Gallery
RF Devices in 3D
The RF devices of SG13G2 as real layouts: the PDK's own cells and the structures this course drew and simulated with an electromagnetic solver, extruded through the process's layer stack. Drag to rotate, scroll to zoom, and take them apart in the explorer.
- Recognize the layers an RF transistor, a MIM capacitor, a varactor, a spiral, a transformer, a transmission line and a probe pad are built from.
- Relate what you see to the RF numbers of the course: gate resistance per cell, emitter size, capacitor feed, inductor Q, line impedance, pad capacitance.
- Explain why RF passives live in the two thick top metals, high above the substrate, over a Metal1 ground.
- Regenerate every model, or turn your own GDS into a 3D model, with the scripts in
tools/3d/.
Contents
How to read these models
A layout is a stack of two-dimensional drawings, one per mask layer; a chip is three-dimensional. Every model on this page was made in the same three steps.
- Take a layout that the course uses, or draw the device with the PDK’s own generator. The spiral, the transformer and the microstrip are the structures Chapter 7 simulated with Palace; the HBT cell, the MIM capacitor and the GSG pads are those of Chapter 18; the matching network is that of Section 17.8. The rest, the metal stack, the RF MOSFET, the high-voltage HBT and the varactor, are drawn by IHP’s KLayout PyCells, the cells you place from KLayout’s library browser [191, 203]. Nothing is drawn by hand.
- Extrude every layer through the process stack. Each GDS layer is merged and lifted to its height: the heights are those of IHP’s EM layer stack, the one gds2palace used for every Palace run of the course [198, 197], which follows the layer thicknesses of the process specification [200] (Table 1).
- Write one 3D file per device in the glTF binary format [212], with one named material per layer, and show it with Google’s
<model-viewer>web component [213]; the explorer uses three.js [214].
Everything above the silicon is to scale, with no vertical exaggeration. The oxide is not drawn, so you see through it. Below the surface the drawing is schematic: the Activ slab stands for the 0.4 µm deep trench-isolation level, and wells, buried layers and the substrate are thin translucent plates. Two simplifications keep the files small: in the pads and in the matching network, every array of small vias is drawn as one solid block (a GSG pad group has about 7000 vias; its model is 20 kB instead of 1.2 MB), and slits in the Metal1 ground plane, drawn on the metal’s slit purpose, are left out, as gds2palace leaves them out (Section 18.3).
How to use the viewers:
- Rotate: drag with the mouse or one finger. Zoom: click into a viewer first, then use the scroll wheel (so that scrolling down the page never gets stuck in a model), or pinch. Pan: right-drag or drag with two fingers.
- Layers: click a layer in the legend under a viewer to hide or show it; shift-click (or alt-click) shows that layer alone. Hide TopMetal2 and the ground plane to look inside a structure.
- Reset view returns to the starting camera. A model loads only when it scrolls into view.
- The explorer in the next section shows the same models with an explode slider, a vertical scale, and top and side views.
Table 1. The SG13G2 layer stack as drawn: z in µm above the silicon, with the 0.4 µm trench-isolation level at the bottom, as in IHP's EM stack and Figure 1.5. Sheet resistances from Appendix C.
