_load 0.1557 the rising driver takes 0.6003 ns, with 157.9 fF in total. That is the effective capacitance, Ceff, of this driver's rising edge.

Driver and edgeTotal load (fF)Ceff (fF)Share
buf_2, rising edge174.6157.990 %
buf_2, falling edge174.4136.978 %
buf_8, rising edge179.7126.170 %

1 mm of met1. Ceff is the set_load total that gives the same driver delay: 0.1557, 0.1349 and 0.1188 pF on net w, plus the pin. The buf_8 drives a buf_8 repeater (design lwire28).

So Ceff belongs to one driver arc and one edge, not to the net. The faster the edge, the less of the far capacitance it sees.

To find it, the calculator first reduces the ladder to a pi model: C2 at the driver, a resistor, then C1. It keeps the ladder's first three moments, which describe how the wire draws current. Here C2 = 29.8 fF, R = 590.4 Ω and C1 = 144.8 fF.

Then it iterates, as in the Dartu, Menezes and Pileggi (DMP) method. It guesses Ceff and looks up the driver ramp for that load. It adjusts Ceff until one capacitor takes the same charge as the pi model up to the 50 % point.

The matching run for the rising edge
Timing pathset_load 0.1557 [get_nets w], No SPEF

ff1 → ff2setup check, latest arrival3 cells

  • Clock latency
  • Cell delay
  • Setup time
set_load 0.1557 [get_nets w], no SPEF00.20.40.60.811.21.41.61.82time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1 dfxtp_1drv/A → drv/X (buf_2): 0.6003 ns2 buf_2rcv/A → rcv/X (buf_1): 0.1396 ns3setup −0.0462 nsslack +0.9010 ns · metarrival 1.0528required 1.9538set_load 0.1557 [get_nets w], no SPEF00.511.52time (ns)launch clk ↑capture clk ↑Data arrivesRequired byff1/CLK → ff1/Q (dfxtp_1): 0.3129 ns1drv/A → drv/X (buf_2): 0.6003 ns2 buf_2rcv/A → rcv/X (buf_1): 0.1396 ns3setup −0.0462 nsslack +0.9010 ns · metarrival 1.0528required 1.9538
Read the path as a table

Launch: clk rise edge at 0.0000 ns, clock latency 0.0000 ns (ideal). Capture: clk rise edge at 2.0000 ns, latency 0.0000 ns, library setup time −0.0462 ns.

Arrival 1.0528 ns, required 1.9538 ns, slack 0.9010 ns (met).

CellArcThroughEdgeDelay (ns)Time (ns)Slew (ns)
1ff1/CLK → ff1/Qdfxtp_1rise0.31290.31290.0340
ff1/Q → drv/Awire q1rise0.00000.31290.0340
2drv/A → drv/Xbuf_2rise0.60030.91320.7585
drv/X → rcv/Awire wrise0.00000.91320.7585
3rcv/A → rcv/Xbuf_1rise0.13961.05280.0452
rcv/X → ff2/Dwire rrise0.00001.05280.0452
The report
Startpoint: ff1 (rising edge-triggered flip-flop clocked by clk)
Endpoint: ff2 (rising edge-triggered flip-flop clocked by clk)
Path Group: clk
Path Type: max

Fanout       Cap      Slew     Delay      Time   Description
-------------------------------------------------------------------------------------
                    0.1000    0.0000    0.0000   clock clk (rise edge)
                              0.0000    0.0000   clock network delay (ideal)
                    0.1000    0.0000    0.0000 ^ ff1/CLK (sky130_fd_sc_hd__dfxtp_1)
     1    0.0018    0.0340    0.3129    0.3129 ^ ff1/Q (sky130_fd_sc_hd__dfxtp_1)
                                                 q1 (net)
                    0.0340    0.0000    0.3129 ^ drv/A (sky130_fd_sc_hd__buf_2)
     1    0.1579    0.7585    0.6003    0.9132 ^ drv/X (sky130_fd_sc_hd__buf_2)
                                                 w (net)
                    0.7585    0.0000    0.9132 ^ rcv/A (sky130_fd_sc_hd__buf_1)
     1    0.0017    0.0452    0.1396    1.0528 ^ rcv/X (sky130_fd_sc_hd__buf_1)
                                                 r (net)
                    0.0452    0.0000    1.0528 ^ ff2/D (sky130_fd_sc_hd__dfxtp_1)
                                        1.0528   data arrival time

