TrapSIMD
J. Ruan, H. Zhang, X. Fang, A. Li et al. TrapSIMD: SIMD-Aware Compiler Optimization for 2D Trapped-Ion Quantum Machines. arXiv 2504.17886 (2025). arXiv:2504.17886 · their code: not published
A grid of X-junctions with a linear trap on every edge, one ion per position. Transport between traps is broadcast: one waveform moves every participating junction the same way at once (a JT-SIMD instruction), the paper-side twin of our own one-waveform rule.
- Rules R4, R4b, R6, R6b, R7b cite it as a source.
- The movement classes of JT-SIMD (12 shifts and 6 swaps on an X-junction grid): our simd_classes control model, with one class per cycle and variable participation.
- Event-scheduled instructions without barriers, and an objective that adds −ln F over the operations.
- The default device's gate and measurement times (25 µs two-qubit, 5 µs one-qubit, 120 µs measurement) and its 600 s coherence time, from its Table 1 and Sec. 6.1. Our device files credit these to H2 and to Jones and Murali, which is wrong; those files are pinned by the released boards, so the next physics release corrects them.
No code, schedules or circuits are published. The two worked examples are specified to the microsecond and are replayed exactly; Table 3 is compared at model level, on circuits we rebuilt from the paper's description.
The paper works two examples out to the microsecond (Figs. 6 and 7). Each is written here event for event and replayed under the paper's rules:
| worked example | the paper | our replay | checks under the paper's rules |
|---|---|---|---|
| Fig. 6(b2), as drawn ▶ watch it in the Studio | 982 µs | 982 µs equal | all 13 checks pass |
| Fig. 6(c2), as drawn ▶ watch it in the Studio | 732 µs | 732 µs equal | all 13 checks pass |
| Fig. 7(b2), as drawn ▶ watch it in the Studio | 823 µs | 823 µs equal | all 13 checks pass |
| Fig. 7(c1), as drawn ▶ watch it in the Studio | 545 µs | 545 µs equal | all 13 checks pass |
All four add up only with a two-qubit gate of 141 µs: the 25 µs of the paper's Table 1 plus a 58 µs shift into the gate zone and one out, which the paper says the gate's latency includes. With 25 µs, every two-qubit gate of all four fails the duration check.
Made by schedule_ours, a scheduler we wrote for the paper's model; it is not the compiler that compiles this website's leaderboard entries.
The paper's results table (Table 3) gives, for four benchmarks on four grid sizes and two trap lengths, the time of its compiler's schedule. With no code or circuits published, this is a model-level comparison: the device, the timing table and the rules are the paper's; the circuits are rebuilt from its description, and the schedules are ours, each checked under the paper's rules (broadcast transport, one waveform at a time, nothing else running during an inter-trap cycle) with strict durations.
The paper's main configuration, A-60 (a 4×4 grid of 40 traps):
| benchmark | the paper (its compiler) | ours | fidelity: paper | ours | checks | |
|---|---|---|---|---|---|---|
| BV-60, L=8 ▶ watch ours | 23,625 µs | 13,861 µs | 1.70× faster | 0.92 | 0.8807 | all 13 checks pass |
| QAOA-60, L=8 ▶ watch ours | 744,943 µs | 143,257 µs | 5.20× faster | 0.01 | 0.01047 | all 13 checks pass |
| RCA-60, L=8 circuit differs: 465 two-qubit gates, the paper implies 857 ▶ watch ours | 278,970 µs | 208,576 µs | 1.34× faster | 0.14 | 0.2452 | all 13 checks pass |
| VQE-60, L=8 ▶ watch ours | 1,255,877 µs | 143,257 µs | 8.77× faster | 0.00 | 0.01047 | all 13 checks pass |
| BV-60, L=14 ▶ watch ours | 28,741 µs | 16,153 µs | 1.78× faster | 0.91 | 0.8766 | all 13 checks pass |
| QAOA-60, L=14 ▶ watch ours | 822,641 µs | 169,705 µs | 4.85× faster | 0.01 | 0.003355 | all 13 checks pass |
| RCA-60, L=14 circuit differs: 465 two-qubit gates, the paper implies 857 ▶ watch ours | 296,592 µs | 236,323 µs | 1.26× faster | 0.14 | 0.194 | all 13 checks pass |
| VQE-60, L=14 ▶ watch ours | 1,551,403 µs | 169,705 µs | 9.14× faster | 0.00 | 0.003355 | all 13 checks pass |
the other grid sizes (A-20, A-40, A-100)
| benchmark | the paper (its compiler) | ours | fidelity: paper | ours | checks | |