| Layer | z, bottom to top (µm) | Thickness | Sheet resistance | Used for |
|---|---|---|---|---|
| Activ (trench-isolation level), GatPoly | 0–0.40, 0.40–0.56 | 0.40, 0.16 µm | transistors, varactor, substrate rings | |
| Metal1 | 1.04–1.46 | 0.42 µm | 110 mΩ/□ | ground planes and shields |
| Metal2 to Metal5 | 2.00–2.49 … 5.09–5.58 | 0.49 µm | 88 mΩ/□ | wiring; Metal5 is the MIM bottom plate |
| MIM top plate | 5.60–5.75 | 0.15 µm | MIM capacitor, contacted by Vmim from TopMetal1 | |
| TopMetal1 | 6.43–8.43 | 2.0 µm | 18 mΩ/□ | inductor crossings, MIM top-plate feed, transformer secondary |
| TopVia2 | 8.43–11.23 | 2.8 µm | between the two thick metals | |
| TopMetal2 | 11.23–14.23 | 3.0 µm | 11 mΩ/□ | spirals, transmission lines, pads |
| Passivation | 14.23–16.13 | 1.5 + 0.4 µm | oxide and nitride, opened over the pads |
Explore
The explorer takes the devices of this page apart. Pull the layer levels apart with Explode, stretch the heights with Vertical scale (the large passives are hundreds of micrometres wide and only 14 µm tall), look straight down with Top (the layout as you draw it) or from the front with Side (the stack in cross-section), and switch layers on and off in the panel, which also lists the height of each one. Every pair of viewers further down has a button that opens its device here.
The 3D view loads when you scroll to it (it needs JavaScript and WebGL).
Drag to rotate; right-drag or two fingers to pan; click into the view, then scroll or pinch to zoom. In the layer panel, shift-click (or alt-click) a layer to show it alone.
Try it · Five things to look for
- The full metal stack, Side. Read the heights in the layer panel: the bottom of TopMetal2 is 9.77 µm above the top of Metal1. That gap is the dielectric of every microstrip and the reason a pad over a Metal1 shield has 23.8 fF of parallel-plate capacitance (Section 18.4).
- RF MOSFET cell, hide Metal2 and Metal1. Find the four poly fingers and the gate ring that contacts each of them at both ends. The fixed 40 Ω of
rfmode=1belongs to this whole cell, which is why the LNA uses sixteen of them.- The two HBTs, Top. Compare the emitter stripes: 0.07 µm × 0.9 µm in
npn13G2, 0.12 µm × 2.5 µm innpn13G2v. The narrow, short emitter is the fast one; the wider one survives a higher collector voltage.- RF MIM capacitor, Side at 10× vertical scale. Follow the signal from the left TopMetal2 line down TopVia2 to the TopMetal1 feed, through Vmim and the MIM plate to the Metal5 bottom plate, and back up TopVia1 and TopVia2 to the right line. The landings are part of the capacitor at 50 GHz.
- Spiral L3 and transformer T1, Explode. In the spiral, TopMetal1 appears only at the crossings; in the transformer it is a whole second winding, one TopVia2 height (2.8 µm) below the first. That small gap is why T1 couples with k = 0.72.
The full metal stack
Start with the back end of line, which Figure 1.5 drew in cross-section: a contact on the active area and a tower of vias up to TopMetal2, drawn by the PyCell via_stack with 2 × 2 via arrays up to Metal5 and single TopVia1 and TopVia2 cuts above.
Footprint 2 × 2 µm, 14.2 µm to the top of TopMetal2. The two thick top metals take more than a third of the height.
Five thin metals and two thick ones. Metal1 to Metal5 are 0.42–0.49 µm of aluminium with 88–110 mΩ/□; TopMetal1 and TopMetal2 are 2 and 3 µm thick with 18 and 11 mΩ/□ [200]. For an RF designer the stack has three jobs:
- Low-loss conductors. A 10 µm wide, 1 mm long TopMetal2 line has 1.1 Ω at DC. Above about 0.93 GHz the skin depth in TopMetal2 is smaller than its 3 µm thickness, so at the course radio’s 2.44 GHz the current already crowds towards the surfaces (Section 7.4).
- Distance from the substrate. TopMetal2 sits 11.23 µm above the silicon, so a spiral on it couples weakly to the lossy 50 Ω·cm substrate.
- A ground plane. Metal1, 9.77 µm below TopMetal2, is the ground of every microstrip, pad shield and matching network in this course (Section 18.1).
