Delaying data transfers speeds up quantum circuit simulations on GPUs

Why moving data matters

A state-vector simulator represents a quantum system by storing its probability amplitudes: complex numbers describing the possible outcomes of measurement. Each additional qubit doubles the number of amplitudes. Distributing this growing array across graphics processing units, or GPUs, makes larger simulations possible, but creates a communication problem.

Some quantum operations can use data already held on each GPU. Others require amplitudes stored elsewhere. Simulators can redistribute the array through a process called qubit reordering, making subsequent operations local. These transfers can be expensive, especially between separate computing nodes rather than GPUs within one node.

Yusuke Teranishi and colleagues propose doing more work between transfers. Their “lazy” approach delays operations requiring redistribution while advancing other operations that can safely run with the current data arrangement.

Scheduling without changing the calculation

The researchers adapt time-space tiling, a technique used to improve data reuse in classical computing. Here, a tile is a group of quantum operations that can run under the same distribution of data. Larger useful tiles mean fewer redistributions.

They develop separate methods for two components of the variational quantum eigensolver, an algorithm used to estimate low-energy states of systems such as molecules. Quantum state update applies a circuit to produce a state. Expectation value computation then calculates quantities such as that state's energy.

For state updates, the scheduler must preserve dependencies: an operation cannot jump ahead of another whose output it needs. The method groups executable gates and greedily chooses a new data mapping for subsequent work. A hierarchical version preferentially exchanges data within a node, accepting extra inexpensive transfers when that avoids slower transfers between nodes.

Expectation-value calculations offer greater scheduling freedom because contributions from different Hamiltonian terms can be evaluated in any order before being added. The method searches possible data mappings and repeatedly selects a large group of compatible terms.

It also exploits the diagonal structure of Pauli Z operators, building blocks of those terms. Unlike X and Y operators, Z contributions on distributed qubits can be handled through local calculations with appropriate signs, without redistributing the state vector. The final partial results still require a collective reduction.

What the experiments found

The implementation used NVIDIA's cuStateVec library and MPI communication on the University of Osaka's SQUID supercomputer, with eight A100 GPUs per node. Tests used double precision, molecular Hamiltonians and GateFabric circuits. Most reported timings were medians of three executions; simulation timings excluded initialization.

For the 30-qubit GateFabric benchmark, the reported 32-GPU results show flat tiling reducing qubit reorderings from 1,920 to 27. Execution time fell from 664.0 seconds to 12.2 seconds, a roughly 54-fold improvement over the authors' in-order GPU baseline. Hierarchical tiling reduced communication time by up to a further 15% relative to flat tiling in the tested cases.

For a 30-qubit water Hamiltonian, the full expectation-value method reduced reorderings from 42,512 to four. Runtime dropped from 6,233 seconds to 3.76 seconds, approximately 1,657-fold faster than the state-update-based baseline.

That exceptionally large gain did not come from scheduling alone. The optimized method also used a specialized Pauli-expectation API instead of treating the terms as general gates, substantially reducing computation time as well as communication.

The authors also tested a less regular, 30-qubit quantum-volume circuit. Hierarchical tiling took 13.5 seconds, compared with 302.3 seconds for their in-order baseline and 128.5 seconds for QuEST. Scheduling overhead was reported below 0.01 seconds for state updates and below one second for expectation values in the evaluated cases.

Limits and publication status

These results accelerate classical simulation, not a physical quantum processor. They do not remove the exponential memory growth of state vectors or establish equivalent gains for every circuit, Hamiltonian or machine.

Performance depended on network topology, and communication between nodes still limited scaling. Comparisons with other simulators also require care: mpiQulacs and Qiskit Aer ran on CPUs in the molecular comparisons, while some competing configurations failed or exceeded practical runtime limits. Realistic noise models remained future work.

The supplied source is arXiv version 2, dated 7 November 2026. However, its full text also identifies an ACM Transactions on Quantum Computing publication from 2025 and records acceptance on 27 June 2025. It therefore should not be described simply as an unreviewed preprint. The evidence supports a practical conclusion: reorganizing calculations around data locality can deliver substantial, workload-dependent savings.