MODULE / 02 · QUANTUM MACHINE LEARNING

ArcQML Rust-native quantum machine learning framework

Connect parameterized circuits, state simulation, automatic differentiation, and classical optimizers in one workflow for building, training, and validating quantum neural networks.

View performance results Read the technical documentation

ArcQML uses hybrid noncommercial source-available licensing, not an OSI-approved open-source license. The official closed-source Runtime is licensed separately. Read the license summary

ArcQMLQuantum–classical training loop

From quantum circuits to trainable modelsEncode input x, execute circuit θ, compute loss L, and update parameters using gradients.

Core capabilities: Automatic Differentiation
Workloads: Quantum Neural Network
Comparison dimensions: Training / Validation

FRAMEWORK / 01

Connect quantum circuits
with classical training

A unified interface organizes data encoding, parameterized circuits, gradient computation, and parameter updates, with consistent experiment entry points for different quantum machine-learning backends.

01Data and parameters
Prepare training samples, circuit structure, and trainable parameters.
02Circuit execution
Evolve and measure quantum states on the target simulation backend.
03Gradient computation
Connect the loss to automatic differentiation to compute parameter gradients.
04Classical updates
Update parameters with the optimizer for the next training iteration.

WORKFLOW / 02

Training and validation
share one execution backend

Training measures complete iteration costs, while validation measures execution with fixed parameters. Together, these results show performance across circuit depths and thread configurations.

01 / TRAINTraining

Repeat forward passes, loss evaluation, gradient computation, and parameter updates.

02 / VALIDATEValidation

Run the trained model and compare total validation time across frameworks.

03 / SCALETwo comparison dimensions

Fix either circuit layers or thread count to examine parallel scaling and circuit-size effects.

Ubuntu 24.04.4 LTS · Intel Xeon Platinum 8253 · ArcQML 0.1.0

Test environment and workload Applies to all QNN and VQE charts on this page, including overviews

All test groups share the software and hardware below. The machine core count describes total capacity; actual QNN thread counts are shown in each chart.

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

Framework and dependency versions
Software	Version tested
Python	3.11.15
ArcQML	0.1.0
MindQuantum	0.12.0
PennyLane	0.45.1
PennyLane Lightning	0.45.0
TensorCircuit	0.12.0
JAX	0.4.38
TensorFlow	2.18.1
TensorFlow Quantum	0.7.6
Cirq-core	1.5.0

QNN circuit configuration
Circuit layers	Parameters	Total gates
1	37	72
10	370	540
20	740	1060
30	1110	1580
40	1480	2100

VQE molecular systems
Molecule	Qubits	Hamiltonian terms
H₂	4	15
LiH	6	62
BeH₂	8	61
HF	10	276
H₂O	12	551
NH₃	14	1750

VQE tests use 6, 17, 34, 67, and 134 circuit layers. The following data comes from 30 original ArcQML VQE test records. Each cell shows parameters / total gates; scroll horizontally on narrow screens to see all layer counts.

VQE circuit parameters
Molecular system	6 layers	17 layers	34 layers	67 layers	134 layers
H₂	48 / 74	136 / 206	272 / 410	536 / 806	1072 / 1610
LiH	72 / 110	204 / 308	408 / 614	804 / 1208	1608 / 2414
BeH₂	96 / 148	272 / 412	544 / 820	1072 / 1612	2144 / 3220
HF	120 / 188	340 / 518	680 / 1028	1340 / 2018	2680 / 4028
H₂O	144 / 224	408 / 620	816 / 1232	1608 / 2420	3216 / 4832
NH₃	168 / 260	476 / 722	952 / 1436	1876 / 2822	3752 / 5636

ArcQML VQE run settings
Single CPU thread; 100 training steps per run, repeated 3 times, reporting the median time. Per-step time is the median run time divided by 100. Adam uses learning rate 0.05; precision is float64 / complex128, with the STO-3G basis and Jordan–Wigner mapping. These settings are documented solely from the original ArcQML VQE records.

RESULTS / 03 · QNN OVERVIEW

QNN overall
Performance overview

Compare framework runtimes for the same QNN task relative to ArcQML (1). Bars use a logarithmic scale; smaller values mean shorter runtimes.

Bar chart of relative QNN runtime: ArcQML 1, MindQuantum 1.656, PennyLane 51.892, TensorCircuit 4.143, and TensorFlow Quantum 11.321

RESULTS / 04 · QNN BENCHMARK

QNN training and validation
Performance results

Hold circuit layers fixed and vary thread count on the horizontal axis to compare parallel scaling across frameworks.

CIRCUIT DEPTH / 01

1 layer

QNN training time versus thread count with 1 circuit layer — Training

QNN validation time versus thread count with 1 circuit layer — Validation

CIRCUIT DEPTH / 10

10 layers

QNN training time versus thread count with 10 circuit layers — Training

QNN validation time versus thread count with 10 circuit layers — Validation

CIRCUIT DEPTH / 20

20 layers

QNN training time versus thread count with 20 circuit layers — Training

QNN validation time versus thread count with 20 circuit layers — Validation

CIRCUIT DEPTH / 30

30 layers

QNN training time versus thread count with 30 circuit layers — Training

QNN validation time versus thread count with 30 circuit layers — Validation

CIRCUIT DEPTH / 40

40 layers

QNN training time versus thread count with 40 circuit layers — Training

QNN validation time versus thread count with 40 circuit layers — Validation

01 / 05

RESULTS / 05 · VQE OVERVIEW

VQE overall
Performance overview

Compare VQE costs across frameworks relative to ArcQML time and peak memory (both 1). The horizontal axis is relative time and the vertical axis relative peak memory; positions closer to the lower left indicate lower time and memory costs.