                    0.1000    2.0000    2.0000   clock clk (rise edge)
                              0.0000    2.0000   clock network delay (ideal)
                              0.0000    2.0000   clock reconvergence pessimism
                                        2.0000 ^ ff2/CLK (sky130_fd_sc_hd__dfxtp_1)
                             -0.0462    1.9538   library setup time
                                        1.9538   data required time
-------------------------------------------------------------------------------------
                                        1.9538   data required time
                                       -1.0528   data arrival time
-------------------------------------------------------------------------------------
                                        0.9010   slack (MET)

The driver line matches the SPEF run. The wire delay is zero, because this model has no resistance.

Repeaters: Making Long Wires Linear

Wire delay grows with the square of the length. A repeater is a buffer that cuts a long wire into shorter pieces. Each piece is short, so its delay is small, but every repeater adds its own delay.

Total wireNo repeater (ns)One repeater (ns)Three repeaters (ns)
2 mm0.77610.8260—
4 mm2.01841.55981.5911
8 mm6.74314.16583.1272

Data path from ff1/Q to ff2/D, buf_8 driver and repeaters, met1 (designs lwire8, lwire28, lwire84).

At 2 mm a repeater does not pay: 0.8260 ns against 0.7761 ns. At 8 mm three repeaters cut the delay from 6.7431 to 3.1272 ns. At a fixed 2 mm spacing, 4 mm takes 1.5598 ns and 8 mm takes 3.1272 ns: almost exactly double. Repeated wires grow linearly, because only the number of stages grows.

There is a best spacing. On 4 mm, three repeaters at 1 mm gave 1.5911 ns, slightly worse than one at 2 mm. Each extra stage cost more than the wire delay it saved. Placement tools insert repeaters this way during design repair, as L4.4 shows.

Calculators, Layers and Coupling

OpenSTA has several delay calculators, and the pd route takes a calculator option to pick one.

CalculatorDriver (ns)Wire (ns)Arrival (ns)
dmp_ceff_elmore0.60030.10431.1580
dmp_ceff_two_pole0.60030.10651.1601
arnoldi0.58620.11301.1530
prima0.60030.10431.1580
ccs_ceff0.60030.10431.1580
lumped_cap0.65520.00001.1083

lwire with lwire-1000 at an 8 ns clock. dmp_ceff_elmore is the default.

Five calculators agree within about 7.1 ps. ccs_ceff falls back to the table model here, because this library has no current-source timing data. lumped_cap ignores the wire's resistance: no wire delay, and the slowest driver.

Layer, 1 mmR (Ω)C (fF)Driver (ns)Wire (ns)Arrival (ns)
met11206172.40.60030.10431.1580
met3166215.00.78610.01821.2592

The same netlist with a met1 and a met3 ladder of 1 mm. R and C per µm are from the platform's setRC.tcl.

met3 has a seventh of the resistance but more capacitance. On 1 mm the driver pays for that capacitance, so met3 is slower. Upper layers win on longer wires.

Part of a wire's capacitance couples to its neighbours. The SPEF lwire-1000-cc50 puts half of every segment's capacitance on a neighbour net. OpenSTA times it exactly like lwire-1000, 1.1580 ns, because without signal-integrity analysis it grounds coupling capacitors. A neighbour switching the other way would slow this net; L6.5 shows that with SPICE.

Before Routing: Wire Load Models

Before placement no wire exists, so older flows guessed parasitics from fanout with a library table: a wire load model. Modern flows estimate wires from the placement, as L4.5 shows, and extract them after routing.

Asked in interviews
  1. What is the Elmore delay?

    The sum, over each resistor on the way to a node, of that resistor times all the capacitance beyond it. For a uniform wire with a load pin it is R·(C/2 + Cpin).

  2. Why is the effective capacitance smaller than the total?

    The wire's resistance shields the far capacitance. During the driver's transition the far end has not charged yet, so the driver feels less load. A faster driver feels even less.

