MODULE / 01 · TENSOR NETWORK

ArcTN Rust-native tensor-network contraction planning

Use multi-chain annealing to find lower-cost tensor-network contraction paths for large-scale quantum-circuit simulation, with efficient search and evaluation implemented in Rust.

View experimental results Read the technical documentation

Company-owned ArcTN code uses the Arclight Non-Commercial Source-Available License 1.0, not an OSI-approved open-source license. Light / Heavy dynamic libraries require separate authorization. Read the license summary

ArcTNCooperative multi-chain search

Explore and exchange to find lower-cost pathsEach node represents a candidate contraction tree; green arrows indicate state exchanges between chains.

Core method: Parallel Tempering
Computational object: Tensor Network
Implementation: Rust-native

METHOD / 01

Multiple search chains
work together for better paths

Search chains at different temperatures balance exploration and convergence, periodically exchanging candidates. Hot chains help escape local optima, while cold chains retain better contraction orders. Rust handles frequent candidate evaluation and parallel scheduling.

01Build the network
Derive tensors and index connectivity from a quantum circuit.
02Parallel path search
Temperature chains independently perturb and evaluate candidate contraction orders.
03Exchange states
Exchange states according to acceptance probabilities to broaden the effective search.
04Return a lower-cost path
Select an executable plan using computation and memory metrics.

RESULTS / 02 · PERFORMANCE OVERVIEW

ArcTN optimization performance

The light optimization scenario compares ArcTN Light with Cotengra Auto; the intensive scenario compares ArcTN Heavy with Cotengra AutoHQ and OMECO. The two overview charts compare contraction-order search time with FLOPs, and total tensor-element reads and writes with the largest intermediate tensor; lower is better for all four metrics.

32 distinct circuits · 5 repeats per circuit · 8 physical cores per run · 3,600-second limit

Test environment and settings Applies to all performance charts on this page

Tests cover contraction-path search only, excluding numerical tensor contraction. Optimization time measures the search, while FLOPs describe the computational cost of the resulting path.

Path-search configuration
Setting	Test setting
Task	Contraction-path search (planning only)
Input networks	EinsumBenchmark32, containing 32 distinct quantum circuits
Repeats	5 repeats per circuit for every method
Optimization objective	FLOPs only
CPU allocation	8 physical cores per run
Time limit	External wall-clock limit of 3,600 seconds per run
Test size	32 distinct circuits × 5 repeats, 160 runs total

Hardware
Item	Specification
CPU	Intel Xeon Platinum 8253 @ 2.20 GHz
Physical cores	8 sockets × 16 cores, 128 cores total
CPU topology	256 logical CPUs, 8 NUMA nodes
System memory	About 2.95 TiB
Operating system	Ubuntu 24.04.4 LTS
Linux kernel	6.8.0-117-generic
Platform	x86_64; glibc 2.39; CPU path search

Framework and dependency versions
Software	Version tested
Python	3.12.3
ArcTN	0.1.0
Cotengra	0.8.2
cotengrust	0.2.1
OMECO	0.2.6
Google TNCO	0.4.0
NumPy	2.2.6
opt_einsum	3.4.0
Optuna	4.9.0
CMA-ES (cmaes)	0.13.0
KaHyPar	1.3.5

Scatter plot comparing ArcTN Light, ArcTN Heavy, and baseline methods by contraction-order search time and FLOPs — Google TNCO SA uses a lightweight path search, but the resulting contraction paths have relatively low quality. Across all 32 test circuits, its geometric mean FLOPs are 11.00× those of ArcTN Light and 15.05× those of ArcTN Heavy. Compared with Google TNCO SA, ArcTN Light and ArcTN Heavy achieve end-to-end speedups of 1.06× and 1.25×, respectively. Google TNCO SA is omitted from the per-circuit comparisons.

Scatter plot comparing ArcTN Light, ArcTN Heavy, and baseline methods by total tensor-element reads and writes and largest intermediate tensor size — Google TNCO SA uses a lightweight path search, but the resulting contraction paths have relatively low quality. Across all 32 test circuits, its geometric mean FLOPs are 11.00× those of ArcTN Light and 15.05× those of ArcTN Heavy. Compared with Google TNCO SA, ArcTN Light and ArcTN Heavy achieve end-to-end speedups of 1.06× and 1.25×, respectively. Google TNCO SA is omitted from the per-circuit comparisons.

RESULTS / 03 · METRIC BREAKDOWN

ArcTN contraction planning
Detailed capabilities

Experiments compare computation, data writes, and intermediate memory across methods. The horizontal axis is a multiple of the lowest value in each chart; lower is better.

Bar chart comparing relative FLOPs across tensor-network contraction planning methods — 01 / COMPUTEFLOPs
Measures the floating-point operations required to complete the contraction.

Bar chart comparing the relative total number of elements written across tensor-network contraction planning methods — 02 / DATA MOVEMENTTotal elements written
Measures the cumulative number of elements written during the contraction.

Bar chart comparing relative largest intermediate tensor sizes across tensor-network contraction planning methods — 03 / INTERMEDIATELargest intermediate tensor
Indicates the memory required by the largest intermediate result.

Bar chart comparing relative peak live element counts across tensor-network contraction planning methods — 04 / PEAK MEMORYPeak live elements
Measures the peak number of elements held in memory at the same time during execution.

LIGHT / ARCTN LIGHT

VS Cotengra Auto

Optimization time1.62× faster
FLOPs61.32% lower
Total elements written64.16% lower
Largest intermediate tensor59.39% lower
Peak live elements61.30% lower

HEAVY / ARCTN HEAVY

VS Cotengra AutoHQ

Optimization time7.21× faster
FLOPs43.83% lower
Total elements written34.96% lower
Largest intermediate tensor39.85% lower
Peak live elements38.81% lower

HEAVY / ARCTN HEAVY

VS OMECO

Optimization time1.33× faster
FLOPs11.85% lower
Total elements written6.89% lower
Largest intermediate tensor16.88% lower
Peak live elements18.72% lower

DETAILS / 04

ArcTN per-circuit
performance (Heavy)

The intensive scenario presents circuits of different sizes and structures, comparing ArcTN Heavy with baselines on contraction work and search time to examine how consistently the overall advantages hold per circuit.

HEAVY · Result quality Compare contraction work across methods for different circuit sizes and structures.

REPRODUCIBILITY / 05

Multiple search chainswork together for better paths