Abstract
Scaling quantum computers, i.e., quantum processing units (QPUs) to enable the execution of large quantum circuits is a major challenge, especially for applications that should provide a quantum advantage over classical algorithms. One approach to scale QPUs is to connect multiple machines through quantum and classical channels to form clusters or even quantum networks. Using this paradigm, several smaller QPUs can collectively execute circuits that each would not be able on its own. However, communication between QPUs is costly as it requires generating and maintaining entanglement, and hence it should be used wisely. In this paper, we evaluate the architectures, and in particular the entanglement patterns, of variational quantum circuits in a distributed quantum computing (DQC) setting. That is, using Qiskit, we simulate the execution of an eight qubit circuit using four QPUs each with two computational and two communication qubits where non-local CX-gates are performed using the remote-CX protocol. We compare the performance of various circuits on a binary classification task where training is executed under ideal and testing under noisy conditions. The study provides initial results on suitable VQC architectures for the DQC paradigm, and indicates that a standard VQC baseline is not always the best choice, and alternative architectures that use entanglement between QPUs sparingly deliver better results under noise.
Introduction: The Challenge of Scaling
Scaling quantum processing units (QPUs) is essential for achieving quantum advantage. While individual NISQ-era machines are growing, large-scale algorithms require more qubits than a single chip can currently provide. Distributed Quantum Computing (DQC) addresses this by clustering smaller QPUs via quantum and classical channels to act as a single large machine.
The primary bottleneck in DQC is communication. Executing a gate between qubits on different chips (a "remote-CX") requires generating and maintaining entanglement (Bell pairs), which is technically difficult and introduces significant noise. This study investigates which VQC architectures best balance machine learning performance with the high overhead of distributed communication.
Methodology and the Remote-CX Protocol
The researchers simulated an 8-qubit VQC distributed across four QPUs using Qiskit. Each QPU was modeled with two data qubits and two communication qubits, totaling 16 qubits in the simulation.
A critical component of this setup is the Remote-CX Protocol. To perform a CNOT gate between qubits on separate QPUs, the system uses a shared Bell pair and local operations to "teleport" the gate's action. Circuits were trained under ideal, monolithic conditions using the SPSA optimizer and then tested in a noisy distributed environment (depolarizing error probability = 0.03) to evaluate robustness.
Architecture Comparison
Four distinct entanglement patterns were evaluated:
- Baseline VQC: Standard linear entanglement.
- Fully Entangled: All-to-all entanglement at the start.
- Alternating Layers: Interleaves "Global" layers (inter-QPU CX gates) with "Local" layers (intra-QPU CX gates).
- Alternating Layers 2: A sparser variation of the alternating approach.
Key Results and Insights
The study revealed that in distributed settings, "less is more." The "Alternating Layers" architecture was the top performer, achieving validation accuracy near 0.9 on synthetic datasets and maintaining the highest median test accuracy under noise.
- The Entanglement Paradox: While entanglement is necessary, excessive global entanglement (as seen in the "Fully Entangled" model) led to poor learning, often failing to exceed random guessing (0.5 accuracy).
- Sparsity and Robustness: Architectures that use inter-QPU entanglement sparingly are naturally more robust to the noise inherent in remote operations. The "Alternating Layers" design showed the smallest performance drop when transitioning from ideal training to noisy testing.
- Strategic Synchronization: Interleaving local processing with occasional global synchronization proved more effective than standard monolithic entanglement patterns.
Limitations and Future Outlook
While promising, these results are based on a simplified depolarizing noise model. Real-world quantum networks face more complex challenges like photon loss and decoherence. Future research will need to incorporate entanglement purification, explore different network topologies (e.g., star or grid), and scale beyond 8 data qubits to fully realize the potential of distributed quantum machine learning.
Summary updated