GPUs accelerate quantum circuit simulation for wider spin and hardware designs

Researchers at VSB-Technical University of Ostrava and Klaipeda University have detailed a comparative analysis of three methods for simulating variational quantum circuits: gate-based statevector simulation, direct Pauli evolution and matrix product states. The work, published in the Journal of Systems Architecture, finds that direct Pauli evolution excels with complex molecular Variational Hamiltonian Ansätze, while graphics processing units (GPUs) significantly accelerate simulations of wider spin and hardware-efficient circuits. The team provides a procedure for jointly selecting simulation representation, precision and resource allocation, finding that the GPU delivers approximately 30-fold acceleration over 18-thread Aer for a 28-qubit, four-layer TFIM circuit in FP64.

Direct Pauli Evolution Accelerates Molecular Variational Hamiltonian Ansatz

Direct Pauli kernel implementations bypass computational ladders inherent in standard rotation methods, offering a performance advantage for specific quantum circuit types. This optimization yielded a 3.45-fold speed increase when applied to two-layer full LiH molecular simulations using NumPy, demonstrating the potential for significant gains through algorithmic refinement. The improvement stems from a direct update formula, R_P(φ)ψ = cos(φ/2)ψ-i sin(φ/2)Pψ, which uses bit permutations and phases instead of dense operator construction, streamlining the calculation.

This suggests that optimal performance isn’t solely dependent on hardware acceleration but requires careful consideration of the circuit’s structure and the chosen simulation method. The work also examined the impact of varying width, depth, precision, and resource allocation on simulation performance, providing a framework for jointly selecting these parameters. Lightning, a native-Pauli CPU implementation, demonstrated strong molecular performance. But the overall trend indicates that a molecular CPU/GPU reversal is possible, suggesting that the best approach depends on the specific characteristics of the quantum circuit being simulated.

GPU Acceleration of Wider Spin and Hardware-Efficient Circuits

GPU acceleration demonstrably improves performance for specific types of quantum circuit simulations, with gains reaching approximately 30-fold over multi-threaded CPU execution for 28-qubit, four-layer Transverse-Field Ising Model (TFIM) simulations. The study systematically varied circuit width, depth, precision, and resource allocation to pinpoint these performance differences across CPU and GPU systems. The research reveals a detailed relationship between simulation method and hardware.

Matrix product states prove useful when limited entanglement allows for accurate compression, while corrected CPU FP32 calculations can achieve the tested tolerance levels, albeit at a higher computational cost than FP64. The work details how workload size, startup overhead, worker memory, and task-specific profiling all influence the preferred resource allocation strategy.

“The large GPU gains for wider spin and HEA circuits show a different operating regime,” the study notes, indicating a shift in the most efficient computational pathway as circuit complexity increases. The researchers emphasize that both the primitive operations and their implementation matter, with a matched NumPy Pauli update accelerating two-layer full LiH simulations by 3.45-fold.

Internal commutation within the simulations enables exact block factorization, a technique that contributes to efficiency. These circuits were tested with varying numbers of qubits, ranging from 8 to 28, and layer counts, allowing for a comprehensive assessment of performance scaling.

The researchers also explored signed-Ising QAOA and product/GHZ ansatzes, providing a broad comparison across different quantum circuit designs. The findings highlight the importance of considering the specific characteristics of the quantum algorithm when choosing a simulation method and hardware platform, ultimately clearing the way for more efficient quantum software development and verification.

CPU/GPU Allocation Impacts Variational Quantum Algorithm Cost

The selection of computational resources, CPU versus GPU, significantly impacts the cost of simulating variational quantum algorithms, with workload characteristics dictating optimal allocation. Beyond CPU optimization, the study reveals substantial gains from using GPUs for specific circuit types. The researchers emphasize that the benefits extend beyond simple speedup, as the choice of representation, precision, and worker allocation are all intertwined and must be considered jointly. The team’s work demonstrates that a careful evaluation of these factors is essential for maximizing the efficiency of quantum circuit simulation.

FP32 vs FP64 Precision Trade-offs in Quantum Circuit Simulation

The selection of floating-point precision, FP32 or FP64, during quantum circuit simulation presents a detailed trade-off between computational cost and accuracy, with corrected CPU FP32 potentially matching the tolerance of FP64 while demanding more resources. This finding challenges the assumption that higher precision always guarantees superior results, particularly when simulating complex quantum systems.

The study details how achieving a target accuracy level does not solely depend on precision, but is also heavily influenced by the specific circuit characteristics and available computational resources. Instead, the best method is contingent on the structure of the quantum circuit being modeled.

Evaluating Accuracy and Resource Use in Quantum Simulation

The study’s focus on LiH, alongside the broader investigation of transverse field Ising model (TFIM) circuits, demonstrates a targeted approach to simulation optimization beyond generic circuit types. This analysis moves beyond simply maximizing speed, acknowledging that trade-offs between these factors are often necessary.

The team’s analysis of Aer’s accuracy revealed that it retains approximately 1.15 x 10-6 after correction, a level of precision achieved through reclassification of NumPy/Lightning accuracy paths. The researchers point out that state export and readout costs must also be considered when evaluating the benefits of correction, as these steps add to the overall computational burden.

NumPy Optimizations Speed LiH Simulation on CPUs

Matched comparisons using NumPy Pauli updates yielded a 3.45-fold acceleration for two-layer full LiH simulations, demonstrating the potential of optimized classical approaches to enhance quantum software development. The research team found that GPU acceleration continues to offer substantial benefits, with a 30-fold speedup observed over an 18-thread Aer implementation for a 28-qubit, four-layer TFIM circuit using FP64 precision. The team’s analysis emphasizes that a careful consideration of these factors alongside task-specific profiling is important for maximizing simulation efficiency.

Procedure for Jointly Selecting Simulation Parameters and Resources

Selecting the appropriate simulation method and computational resources requires a joint consideration of representation, precision and worker allocation. This acceleration stems from streamlining the calculation by removing unnecessary steps, without altering the underlying circuit or its resulting energy.

These models demonstrate that even with corrected calculations, CPU-based simulations using 32-bit floating point precision require approximately 3 to 24 percent more runtime than their 64-bit counterparts to meet tested tolerance levels. The findings suggest that a careful balance between computational resources, simulation representation, and numerical precision is essential for efficiently simulating and advancing the field of quantum software development.

👉 More information
🗞 Efficient simulation of variational quantum circuits: representation, accuracy and CPU/GPU allocation
✍️ Vojtěch Novák, Radovan Rečka, Ondřej Knebl and Ivan Zelinka (Department of Computer Science, Faculty of Electrical Engineering and Computer Science, VSB-Technical University of Ostrava, Ostrava, Czech Republic)
🧠 ArXiv: https://arxiv.org/abs/2611.24444v1

Stay current

See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.

Avatar of Dr. Donovan

Latest Posts by Dr. Donovan: