Leakage: Threshold, Temperature and Multi-Vt
Explain why a lower threshold makes a transistor faster and leakier, and decide where a design should spend it.
You Will Be Able To
- Compare the off current and the on current of each threshold flavour.
- Predict how temperature and input state change the leakage of a gate.
- Choose which cells of a design get the fast, leaky flavour.
- Lesson
- W1.3
- Time
- 30 min
- Runs on
- ngspice
- Check
- 5 questions
A team wanted more speed, so it swapped every cell of a block to the low-threshold library. Timing passed. The standby current of the block rose more than ten times, and the battery target was lost.
- standard
- HVT
- LVT
Read the plot as a table
| Source-to-gate voltage | standard | HVT | LVT |
|---|---|---|---|
| 0 V | 280.2 pA | 1.514 pA | 3.007 nA |
| 200 mV | 6.282 nA | 2.218 pA | 222.9 nA |
| 400 mV | 109.7 nA | 540.2 pA | 2.659 µA |
| 600 mV | 1.951 µA | 46.65 nA | 10.09 µA |
| 800 mV | 11.71 µA | 1.026 µA | 23.28 µA |
| 1 V | 31.94 µA | 9.399 µA | 42.78 µA |
| 1.2 V | 61.58 µA | 30.77 µA | 68.35 µA |
| 1.4 V | 98.52 µA | 61.61 µA | 98.5 µA |
| 1.6 V | 140 µA | 97.28 µA | 130.6 µA |
| 1.8 V | 183 µA | 133.7 µA | 161.9 µA |
Measured on ngspice 42 with the SKY130 tt models. The curves end close together and start far apart.
The plot shows why: the flavours differ a little when on and enormously when off.
A Lower Threshold Buys Speed and Costs Leakage
The threshold voltage Vt is the gate voltage where a transistor starts to conduct well. A foundry offers the same transistor with several thresholds. These are the flavours: standard, high-threshold (HVT) and low-threshold (LVT).
Read the plot from the left. With no gate drive the device is off, and the current there is the off current Ioff. Below the threshold the current grows several times for every 0.1 V of gate drive. Here that is about 4 times for the standard PMOS and 13 times for the NMOS. So a threshold that is 0.25 V higher can mean hundreds of times less current when off.
Read it from the right. Fully on, the current is the on current Ion. It sets how fast the device charges a load. Here the threshold matters much less.
| Device (W = 1 µm) | Vt | Ioff | Ion | Ioff at 100 °C |
|---|---|---|---|---|
| PMOS standard | 0.520 V | 280 pA | 183 µA | 11.5 nA |
| PMOS HVT | 0.769 V | 1.51 pA | 134 µA | 7.17 pA |
| PMOS LVT (L = 0.35 µm) | 0.215 V | 3.01 nA | 162 µA | 98.9 nA |
| NMOS standard | 0.600 V | 216 fA | 477 µA | — |
| NMOS LVT | 0.502 V | 2.90 pA | 545 µA | — |
tt corner, 1.8 V. Vt is read where the current reaches 100 nA × W/L. That is our definition; other definitions give other numbers.
Compare the standard PMOS with the HVT PMOS. The HVT device leaks 185 times less. It also delivers 27 % less on current, so a gate built from it is slower.
Now compare the two NMOS devices. The LVT device gives 14 % more on current and leaks 13.5 times more.
Heat Multiplies Leakage
.temp changed. Compare the left ends of the curves with the 25 °C plot.- standard
- HVT
- LVT
Read the plot as a table
| Source-to-gate voltage | standard | HVT | LVT |
|---|---|---|---|
| 0 V | 11.52 nA | 7.167 pA | 98.91 nA |
| 200 mV | 129.8 nA | 596.6 pA | 1.354 µA |
| 400 mV | 1.414 µA | 35.78 nA | 5.959 µA |
| 600 mV | 7.954 µA | 656.2 nA | 14.82 µA |
| 800 mV | 22.82 µA | 5.446 µA | 28.49 µA |
| 1 V | 46.17 µA | 20.19 µA | 47.39 µA |
| 1.2 V | 77.13 µA | 44.69 µA | 71.32 µA |
| 1.4 V | 114 µA | 75.95 µA | 99.17 µA |
| 1.6 V | 154.6 µA | 110.6 µA | 129 µA |
| 1.8 V | 196.2 µA | 145.7 µA | 158.6 µA |
Measured on ngspice 42. The on currents barely move. The off currents of the standard and LVT devices rise by more than a decade.
