Corners and temperature
Run a transistor in the five SKY130 process corners and at three temperatures, read which way each one moves the current, and check what a corner's name really means.
- Lesson
- E2.7
- Time
- 25 min
- Runs on
- ngspice
- Check
- 5 questions
A bias block was signed off in the typical corner at room temperature. Its NMOS (W = 2 µm, L = 0.15 µm) carried 113.7 µA at VGS = 0.9 V. On a wafer from the slow end of the factory's spread the same device carried 80.5 µA, 29 % less, and the circuit it fed ran out of time.
.lib sky130 ss. The dashed line is the typical corner's current at VGS = 0.9 V.- Gate voltage V_GS 900 mVI_D in ss = 80.49 µA
At VGS = 0.9 V the slow corner gives 80.49 µA. The typical corner gave 113.7 µA: a 29.2 % shortfall from the models alone.
- Gate voltage V_GS 1.8 VI_D in ss = 878 µA
With the full 1.8 V on the gate the slow device carries 878.0 µA. The gap shrinks with more gate drive, but it never closes.
Read the plot as a table
| Gate voltage V_GS | I_D in ss |
|---|---|
| 0 V | 2.02 pA |
| 200 mV | 39.21 pA |
| 400 mV | 6.532 nA |
| 600 mV | 890.2 nA |
| 800 mV | 30.24 µA |
| 900 mV | 80.49 µA |
| 1 V | 151.2 µA |
| 1.2 V | 319.7 µA |
| 1.4 V | 502.2 µA |
| 1.6 V | 689.6 µA |
| 1.8 V | 878 µA |
Measured on ngspice 42 with the SKY130 ss models (181 points). The tt value is the lab's in_short.
The typical model describes the middle of the spread, and the chip on your bench can sit anywhere in it.
Five corners of one process
No two wafers come out of a factory alike. The gate oxide is a little thicker or thinner, the implants a little heavier or lighter, so thresholds and mobilities shift. A process corner is a set of model parameters for a wafer at the edge of that spread. In a fast corner a device carries more current at the same bias; in a slow one, less.
SKY130 ships five FET corners, chosen with .lib sky130 tt|ff|ss|sf|fs: typical (tt), both fast (ff), both slow (ss) and two skewed corners in which one kind is fast and the other slow. A corner moves every device on the die together. Two neighbouring devices that differ from each other show mismatch, a separate effect (E7).
| Corner | NMOS 2/0.15, 0.9 V | PMOS 2/0.15, 0.9 V | NMOS 5/1, 1.2 V | PMOS 5/1, 1.2 V |
|---|---|---|---|---|
| tt | 113.7 µA | 39.82 µA | 215.3 µA | 7.666 µA |
| ff | 156.5 µA (+38 %) | 67.03 µA (+68 %) | 230.4 µA (+7 %) | 8.696 µA (+13 %) |
| ss | 80.49 µA (−29 %) | 21.20 µA (−47 %) | 200.9 µA (−7 %) | 6.726 µA (−12 %) |
| sf | 154.7 µA (+36 %) | 24.80 µA (−38 %) | 235.5 µA (+9 %) | 4.979 µA (−35 %) |
| fs | 77.44 µA (−32 %) | 56.85 µA (+43 %) | 195.9 µA (−9 %) | 11.06 µA (+44 %) |
Drain current at |VDS| = 1.8 V and the stated |VGS|, from five runs of the lab bench (one per corner), 27 °C. W/L in µm. Percentages against tt.
Short devices swing much more than long ones, and the PMOS more than the NMOS. Now read the skewed rows. In the corner named sf the NMOS is fast at both sizes, as fast as ff at L = 0.15 µm and faster than ff at L = 1 µm. Its PMOS is slow, and at L = 1 µm slower than in ss. The corner named fs is the mirror image.