The RF MOSFET cell
The PDK’s rfnmos cell draws the structure IHP measured to fit its RF model: four 2 µm fingers (W = 8 µm, L = 0.13 µm) with a gate ring that contacts every poly finger at both ends, and a p+ guard ring around the cell. This is one of the sixteen cells of the Chapter 9 LNA’s 128 µm input device, as Figure 18.1b drew them.
W = 8 µm (4 fingers of 2 µm), L = 0.13 µm, 4.28 × 4.88 µm with gate ring and guard ring. Hide Metal2 and Metal1 to see the fingers.
What matters for RF here is the gate resistance. With rfmode=1, SG13G2’s PSP model adds a fixed RGO = 40 Ω per cell, which stands for the gate ring and contacts you see, on top of the finger terms of Eq. (6.6); it divides by the number of cells \(m\), not by the number of fingers [205]. The LNA’s device as sixteen of these cells has a model gate resistance of 3.36 Ω; as one cell of 64 fingers it would be above 40 Ω (Section 6.2, Section 18.2). The same mode adds the non-quasi-static channel, a substrate network and the gate’s end-cap capacitance; without it, the reference device of Chapter 6 shows fmax = 216 GHz instead of 89.0 GHz. What the model cannot know is the wiring between the cells: sixteen of them in a 4 × 4 array occupy 31.3 µm × 25.9 µm, and a 2 µm Metal2 gate bus across them, fed from one end, adds 0.44 Ω (Worked example 18.1).
SiGe HBTs: npn13G2 and npn13G2v
SG13G2 is a BiCMOS process, and its heterojunction bipolar transistors are the reason the course can build a 50 GHz amplifier. On the left, the npn13G2 cell with four emitters that Chapter 18 drew (Figure 18.1c) and the Chapter 17 amplifier uses four times in parallel; on the right, the high-voltage npn13G2v as one of the four instances of each transistor of the Chapter 16 PA.
Four emitters of 0.07 × 0.9 µm, 12.2 × 7.11 µm. Hide Metal2 and Metal1 to see the pink emitter windows.
Four emitters of 0.12 × 2.5 µm, 4.8 µm² for the four instances of one PA transistor; cell 14.8 × 8.7 µm.
The models show only the drawn layers. What makes the HBT fast, the thin graded SiGe base, the polysilicon emitter that fills the emitter window, the buried layer and the deep trenches (Figure 6.6), is formed by the process from these drawings, so the pink stripes stand in for the emitters. The two cells trade speed for voltage [200]:
- npn13G2: fT = 353 GHz at 42.8 mA/µm² and fmax = 456 GHz in simulation (Nx = 4, VCE = 1.2 V), but BVCEO = 1.6 V (Section 6.4). It is the device of the 50 GHz amplifier.
- npn13G2v: fT = 104 GHz and fmax = 265 GHz at only 3.51 mA/µm², with BVCEO = 2.5 V, which lets the course PA’s cascode run from 2.0 V with a load near 50 Ω.
The layout decides two things the model cannot. Parallel cells share their current only if they share their temperature: in Chapter 18’s ngspice check, one cell 10 K warmer than the other three carries 7.48 mA against 6.67 mA. And at 50 GHz each cell’s emitter must reach the ground plane through the shortest possible via stack, because 3 pH of emitter inductance, 10 µm of wire, costs about a decibel of gain (Section 18.2).
RF MIM capacitor
A MIM capacitor is a thin top plate over Metal5 with a thin, high-permittivity dielectric in between, 1.5 fF/µm² (Section 7.1). The model is the middle of Chapter 18’s MIM10 test structure: a 10 µm rfcmim cell between two 16 µm TopMetal2 lines, over the Metal1 ground plane.
The middle 66 × 44 µm of MIM10: a 10 × 10 µm plate with 2 µm feeds and its two 8 × 16 µm landings. The substrate-ring contacts are drawn as one band. Hide TopMetal2 and TopMetal1 to see the plate.