A warm chip leaks more for two reasons. The threshold falls as the temperature rises. The slope below the threshold also gets flatter, so more current remains at zero gate drive. Lesson E2.6 measures both effects.
From 25 °C to 100 °C the standard PMOS goes from 280 pA to 11.5 nA: 41.1 times more.
The HVT reading at 0 V rises only 4.74 times, but that does not make HVT less sensitive to heat. At 25 °C it reads the floor; at 100 °C the subthreshold current has risen above it. Compare at 0.3 V, where both devices are below their thresholds. There the HVT current rises 162 times, and the standard one 17 times.
So budget HVT leakage at the hot corner as well. At 100 °C the HVT device still leaks 1,610 times less than the standard one.
A phone in a pocket or a chip inside a closed box is hot. So leakage is budgeted at the hot corner, never at room temperature.
The Input State Changes the Leakage of a Gate
A gate has several transistors, and its inputs decide which of them are off. So one gate has several leakage currents. The run below holds four NAND2 gates in the four input states.
- A B = 0 0
- A B = 0 1
- A B = 1 0
- A B = 1 1
- Supply voltage 1.8 VA B = 0 0 = 41.31 fA · A B = 0 1 = 215.6 fA · A B = 1 0 = 186.5 fA · A B = 1 1 = 560.5 pA
At 1.8 V the four states read 41.3 fA, 216 fA, 186 fA and 560 pA.
Read the plot as a table
| Supply voltage | A B = 0 0 | A B = 0 1 | A B = 1 0 | A B = 1 1 |
|---|---|---|---|---|
| 600 mV | 30.86 fA | 63.24 fA | 71.73 fA | 22.99 pA |
| 800 mV | 33.4 fA | 80.98 fA | 84.97 fA | 41.99 pA |
| 1 V | 35.55 fA | 101.4 fA | 100.2 fA | 74.73 pA |
| 1.2 V | 37.34 fA | 124.7 fA | 117.8 fA | 129.3 pA |
| 1.4 V | 38.86 fA | 151.3 fA | 137.8 fA | 217.1 pA |
| 1.6 V | 40.11 fA | 181.4 fA | 160.6 fA | 353.8 pA |
| 1.8 V | 41.31 fA | 215.6 fA | 186.5 fA | 560.5 pA |
Measured on ngspice 42. The leakiest state passes 13,600 times the current of the quietest.
| A B | Output | Which transistors are off | Supply current at 1.8 V |
|---|---|---|---|
| 0 0 | 1 | Both NMOS, in series | 41.3 fA |
| 0 1 | 1 | The upper NMOS | 216 fA |
| 1 0 | 1 | The lower NMOS | 186 fA |
| 1 1 | 0 | Both PMOS, side by side | 560 pA |
The NAND2 has two PMOS in parallel and two NMOS in series.
With both inputs high, the two PMOS are off. They sit side by side and both leak, and this PMOS is the leaky device. That state passes 560 pA.
With both inputs low, the two NMOS are off in series. This is the stack effect. The node between them rises to 56.6 mV. That small voltage gives the upper NMOS a negative gate-to-source voltage, and the body effect raises its threshold. The same voltage also lowers the drain voltage of the lower one. The pair leaks 5.22 times less than one NMOS alone.
A designer can use this. A block that sleeps can park its inputs in the state that leaks least.
Multi-Vt: Fast Cells Only Where Timing Needs Them
- Storage
- Datapath
- Data
- The long path sets the speed
These gates have no time to spare. They keep the standard or low-threshold flavour, and they pay its leakage.
- The short path can be slow
This data arrives early. Swap its cells to HVT: each PMOS leaks 185 times less, and the path still meets timing.
Read the diagram as text
- Launch register: Starts both paths
- Long path: No slack: fast cells
- Short path: Slack to spare: HVT
- Capture register: Samples both
- Launch register → Long path: many gates
- Long path → Capture register: just in time
- Launch register → Short path: few gates
- Short path → Capture register: early
Most paths of a real block are short. So most cells can take the low-leakage flavour.
This method is called multi-Vt design. Some flows start with the low-leakage flavour everywhere and swap cells to a faster flavour on the paths that fail timing. Others start fast and swap the cells on paths with slack to HVT. Both end with fast cells only where timing needs them.