.lib sky130 sf. Both devices are swept by the same |VGS|. The dashed lines are the tt currents at 0.9 V.- NMOS in sf
- PMOS in sf
- |VGS| for both devices 900 mVNMOS in sf = 154.7 µA · PMOS in sf = 24.8 µA
At 0.9 V the sf NMOS carries 154.7 µA, above its tt line; the sf PMOS carries 24.80 µA, below its tt line. Fast NMOS, slow PMOS.
- |VGS| for both devices 1.8 VNMOS in sf = 1.044 mA · PMOS in sf = 328.2 µA
At full drive: 1044 µA for the NMOS and 328.2 µA for the PMOS.
Read the plot as a table
| |VGS| for both devices | NMOS in sf | PMOS in sf |
|---|---|---|
| 0 V | 5.238 pA | 66.53 pA |
| 200 mV | 476.3 pA | 2.843 nA |
| 400 mV | 62.51 nA | 49.61 nA |
| 600 mV | 5.179 µA | 912.7 nA |
| 800 mV | 77.59 µA | 11.24 µA |
| 900 mV | 154.7 µA | 24.8 µA |
| 1 V | 244.6 µA | 44.27 µA |
| 1.2 V | 438.2 µA | 98.42 µA |
| 1.4 V | 639 µA | 167.9 µA |
| 1.6 V | 841.9 µA | 246.4 µA |
| 1.8 V | 1.044 mA | 328.2 µA |
Measured on ngspice 42 with the SKY130 sf models (181 points per curve).
The comment at the top of the PDK's sky130.lib.spice describes sf as slow N, fast P. The parameters in parameters_fet_sf.spice do the opposite, at both sizes we measured. So read a corner name as a label, and learn what each corner does from your own devices. In this track sf means fast NMOS, slow PMOS, because that is what it simulates.
- The middle
ttis the centre of the spread: 113.7 µA for the NMOS and 39.8 µA for the PMOS. - Both together
ffandssmove both kinds the same way. They bound the speed of a gate that needs both, such as an inverter driving a long wire. - Skewed
sfandfsmove the two kinds apart. By measurement,sfis the fast-NMOS, slow-PMOS corner, so it sits bottom right. These corners decide the switching threshold of an inverter and the balance of a latch.
Read the diagram as text
- fs: NMOS −32 %, PMOS +43 %
- ff: NMOS +38 %, PMOS +68 %
- tt: NMOS 113.7 µA, PMOS 39.8 µA
- ss: NMOS −29 %, PMOS −47 %
- sf: NMOS +36 %, PMOS −38 %
- tt → ff
- tt → ss
- tt → sf
- tt → fs
Arrows point from the typical corner to each process corner.
Temperature pulls two ways
Heat changes the same device in two opposite ways. The threshold falls: the lab extracts 0.726 V at −40 °C, 0.676 V at 27 °C and 0.612 V at 125 °C, about −0.69 mV per kelvin. Lower threshold, more current. Mobility also falls, because a hotter lattice scatters the carriers more often. Lower mobility, less current.
tt. Below the crossing the hot device carries more; above it, less.- −40 °C
- 27 °C
- 125 °C
- Gate voltage V_GS 900 mV−40 °C = 102.8 µA · 27 °C = 113.7 µA · 125 °C = 121.1 µA
At 0.9 V the threshold wins: 102.8 µA at −40 °C, 113.7 µA at 27 °C, 121.1 µA at 125 °C. Hot is fast.
- Gate voltage V_GS 977.6 mV−40 °C = 174.8 µA · 27 °C = 178.1 µA · 125 °C = 178.1 µA
At 0.978 V the 27 °C and 125 °C curves cross at 178.1 µA. This is the zero-temperature-coefficient (ZTC) point: the two effects cancel.
- Gate voltage V_GS 1.8 V−40 °C = 1.089 mA · 27 °C = 985.3 µA · 125 °C = 860 µA
At 1.8 V mobility wins: 1089 µA at −40 °C, 985.3 µA at 27 °C, 860.0 µA at 125 °C. Hot is slow.
