Tutorial 1: Chip 12

examples/tutorials/01_chip12_coupled_chip.ipynb rebuilds a legacy nine-qubit test chip, “chip 12”, in the Quantcore PDK, step by step, and runs the whole loop on it. The chip is an internal Quantum Metal design; its positions and options are in examples/designs/legacy/_legacy_b.py and its spec is Chip 12. The design notebook builds the same chip in one cell; this tutorial builds it in small steps. It is the first tutorial where the layout comes from a source and the PDK rules change it.

../_images/t01_chip.png

Section 6: all nine qubits, twelve launchpads, readouts and control lines, with device ID D12.

What the chip is

A 7.5 × 7.5 mm coupled-qubit test chip:

  • one 20/12 µm feedline across the die centre with nine λ/4 readouts, 4.3 to 6.3 mm long;

  • nine TransmonCross Xmons (metal: wrapper);

  • Q1-Q2, a direct pair: their arm tips face each other across a 20 µm ground strip;

  • Q5-Q8-Q6, a tunable-coupler triple: Q8 is a flux-tunable Xmon between Q5 and Q6, with its own readout;

  • five flux lines and five drive lines, each to its own launchpad (ten line pads and two feed pads).

The PDK wins wherever it has a rule, so the rebuild differs from the source:

From the source

What the PDK changes

Pads

Legacy launchpad positions

Outer etch edge 100 µm from the die edge

Junctions

External junction cells, no values

Trilayer Quantcore Junction / Quantcore Squid at the PDK nominal R_n 45 kΩ (L_J 6.73 nH)

Cheese

None

The PDK band around the 3/0 keep-outs

Readouts

Lengths and couplers

Couplers pinned with near= so they stay put when pads move

Values

No qubit targets

Out-of-margin values are flagged (DRC design warnings), never retuned

The sections

Section

What it does

What comes out

1 PDK facts

The PDK rules this chip uses

A table of the rules and what each is used for

2 Die and frame

Designer("t01_chip12", box=..., chip={"name_chip": "D12"}); at(x, y) moves a legacy coordinate into the die frame

The empty die with its frame

3 Pads

Two feed launchpads and ten line pads, edge="auto"

A table: every pad’s outer etch edge is 100.0 µm from the die edge

4 Feedline

d.route("cpw_main", p1.tie, p2.tie, a=20, b=12)

A 5380 µm feedline; it runs right to left, so its “right” side is the top half of the chip, which decides the hanger side

5 Components

Nine Xmons with junctions; d.couple(...) for the three couplings no route carries

Junction L_J 6.7344 nH, C_J 15 fF; arm-to-arm estimate C_arm = 1.67 fF

6 Routes

Nine λ/4 readouts through fl.hanger(near=...), ten flux and drive lines

A table of drawn against legacy lengths (all equal, R0 5493 µm to R8 4300 µm) and the full chip picture

7 DRC

Errors, warnings and PDK design warnings

0 errors, 30 warnings (all dangling_port: bond pads and junction attachment points), 0 design warnings

8 Netlist and LOM

d.lom() turns components and routes into a lumped circuit

24 nodes, 74 branches; a small netlist of Q1, Q2, R1, R2

9 LOM simulation

Each qubit with its readout, the pair, the triple

Tables of f_q, α, f_r, κ, χ; J and ZZ of Q1-Q2; three hybridised modes of Q5-Q8-Q6

10 FEM stage 1

Sim groups and the subset check

Convergence, C-matrices, eigenmodes, FEM-filled capacitances

11 FEM stage 2

The subset values in the whole netlist, then feedback

LOM versus FEM, J on FEM values, a revision with new readout lengths

12 Iterate

Claw 150 to 180 µm on a copy of the design

A before/after table, design hash and DRC

13 Build and export

export_fab with and without cheese and rounding; CLI equivalents

Fab layer table

14 Snapshots

The pair, a SQUID, a readout short, the fab view, the legacy screenshot

Pictures

15 Summary

Matches and differences against the source

A table

3 to 6: pads, feedline, qubits, routes

Twelve pads are added in one with d.batch(): block, so the design is rebuilt once instead of twelve times. Every pad is asked at its legacy position and slid to the 100 µm rule. The feed pads are 400 × 400 µm (20/12 µm trace); the ten line pads are 400 × 300 µm (10/6 µm trace).

