ArcLight Quantum at DAC 2026: 1000× Faster CNOT Synthesis, 606× Faster Simplification, 95% Error-Rate Cut for Neutral-Atom QEC

Forum topic · 小凯 · 2026-08-15

Summary

ArcLight Quantum had three papers accepted at DAC 2026, all delivering order-of-magnitude gains for quantum compilation. (1) Lin-search performs optimal CNOT synthesis by combining two lower-bound prunings with canonical-form deduplication; on a laptop under a 600-second time limit it pushes the tractable frontier by 3 qubits over Qiskit-SAT and accelerates small-to-medium circuits by up to ~1000×, while embedded in Clifford+T pipelines it cuts CNOT count by up to 26% and timeouts by 30%. (2) A hypergraph-based simplifier represents same-direction adjacent gates as hyper-nodes and preserves cross-qubit connections via hyper-edges, outperforming Qiskit's DAG template matcher by 11-65× on random circuits and 606× (3.5 hours vs 21 seconds) on a 2,707-gate 48-qubit arithmetic circuit with no quality loss. (3) NEAT jointly optimizes data-atom placement, shuttler movement, and gate timing for neutral-atom QEC under a unified symbolic constraint model, reducing syndrome-extraction move rounds from 100-600 to 3-4 and total distance by roughly an order of magnitude, yielding logical error rates 2-20× lower than Enola (up to 95% reduction).

Background

At DAC 2026 (Design Automation Conference), ArcLight Quantum had three papers accepted covering CNOT synthesis, circuit simplification, and neutral-atom quantum error correction (QEC) compilation. Although the entry points differ, all three target the same goal: make quantum programs run with fewer resources, lower cost, and higher fidelity on real hardware. In the best cases each delivers gains in the hundreds.

Breakthrough 1 — Lin-search: Up to ~1000× Faster Optimal CNOT Synthesis

CNOT gates are the most common two-qubit operations and the noisiest. Reducing CNOT count directly improves circuit fidelity.

Compilers such as Qiskit-SAT find "fairly short" CNOT circuits but cannot guarantee optimality. ArcLight instead performs exact synthesis: given a target, return a circuit that uses the fewest CNOTs.

Their Lin-search algorithm searches the CNOT circuit space directly and prunes with two lower bounds:

Distance-to-target-rows bound: back-propagates from output differences to estimate the minimum remaining gates.
Cut-rank bound: estimates how much independent information must still flow across regions via CNOTs.

A canonical-form routine automatically merges equivalent circuits.

Benchmark results. For a 4-qubit, 5-CNOT target, a naive search must inspect 248,832 candidates; with both lower bounds plus equivalence merging, only 18 candidates are checked and the optimum is found in 0.27 ms. On random 15-qubit, 15-CNOT tests, Lin-search pushes the tractable frontier by 3 qubits beyond Qiskit-SAT and yields hundreds-fold speedups, up to ~1000× for small-to-medium instances.

Deployment. Embedded in a Clifford+T optimization pipeline, Lin-search reduces consecutive CNOT sub-circuits by an average of 16% (max 26%) and cuts timeout failures by 30% relative to Qiskit-SAT.

Breakthrough 2 — Hypergraph Simplifier: Up to 606× Speedup

Before mapping to hardware, quantum circuits are typically simplified. Exchange relations between gates explode the search space. Existing methods use a DAG that keeps non-commuting dependencies and deletes commuting edges, which turns exchangeable gates into isolated nodes and forces blind global search at scale.

ArcLight introduces a hypergraph representation:

Merge adjacent same-direction gates on the same qubit into hyper-nodes.
Use hyper-edges to restore cross-qubit connections.
Match templates by expanding outward from a local neighborhood.

Head-to-head against Qiskit's DAG template matcher:

| Circuit size | Qiskit | Qiskit (heuristic) | Hypergraph | Speedup | |---|---|---|---|---| | 100 random gates | 27.64s | 18.97s | 1.75s | 15-65× | | 300 random gates | 177.5s | 93.2s | 5.67s | 15-65× | | 500 random gates | 583.34s | 293.85s | 9.19s | 11-29× | | 2,707 gates / 48-qubit arithmetic | ~3.5 h | — | ~21 s | 606× |

3.5 hours vs 21 seconds at parity gate count — the advantage grows with scale, indicating strong potential for large circuits.

Breakthrough 3 — NEAT: Neutral-Atom QEC Compiler Cuts Logical Error Rate by Up to 95%

Neutral-atom platforms (QuEra, Atom Computing, Pasqal, Vector Origin / 向量奇点, etc.) rely on auxiliary atoms shuttling through arrays to extract error syndromes. Prior compilers optimize gate ordering and atom movement separately, like designing bus timetables and routes independently — wasting shuttling time and amplifying errors.

The NEAT compiler places three variables inside one symbolic constraint model:

Where data atoms sit (fixed coordinates).
How auxiliary atoms move (position at every timestep).
When each gate executes (parallel timestep).

The solver simultaneously satisfies logical constraints (code parity checks) and hardware constraints (Rydberg interaction ranges, laser timing, no-collision, relative ordering of parallel trajectories). A two-tier symmetry-breaking scheme first fixes time directions and spatial orientations with light constraints, then adds lexicographic ordering only for hard instances.

Benchmark results vs the Enola compiler:

| Code | Enola move rounds | NEAT move rounds | Enola total distance | NEAT total distance | |---|---|---|---|---| | Surface7 | 177 | 3 | 3907.2 | 568.6 | | Surface13 | 624 | 4 | — | — | | Planar7 | 100 | 3 | 1291.6 | 355.8 | | Planar13 | 292 | 3 | 11115.3 | 1580.1 |

End-to-end logical error rate: under matched noise, NEAT-compiled circuits reach error rates of 1/2 to 1/20 of Enola's, a maximum reduction of 95%.

Three Takeaways

1. Compilation engineering is the real battlefield. With hardware scaling, whether algorithms actually run on machines is decided by the compiler. ArcLight covers synthesis, simplification, and hardware-aware scheduling, each with ~100× gains. 2. Neutral-atom platforms get an engineering acceleration. Compressing syndrome-extraction rounds from 100-600 to 3-4 and shortening total distance by roughly an order of magnitude lets QuEra, Atom Computing, Vector Origin, and peers reach "engineered QEC" sooner without waiting for many more qubits. 3. DAC is the true home of quantum compilation. As the top EDA venue and a primary QDA venue, DAC's acceptance of all three ArcLight papers — co-authored with CAS, Tsinghua, and ECNU collaborators — signals steady domestic output in quantum compilation engineering.

Signals to watch over the next 6–12 months: (1) whether Lin-search, the hypergraph simplifier, and NEAT are integrated into Qiskit, Cirq, PennyLane, or Tket; (2) whether neutral-atom vendors (QuEra, Atom Computing, Vector Origin, 启科量子) cite NEAT numbers in public benchmarks.

Sources:

Tencent News 8-14 ArcLight Quantum DAC report
Photon Box QUANTUMCHINA 8-14
DAC 2026 paper links