Qiskit¶

cuQuantum Appliance 23.03 contains Qiskit and cusvaer. cusvaer is seamlessly integrated into Qiskit Aer, so that users are able to use run multi-node simulations through Qiskit Aer without any modifications to their source code.

For details of the cusvaer backend solver, please refer to cusvaer.

Getting started¶

The following python script is an example to run Greenberger–Horne–Zeilinger (GHZ) circuit.

from qiskit import QuantumCircuit, transpile
from qiskit import Aer

def create_ghz_circuit(n_qubits):
    circuit = QuantumCircuit(n_qubits)
    circuit.h(0)
    for qubit in range(n_qubits - 1):
        circuit.cx(qubit, qubit + 1)
    return circuit

simulator = Aer.get_backend('aer_simulator_statevector')
circuit = create_ghz_circuit(n_qubits=20)
circuit.measure_all()
circuit = transpile(circuit, simulator)
job = simulator.run(circuit)
result = job.result()

print(result.get_counts())
print(f'backend: {result.backend_name}')

This script works with Qiskit Aer. When the script is executed in the cuQuantum Appliance 23.03, the multi-node backend is automatically selected and runs the simulation. The backend name is "cusvaer_simulator_statevector".

$ python ghz.py
{'00000000000000000000': 493, '11111111111111111111': 531}
backend: cusvaer_simulator_statevector

By launching a script with mpirun, simulation will be distributed to multiple processes. In the following examples, two processes are used for simulation, and result will be output twice.

$ mpirun -n 2 python ghz.py
{'00000000000000000000': 536, '11111111111111111111': 488}
cusvaer_simulator_statevector
{'00000000000000000000': 536, '11111111111111111111': 488}
cusvaer_simulator_statevector

By adding the lines shown below, the process where mpi_rank == 0 is selected to output the result.

if result.mpi_rank == 0:
    print(result.get_counts())
    print(f'backend: {result.backend_name}')

$ mpirun -n 2 python ghz.py
{'00000000000000000000': 532, '11111111111111111111': 492}
cusvaer_simulator_statevector

Note

GHZ circuit is used here as an example, which does not mean for demonstrating performance. It is known that the GHZ circuit is too shallow to show good scaling when simulation is distributed to multiple-GPUs and/or multiple-nodes.

Selecting simulator¶

The version of Qiskit Aer installed in cuQuantum Appliance 23.03 is extended to instantiate the cusvaer backend solver when all the following conditions are met:

cusvaer_enable == True

method == "statevector" or method == "automatic"

noise_model == None

cusvaer_enable is an extension to Qiskit Aer that enables cusvaer. The default value is True. The method and noise_model are options provided by Qiskit Aer. cusvaer is selected only if the noise_model option is None as the first release of cusvaer does not support noise model.

cusvaer-specific options¶

Please refer to cusvaer options.

Modifications for Qiskit Aer options¶

The following table shows the modified and added options to Qiskit Aer.

Option	Description
`cusvaer_enable`	If set to `True`, cusvaer is enabled. Otherwise, Qiskit Aer simulators are used. The default value is set to `True`.
`device`	The default value is changed to `"GPU"`.
`cuStateVec_enable`	The default value is changed to `True`.

Interoperability with mpi4py¶

mpi4py is interoperable with cusvaer backend. Please refer to Interoperability with mpi4py for details.

Limitations¶

The 23.03 release has the following limitations:

Noise model and noisy circuit simulations are not supported.
Simulations are executed in serial even if Qiskit Aer’s parallel simulation options (e.g., batched_shots_gpu, batched_shots_gpu_max_qubits) are provided.
Single-process multi-GPU simulation is not supported. Users need to launch multiple MPI processes even in a single node.