|---|---|---|---|---|---|---|
| BV-20, L=8 | 8,373 µs | 4,145 µs | 2.02× faster | 0.97 | 0.9614 | all 13 checks pass |
| QAOA-20, L=8 | 93,845 µs | 32,697 µs | 2.87× faster | 0.61 | 0.6142 | all 13 checks pass |
| RCA-20, L=8 circuit differs: 145 two-qubit gates, the paper implies 257 | 70,624 µs | 62,686 µs | 1.13× faster | 0.57 | 0.6508 | all 13 checks pass |
| VQE-20, L=8 | 130,538 µs | 32,697 µs | 3.99× faster | 0.56 | 0.6142 | all 13 checks pass |
| BV-40, L=8 | 16,175 µs | 9,003 µs | 1.80× faster | 0.94 | 0.9203 | all 13 checks pass |
| QAOA-40, L=8 | 327,601 µs | 90,977 µs | 3.60× faster | 0.13 | 0.1343 | all 13 checks pass |
| RCA-40, L=8 circuit differs: 305 two-qubit gates, the paper implies 557 | 180,950 µs | 148,791 µs | 1.22× faster | 0.28 | 0.3969 | all 13 checks pass |
| VQE-40, L=8 | 532,440 µs | 90,977 µs | 5.85× faster | 0.09 | 0.1343 | all 13 checks pass |
| BV-100, L=8 | 38,679 µs | 23,577 µs | 1.64× faster | 0.86 | 0.8057 | all 13 checks pass |
| QAOA-100, L=8 | 2,256,313 µs | 277,950 µs | 8.12× faster | 0.00 | 3.042e-06 | all 13 checks pass |
| RCA-100, L=8 circuit differs: 785 two-qubit gates, the paper implies 1457 | 561,586 µs | 354,216 µs | 1.59× faster | 0.03 | 0.09061 | all 13 checks pass |
| VQE-100, L=8 | 3,570,854 µs | 277,950 µs | 12.85× faster | 0.00 | 3.042e-06 | all 13 checks pass |
| BV-20, L=14 | 11,745 µs | 4,793 µs | 2.45× faster | 0.97 | 0.9605 | all 13 checks pass |
| QAOA-20, L=14 | 128,886 µs | 40,585 µs | 3.18× faster | 0.58 | 0.5477 | all 13 checks pass |
| RCA-20, L=14 circuit differs: 145 two-qubit gates, the paper implies 257 | 85,702 µs | 74,363 µs | 1.15× faster | 0.56 | 0.6125 | all 13 checks pass |
| VQE-20, L=14 | 134,186 µs | 40,585 µs | 3.31× faster | 0.55 | 0.5477 | all 13 checks pass |
| BV-40, L=14 | 18,479 µs | 10,473 µs | 1.76× faster | 0.94 | 0.9178 | all 13 checks pass |
| QAOA-40, L=14 | 370,858 µs | 108,145 µs | 3.43× faster | 0.12 | 0.08197 | all 13 checks pass |
| RCA-40, L=14 circuit differs: 305 two-qubit gates, the paper implies 557 | 204,547 µs | 156,135 µs | 1.31× faster | 0.27 | 0.3427 | all 13 checks pass |
| VQE-40, L=14 | 784,346 µs | 108,145 µs | 7.25× faster | 0.07 | 0.08197 | all 13 checks pass |
| BV-100, L=14 | 60,293 µs | 27,513 µs | 2.19× faster | 0.83 | 0.7988 | all 13 checks pass |
| QAOA-100, L=14 | 2,568,088 µs | 343,026 µs | 7.49× faster | 0.00 | 2.344e-07 | all 13 checks pass |
| RCA-100, L=14 circuit differs: 785 two-qubit gates, the paper implies 1457 | 572,560 µs | 396,199 µs | 1.45× faster | 0.03 | 0.06124 | all 13 checks pass |
| VQE-100, L=14 | 4,329,181 µs | 343,026 µs | 12.62× faster | 0.00 | 2.344e-07 | all 13 checks pass |
The paper prints its fidelities to two decimals; ours is the same estimate (Table 1's fidelities, transport counted per ion moved, a swap once, the paper's decoherence term).
What moves these numbers, each run both ways on A-60 at L = 8:
- Two gates in one trap at once are allowed (the paper's figures run them). Forbidding it slows VQE-60 by 31%; it stays 6.7× faster than the paper.
- Where the two gate zones sit in a trap is not stated; we space them evenly. Next to the trap ends they cost up to 23% (RCA; 5% on QAOA).
- A swap needs one of its ions in a gate zone: not stated, but every swap the paper draws has one. Kept on; off, RCA is 6% faster.
- The two-qubit gate is 141 µs, as the worked examples need. With Table 1's 25 µs, ours would be 10–24% faster still.
- QAOA is run on the complete graph, which matches the gate counts the paper plots; its text says 3-regular graphs, which we also ran (128.5 ms at L = 8).
- Every schedule of ours runs each qubit's gates in the circuit's order, even where two commute: the order is the circuit's, not ours to change. An earlier version let QAOA's commuting gates run in any order; keeping the order costs it up to 41% (QAOA-60 at L = 8: 101.3 ms then, 143.3 ms now).
Where ours is fast, it is because all-to-all circuits (VQE, QAOA) are run as a line of ions swept by an odd-even swap network laid along a trail of traps that needs only five broadcast classes.