The Xmons have a 220 × 50 µm cross in a 20 µm gap and a 150 µm readout claw (6 µm gap). Q1-Q2, Q5-Q8 and Q8-Q6 couple through their facing arm tips; no route joins them, so d.couple(...) declares each with a coupled-strip estimate of 1.67 fF. The Q1_Q2 FEM group gives the real value.

Readout lengths came out equal to the legacy ones. The flux and drive lines run through the legacy corner points as waypoints; the pads moved to the edge, so the first straight of each line is longer.

7 to 9: DRC, netlist and the LOM

The legacy readout lengths were drawn for a bare λ/4 (no claw), so each legacy length’s bare frequency is the readout target. Each qubit with its readout, the feedline acting as a load:

Qubit

f_q (GHz)

α (MHz)

Readout f_r, LOM (target) GHz

Q0

5.0591

-196.9

4.9357 (5.7361)

Q1

5.0411

-146

4.4242 (5.0317)

Q2

5.0401

-145.2

4.5015 (5.1317)

Q3

5.0325

-146

5.7585 (6.7936)

Q4

5.0331

-146.4

5.937 (7.041)

Q5

5.0248

-135.7

5.226 (6.0594)

Q6

5.0277

-139.6

5.283 (6.1396)

Q7

5.046

-124.6

4.8085 (5.5375)

Q8

5.0352

-146.8

6.1513 (7.3276)

Every Xmon lands near 5.0 GHz (inside the PDK band) with α about -145 MHz. The claw adds about 125 fF at each readout’s open end, so the loaded readouts sit 0.6 to 1.2 GHz below their bare λ/4. Q0 and R0 come out nearly degenerate and hybridise (χ = 64.7 MHz), and Q7 and R7 are close (χ = -22.7 MHz). The notebook does not retune them: that is a design decision.

The direct pair Q1-Q2, with the same junction on both, shows two hybridised modes at 5.1220 and 5.1828 GHz and J = 29.5 MHz. That J comes from the rough 1.7 fF arm estimate. With Q2 detuned (L_J 7.4 nH) the static ZZ is -1.12e+04 kHz, in the MHz range: the reason Q8 sits between Q5 and Q6 as a tunable coupler. The triple at zero flux (all three resonant) gives three hybridised modes at 5.0932, 5.1529 and 5.1836 GHz.

10 and 11: the FEM stages

The sim groups cut the chip into jobs a workstation can solve:

Group

Analysis

Box

Gives

Q1_R1 and one Qn_Rn per qubit

q3d_cmatrix

The qubit and 300 µm; the readout is cut at the claw

C_island, C_g, C_claw

Q1_q

eigenmode_epr

Q1 and 150 µm; the readout and flux lines leave the box as 50 Ω lines

The qubit’s own mode: f_q, α, p_J

Q1_R1e

eigenmode_epr

Q1, its whole readout and 100 µm, with the feedline cut at both ends

f_q, f_r, χ, κ

Q1_Q2

q3d_cmatrix

Both qubits and 300 µm

The arm-to-arm capacitance, hence J

Q5_Q8_Q6

q3d_cmatrix

The triple and 300 µm

The tunable-coupler capacitances

The eigenmode job of the whole Q1_R1 box (a 1.3 × 2.9 mm box) did not finish within 20 minutes on the workstation, which is why the notebook splits it into Q1_q and Q1_R1e.