Everything cap_rfcmim models is in the cell (Figure 7.1): the MIM plate over the Metal5 bottom plate, the Vmim array up to the TopMetal1 top-plate connection, a feed of width wfeed on each side (TopMetal1 on one, Metal5 on the other), and a ring of substrate contacts around the bottom plate for the model’s substrate network. The feed is a model parameter: its resistance is 0.96 Ω·µm / wfeed, and a 30 µm capacitor has Q = 35.6 at 4.88 GHz with the default 2 µm feed against 87.3 with a 10 µm feed. The plate size matters as much: this 10 µm plate is 0.152 pF with Q = 373 at 4.88 GHz and a self-resonance above 100 GHz, while a 75 µm plate resonates at 7.94 GHz (Table 7.1). The landings, which the model does not contain, connect a TopMetal2 line to the top plate through TopVia2 and to the bottom plate through TopVia1, TopMetal1 and TopVia2; at 50 GHz, where this capacitor is only −j16.7 Ω, their few picohenries move its apparent value (Section 18.3).
Varactor
The accumulation-mode MOS varactor sg13_hv_svaricap tunes the course VCO. It is built with the 3.3 V gate oxide (the orange marker sheet): n+ gates over an n-well, so the surface accumulates or depletes but never inverts, and its capacitance changes monotonically with the gate-to-well voltage. Its two gates, G1 and G2, are the two interleaved poly combs you see, and they share one well, so the device is naturally a differential capacitor between G1 and G2, tuned by the well voltage (Section 7.6).
W = 9.74 µm, L = 0.3 µm, Nx = 10: the varactor of the course VCO, 22.3 × 12.4 µm. Hide Metal1 and Cont to see the two gate combs.
This is the device of Chapter 11’s VCO, between its two outputs with the well on Vtune. Its small-signal capacitance falls from 87.3 to 59.2 fF as Vtune goes from 0 to 1.2 V, a steepest slope of 39.7 fF/V that predicts a peak KVCO of 180 MHz/V (Worked example 11.2), and in the tank it has Q = 98.3, far above the spiral’s. The gate length sets the trade: with L = 0.3 µm the tuning ratio over ±1.5 V is 1.99 and Q ≥ 89 at 4.88 GHz; with L = 0.8 µm the ratio is 2.34 but Q drops to as little as 37.9, because the AC current crosses a longer, more resistive path under each gate (Figure 7.8).
Spiral inductor
The inductor of the course VCO’s tank: L3 from Chapter 7’s family, a symmetric octagonal spiral of three turns on TopMetal2, with TopMetal1 crossings where the turns swap sides at the top and the bottom, and both terminals at the bottom of the outer turn. Around it lies the 20 µm wide Metal1 frame of the EM simulation, half a diameter away.
dout = 170 µm, w = 8 µm, s = 2.1 µm, as simulated in Palace; frame 380 µm across. Zoom in on the crossings.
From the Palace results of Table 7.2: L = 1.92 nH, Q = 10.6 at 2.44 GHz and 16 at 4.88 GHz (17.7 driven differentially), a peak Q of 17.1 at 6.75 GHz, and a self-resonance at 16.9 GHz. In the VCO tank it has Q = 17.8 and causes 84 % of the tank’s loss (Section 11.2): the spiral sets the phase noise. Look at what the model shows about its losses. The track is on the thickest, highest metal, 11.23 µm above the substrate; the crossings drop to the thinner TopMetal1 and add resistance; and nothing else lies inside the frame, because an inductor’s field reaches about its own diameter. The PDK’s own inductor2 and inductor3 cells draw the same shape and add marker layers that keep fill and substrate contacts out of that area (Section 18.3).
Stacked transformer
Transformer T1 of Section 7.5 is the simplest stacked one: a single 200 µm octagonal turn on TopMetal2 over an identical turn on TopMetal1, with the feeds of the primary at the bottom and of the secondary at the top.