The simulator can hide all of this. The plot below repeats the standard NMOS with the default gmin.
gmin left at its default.- standard
- LVT
- standard, default gmin
- Gate voltage 0 Vstandard = 215.5 fA · LVT = 2.904 pA · standard, default gmin = 2.016 pA
Off: the standard device carries 216 fA and the LVT device 2.90 pA. With the default
gminthe standard device reads 2.02 pA. - Gate voltage 1.8 Vstandard = 476.9 µA · LVT = 544.6 µA · standard, default gmin = 476.9 µA
On: 477 µA and 545 µA. The
gmincurve has joined the real one long before.
Read the plot as a table
| Gate voltage | standard | LVT | standard, default gmin |
|---|---|---|---|
| 0 V | 215.5 fA | 2.904 pA | 2.016 pA |
| 200 mV | 34.37 pA | 458.3 pA | 36.17 pA |
| 400 mV | 5.591 nA | 64.58 nA | 5.593 nA |
| 600 mV | 673.8 nA | 4.704 µA | 673.8 nA |
| 800 mV | 18.62 µA | 54.16 µA | 18.62 µA |
| 1 V | 85.23 µA | 142.7 µA | 85.23 µA |
| 1.2 V | 176.4 µA | 241 µA | 176.4 µA |
| 1.4 V | 274.7 µA | 342.4 µA | 274.7 µA |
| 1.6 V | 375.5 µA | 444.1 µA | 375.5 µA |
| 1.8 V | 476.9 µA | 544.6 µA | 476.9 µA |
Measured on ngspice 42. The default gmin adds about 1.8 pA, nine times this device's own leakage.
Asked in interviews
What is multi-Vt design?
The library offers each cell with several thresholds. A low threshold is fast and leaky. A high threshold is slow and leaks little. The flow keeps the fast flavour only on paths that would fail timing, and uses the low-leakage flavour everywhere else.
Why does leakage depend on the inputs of a gate?
The inputs decide which transistors are off and how they are connected. Two off transistors in series leak less than one: the stack effect. Measured here, one NAND2 leaks from 41.3 fA to 560 pA across its four states.
At which corner do you sign off leakage?
At the fast process corner, the highest supply and the hottest temperature. Leakage rises steeply with temperature: 41.1 times from 25 °C to 100 °C for the standard PMOS here.
Try It
Run the netlist. Vsg sweeps the gate drive of all three devices together. Each device has a 0 V source in its drain, and SPICE reports the current through that source.
Divide ioff_std by ioff_hvt, then ion_std by ion_hvt. The first ratio is the leakage you save. The second is the speed you pay.
Change .temp 25 to .temp 100 and run again. Predict first which ioff moves least. Then delete the .options gmin=1e-18 line and see which reading it spoils.
Run the netlist to draw its plots here.
ioff_std = 2.802369e-10 ion_std = 1.829590e-04 at= 1.800000e+00 ioff_hvt = 1.513651e-12 ion_hvt = 1.337345e-04 at= 1.800000e+00 ioff_lvt = 3.006908e-09 ion_lvt = 1.618616e-04 at= 1.800000e+00 vt_std = 5.200544e-01 vt_hvt = 7.694921e-01 vt_lvt = 2.153764e-01Check Yourself
Common Mistakes
Wrong Use the low-threshold flavour for every cell.
The LVT NMOS leaks 13.5 times more for 14 % more on current.
Fix Spend the fast flavour only on paths that fail timing without it.
Wrong Quote leakage measured at 25 °C.
The standard PMOS leaks 41.1 times more at 100 °C.
Fix Measure leakage at the hottest corner the product must work in, with the fast process and the highest supply.
Wrong Give one leakage number for a gate.
The NAND2 here leaks 41.3 fA in one input state and 560 pA in another.
Fix Ask which state the gate rests in, and choose that state when you can.
Wrong Carry these 130 nm numbers to a modern process.
Newer processes use lower thresholds and thinner gate oxides, which add gate leakage, and a chip holds far more transistors.
Fix Keep the ratios and the mechanisms, and read the absolute values from the library you use.
Remember
- A higher threshold cuts the off current hugely and the on current a little: 185 times against 27 % here.
- Leakage rises steeply with temperature: 41.1 times from 25 °C to 100 °C for the standard PMOS.
- One gate has several leakage currents. Off transistors in series leak least: the stack effect.
- Multi-Vt design keeps fast, leaky cells on the critical paths and low-leakage cells everywhere else.
- SKY130 leaks little. Take the ratios with you, not the picoamps.