The kept results (fast settings):

Job

Passes

Last Δ / target

Converged

Wall time

Q1_R1 Q3D

8

1.24 / 2.0 %

no

60.6 s

Q1_Q2 Q3D

8

1.33 / 2.0 %

no

41.1 s

Q1_q eigenmode

3

0.591 / 1.0 %

yes

123.6 s

Q1_R1e eigenmode

3

1.67 / 1.0 %

no

276.8 s

The Q1_R1 C-matrix gives C_g 10.883 fF against the 5.2 fF estimate and C_claw 111.388 fF. The eigenmode of Q1 alone gives a dressed f_q of 4.7564 GHz and α of -126.9 MHz (p_J 0.9876). Q1_R1e gives f_q 4.7212 GHz and a cross-Kerr χ of -26.451 MHz.

Stage 2 fills the branches the Q3D jobs simulated (9 branches) into the whole netlist; 40 branches keep their estimates. The LOM then runs pair by pair, because QuCat on all nine qubits at once is far too slow. For Q1 and Q2 the LOM on the subset Q3D moves f_q by +1.6 % and +3.1 %, and the readout χ from -1.644 to -9.158 MHz (Q1) and from -2.369 to -10.86 MHz (Q2). Against pyEPR on Q1_q the filled LOM f_q is 7.2 % high (5.123 against 4.756 GHz). For the pair, the FEM arm-to-arm capacitance is 1.153 fF (against 1.67 fF) and J drops from 29.53 to 21.91 MHz.

../_images/t01_pair.png

Section 14: the direct-coupled pair Q1-Q2 with their readouts and control lines.

Feedback moves each readout’s length: the λ/4 scales with f / target. For example R0 goes from 5493 to 4727 µm and R8 from 4300 to 3610 µm. The result is a revision examples/tutorials/designs/t01_chip12_fb.json (DRC clean), exported to a fab file. The consensus on the whole netlist reports 0 of 9 targets within tolerance: the revision is the first step, and the loop closes by running stage 2 on it.

12: iterate

Lengthening Q1’s claw from 150 to 180 µm on a copy of the design (Designer.load):

Quantity

Before

After

Change

Q1 f_q (GHz)

5.0411

5.0264

-0.3 %

Q1 α (MHz)

-146.02

-144.81

+0.8 %

R1 f_r (GHz)

4.4242

4.3754

-1.1 %

R1 χ (MHz)

-1.6439

-1.883

-14.5 %

R1 κ (MHz)

0.15343

0.14668

-4.4 %

The design hash changes from b08ee7bf... to 818a245f... and the DRC stays clean. One parameter moved; the layout, DRC, netlist and LOM followed.

../_images/t01_claw_before1.png

Before: Q1’s readout claw at 150 µm.

../_images/t01_claw_after1.png

After: the claw at 180 µm.

13: export

export_fab keeps the build’s cheese (14723 holes) and rounds corners. With cheese=False, rounding=False it writes 0 holes and 42 M1 polygons. The CLI equivalents are in the notebook: qcc build examples/tutorials/designs/t01_chip12.json --show, qcc design drc ..., qcc circuit export ..., qcc circuit simulate ....

Pitfalls

  • Hangers by arclength (s=) slide when a pad moves; use near= as the notebook does.

  • squid_a and squid_b are junction attachment points, not route ends. They show as dangling_port warnings. That is expected.

  • A SQUID gets the same total L_J as a single junction, so each of its two junctions has twice the R_n. The PDK window judges the device’s parallel R_n.

  • At a 200 µm claw the claw would reach the cross’s side arms, and the DRC flags it as pdk_claw. The notebook stops at 180 µm.

  • The FEM results of sections 10 and 11 belong to the original geometry. After section 12 changes the design hash they are stale: re-run them to refresh.

  • The Q3D jobs are fast-setting results, not converged (Δ 1.2 to 1.3 % against 2 %).

  • FEM_WHOLE = True also solves the whole chip as one Q3D (chip_q3d_cmatrix is kept). Whole-chip eigenmode or driven runs take hours and over 30 GB.

Tutorial 2 uses a chip with no public layout, built from device data.