Primary on TopMetal2, secondary on TopMetal1, 2.8 µm apart; frame 440 µm across. Hide TopMetal2 to see the secondary.
Stacking puts the two windings one TopVia2 height apart, so most of the flux of one links the other: L1 = 0.437 nH, L2 = 0.441 nH and k = 0.72. The price is the capacitance between the windings, and the thinner TopMetal1 gives the secondary a lower Q, 9.64 against 13.4 at 4.88 GHz. The minimum insertion loss is 1.78 dB at 2.44 GHz but only 0.372 dB near 38 GHz, below the self-resonance at 42.2 GHz: one turn is too little inductance for 2.4 GHz and just right for millimetre waves (Figure 7.9).
TopMetal2 microstrip
The transmission line of the course: a 16 µm TopMetal2 strip over a Metal1 ground, here the 300 µm line of Section 7.7 over its 80 µm wide ground.
16 µm × 300 µm TopMetal2 strip, 80 µm wide Metal1 ground 9.77 µm below. Side view: the field lives in the oxide between them.
The ground plane shields the lossy substrate, so the field is mostly in the oxide and the loss mostly in the metals. From two lines of 300 and 900 µm, Palace gives Z0 = 50.7 Ω at 4.88 GHz and 47.8 Ω at 50 GHz, εeff = 3.73, and a loss of 0.283, 0.325 and 0.583 dB/mm at 2.44, 4.88 and 50 GHz. A quarter wavelength at 50 GHz is 776 µm. Seen as a wire, the same line is 0.308 pH/µm and 0.135 fF/µm (Eq. 18.1): ten micrometres are 3.08 pH, which is why every connection of the 50 GHz amplifier is drawn as this line (Section 18.1).
GSG probe pads
A millimetre-wave chip is measured with ground–signal–ground probes. Chapter 18’s pad group has an 80 µm TopMetal2-only signal pad (the PDK’s bondpad cell without its metal stack) between two ground pads at a 150 µm pitch, which are stacked down to a Metal1 shield under the whole group.
Signal pad on TopMetal2 only; ground pads stacked from TopMetal2 to Metal1; via arrays drawn as solid blocks; passivation opened over each pad. 120 × 420 µm.
The signal pad and the shield 9.77 µm below it form a capacitor that becomes part of the matching network. The parallel-plate estimate is 23.8 fF; Palace gives 37.2 fF at 2.44 GHz and 37.7 fF at 50 GHz, 56 % more, because of the fringing field to the shield and the ground pads (Table 18.1). Chapters 9 and 17 had assumed 30 fF; with the real pad the LNA’s worst-case S11 moves from −15.4 to −13.5 dB, still within its specification. The PDK’s default bondpad stacks Metal3 to TopMetal2 for wire bonding, and its Metal3 plate, only 1.57 µm above the shield, would have about 123 fF: for RF, use a top-metal pad over a shield, as small as the probe allows.
Showcase: a 50 GHz matching network
Everything above, put together: the input network of the redesigned 50 GHz amplifier of Section 17.8. From the GSG pads on the left, a DC block (a 20 µm cap_rfcmim with 16 µm feeds), a T-junction, the series line ℓi = 107 µm towards the transistor’s base on the right, and the stub ℓsi = 117 µm down to a 30 µm decoupling capacitor stacked to the ground plane, with its 8 µm bias feed.
330 × 420 µm, every line the 16 µm TopMetal2 microstrip over Metal1; via arrays drawn as solid blocks. The transistors (not drawn) sit at the right end of the series line.
Each piece of this network was simulated in Palace on its own (the pad, the DC block, the T-junction, the decoupled stub end), the line lengths were designed on those S-parameters, and the finished network was simulated once more, in 56 minutes. With both networks from the full EM simulation, the amplifier keeps 12.4 dB of gain at 50 GHz, with S11 = −14.0 dB and S22 = −30.6 dB: the layout costs 0.7 dB, where the first design of Chapter 17 lost 4.3 dB to its layout (Section 18.6). Hide TopMetal2 and the passivation to find the two capacitors; one of them, the 30 µm decoupling capacitor, replaced a 20 µm one that resonated with its own feed near 50 GHz.