  3. Why insert repeaters on a long wire?

    Unbuffered wire delay grows with the square of the length. Repeaters at a fixed spacing make it grow linearly, and they restore a sharp edge at each stage.

Try It

This run reads the 1 mm SPEF before your SDC. Run it and open Paths: the drv line is 0.6003 ns and the net line after it is 0.1043 ns.

Now add set_load 0.172375 [get_nets w] and run again. In OpenSTA a net's set_load replaces its SPEF. The wire line drops to zero, and the run matches the lumped row. Then remove it and shorten the period until the path fails.

STAOpenSTA in OpenROAD 26Q2 · sky130hd tt_025C_1v80 Liberty · simulator service
Edit the constraints, then run them

          
Expected: WNS 0.7954 ns
Read the netlist: lwire.v (the bundled design lwire)
// lwire: ff1 -> drv (buf_2) -> net w (a long wire, see make-spef.py) -> rcv (buf_1) -> ff2.
// lwire2 is the same with a repeater (buf_2) in the middle of the wire: nets w and w2.
module lwire (clk, d, q);
  input clk; input d; output q;
  wire q1, w, r;
  sky130_fd_sc_hd__dfxtp_1 ff1 (.CLK(clk), .D(d), .Q(q1));
  sky130_fd_sc_hd__buf_2 drv (.A(q1), .X(w));
  sky130_fd_sc_hd__buf_1 rcv (.A(w), .X(r));
  sky130_fd_sc_hd__dfxtp_1 ff2 (.CLK(clk), .D(r), .Q(q));
endmodule

Check Yourself

Predict the outputThe same lwire block with the 1 mm SPEF, but a 1.2 ns clock. What worst setup slack (WNS) does OpenSTA report, in ns?
create_clock -name clk -period 1.2 [get_ports clk]
set_clock_transition 0.1 [all_clocks]
ConceptA driver feeds R1 = 100 Ω to node n (10 fF). From n, R2 = 100 Ω goes to load A (20 fF) and R3 = 200 Ω goes to load B (10 fF). What is the Elmore delay to load A?
ConceptWith the 1 mm SPEF the driver delay is 0.6003 ns. With the whole 172.4 fF lumped on the net it is 0.6552 ns. Why is the lumped number larger?
ConceptIn this lesson the lumped model gave an arrival only a little earlier than the SPEF. What does that tell you about lumped models?
ConceptA wire with repeaters every 2 mm grows from 4 mm to 8 mm. About how does its delay change?

Common Mistakes

Wrong Model a long wire as a capacitor only, with set_load.

The driver then sees all 172.4 fF at once and the wire adds no delay, so both report lines are wrong.

Fix Give the timer the wire's resistance too: read a SPEF from extraction, or an estimate from placement.

Wrong Quote "0.69·RC" and "0.5·RC" for a wire without saying which model each belongs to.

0.69·RC belongs to one lumped RC under a step, and R·C/2 to the Elmore delay of a uniform wire.

Fix Name the model with the number: lumped RC or distributed line, step input or real driver.

Wrong Leave out the receiver pins when you add up the downstream capacitance.

Each wire resistor also charges the far-end pins, and a big receiver or high fanout adds far more than 2.2 fF.

Fix Add every load pin's capacitance at its node before you multiply by the resistance upstream.

Wrong Add more repeaters to every long wire, because repeaters make wires linear.

Each repeater adds its own delay, and on 4 mm three repeaters gave 1.5911 ns against 1.5598 ns for one.

Fix Space repeaters at the length where one more stage costs as much as the wire delay it saves.

Remember

  • A wire is a distributed resistor and capacitor; a timer models it as an RC ladder read from a SPEF.
  • The Elmore delay multiplies each resistor by all the capacitance beyond it; for a uniform wire it is R·(C/2 + Cpin).
  • A long wire also stretches the edge, so the receiving cell gets slower too.
  • The driver sees an effective capacitance below the total; it is smaller for a faster driver or edge.
  • Unbuffered wire delay grows with length squared; repeaters at a good spacing make it linear.

Practice

Report a problem with this page