Read the plot as a table
| Gate voltage V_GS | −40 °C | 27 °C | 125 °C |
|---|---|---|---|
| 0 V | 1.813 pA | 2.82 pA | 154.5 pA |
| 200 mV | 3.441 pA | 141.9 pA | 6.219 nA |
| 400 mV | 929.7 pA | 18.92 nA | 235.5 nA |
| 600 mV | 433.7 nA | 1.929 µA | 6.201 µA |
| 800 mV | 34.54 µA | 47.97 µA | 60.58 µA |
| 900 mV | 102.8 µA | 113.7 µA | 121.1 µA |
| 977.6 mV | 174.8 µA | 178.1 µA | 178.1 µA |
| 1 V | 197.3 µA | 197.9 µA | 195.5 µA |
| 1.2 V | 411.9 µA | 386.4 µA | 358 µA |
| 1.4 V | 636.3 µA | 584.1 µA | 524.9 µA |
| 1.6 V | 862.8 µA | 784.7 µA | 692.8 µA |
| 1.8 V | 1.089 mA | 985.3 µA | 860 µA |
Measured on ngspice 42 with the SKY130 tt models (361 points per curve).
So "hot is slow" holds only with plenty of gate drive. A gate driven from the full 1.8 V supply is slowest at 125 °C. A circuit whose devices see less than about 1 V of drive is slowest when cold. That reversal is called temperature inversion, and it is why timing sign-off checks both ends of the temperature range.
Corners and temperature combine. A design is checked at process, voltage and temperature together (PVT): for example ss at 125 °C for a fully driven gate, and ss at −40 °C for a weakly driven one. E7 builds that sign-off table for an amplifier.
Asked in interviews
What are process corners? What do ff, ss, sf and fs mean?
Model sets for wafers at the edges of the manufacturing spread: both kinds fast, both slow, or one fast and one slow. Check the naming in your PDK: in SKY130 the corner named sf measures as fast NMOS and slow PMOS, the reverse of its header comment. Corners move all devices together; mismatch is device to device.
How does temperature change drain current?
The threshold falls (here about −0.69 mV/K), which adds current, and mobility falls, which removes it. At low overdrive the threshold wins and current rises with heat; at high overdrive mobility wins. They cancel at the zero-temperature-coefficient point, 0.978 V for this device.
What is temperature inversion?
In a low-voltage or weakly driven circuit, the gates are slowest cold instead of hot, because the falling threshold matters more than the falling mobility. Sign-off must check both temperature extremes, not only the hot one.
Try it
Run the netlist: the first sweep moves every gate together, and meas reads the four currents of the table's tt row. The second part extracts the threshold at three temperatures with the constant-current method of E2.2.
Now change the first line after the title to .lib sky130 sf and run again; then try fs. Before each run, predict the sign of every change against tt. One run loads one corner: the library is read when the netlist loads, so five corners take five runs.
Run the netlist to draw its plots here.
in_short = 1.136616e-04 ip_short = 3.981684e-05 in_long = 2.153058e-04 ip_long = 7.665913e-06 vth_27 = 6.760515e-01 vth_m40 = 7.255910e-01 vth_125 = 6.119765e-01Check yourself
Remember
- A process corner is a model set for a wafer at the edge of the spread; corners move every device together, while mismatch is device to device.
- SKY130 has tt, ff, ss and two skewed corners; by measurement its
sfis fast NMOS, slow PMOS (the header comment says the opposite), and fs is the reverse. - Short devices and PMOS devices swing most between corners: −29 % to +38 % for the short NMOS, −47 % to +68 % for the short PMOS.
- Heat lowers both threshold and mobility: below the zero-temperature-coefficient point current rises with temperature, above it current falls.
- Sign-off checks process, voltage and temperature together, at both temperature ends.