Make your own 3D models
Everything on this page is regenerated by one script, which runs in IIC-OSIC-TOOLS [193] (or starts the container with Docker if you call it from outside):
tools/3d/make_devices.sh # PyCell GDS -> tools/3d/gds/, 3D models -> assets/3d/ (well under a minute)
It draws the metal stack, the RF MOSFET, the npn13G2v and the varactor with the SG13G2 PyCells (tools/3d/gen_sg13g2.py), takes the other structures from the course’s own layouts (tools/em/ch07_inductors/gds/, tools/em/ch17_pa/gds/, tools/em/ch18_pa/gds/), and converts each GDS with tools/3d/gds2glb.py. The converter works on any SG13G2 layout, for example the whole Chapter 18 amplifier:
python3 tools/3d/gds2glb.py tools/em/ch18_pa/gds/PA.gds pa.glb --merge-cuts 0.6 # via arrays as blocks
python3 tools/3d/gds2glb.py my_lna.gds core.glb --clip 0 0 40 30 --skip Substrate # a window of a layout
python3 tools/3d/preview.py pa.glb pa.png --orbit 35 60 # PNG check, no browser
Options: --cell (which cell), --zscale (vertical exaggeration, default 1), --clip X0 Y0 X1 Y1 (a window in µm), --only and --skip (layers by legend name), --merge-cuts G (draw arrays of vias closer than 2G µm as one block) with --keep-cuts (via layers to leave as single cuts), and --margin (substrate margin). The layer stack lives in tools/3d/stacks.py, with each layer’s GDS number, a Boolean expression, its bottom and top height, colour and opacity. Open the .glb in any glTF viewer, in Blender, or on a page of your own with <model-viewer>.
Try it · Your own structure in 3D
Draw a spiral of your own with Chapter 7’s
tools/em/ch07_inductors/ch07_geom.py(the second experiment of Chapter 7’s Try-it box) or a pad group withtools/em/ch18_pa/ch18_layout.py, convert it withgds2glb.pyand look at it from below TopMetal2: it is the quickest way to spot a crossing without its via, or a ground pad whose stack stops short of the shield.
- SG13G2’s RF devices are built on its back end: two thick top metals (2 and 3 µm, 18 and 11 mΩ/□) high above the substrate for spirals, lines and pads, and Metal1, 9.77 µm below TopMetal2, as the ground plane.
- An RF MOSFET is a cell with a gate ring and a guard ring;
rfmode=1charges 40 Ω per cell, so wide devices are built from many cells.- The HBT’s emitter stripes (0.07 µm wide in npn13G2, 0.12 µm in npn13G2v) trade speed for breakdown voltage: 353 against 104 GHz of fT, 1.6 against 2.5 V of BVCEO.
- A MIM capacitor is the plate plus its feeds and landings; the feed width and the plate size set its Q, the landings its value at 50 GHz.
- Spirals, transformers, lines and pads are large flat structures whose RF behaviour (L, Q, k, Z0, Cpad) only an EM simulation of the drawn layout gives.
Further reading
- The SG13G2 process specification [200] gives the layer thicknesses and device parameters, and IHP’s EM workflows [198] contain the layer stack used for the heights here and for every Palace run of the course.
- KLayout [203] shows a layout in 3D inside the editor with its 2.5D view; the PDK ships the script for SG13G2 (
libs.tech/klayout/tech/d25/sg13g2_beol.lyd25). - The sibling course’s gallery, PDK Devices in 3D in The Art of Analog IC Design, shows the analog devices of SG13G2 (core MOSFETs, resistors, the substrate PNP) beside those of SKY130, with the same tools.