Today, I&#39;d like to conduct various benchmarks using Qiskit + cuQuantum. Given that cuQuantum is free, one would expect more articles about it. However, there aren&#39;t many, probably due to the niche nature of the quantum computing industry. First, regarding the installation, you&#39;ll set up the cuQuantum in a conda environment and then install cuQuantum and cuQuantum Python. For Qiskit, you can install qiskit-aer and qiskit-aer-gpu using pip. I am using some references<a href="https://medium.com/qiskit/improve-quantum-simulations-with-qiskit-aer-cuquantum-9cd4bf69f042" rel="noopener noreferrer" target="_blank">https://medium.com/qiskit/improve-quantum-simulations-with-qiskit-aer-cuquantum-9cd4bf69f042</a> <pre class="ql-syntax" spellcheck="false">from qiskit import *
from qiskit.circuit.library import *
from qiskit_aer import *
import time
import numpy as np
import matplotlib.pyplot as plt

def exp_qv(qubits=15, depth=10, device_num=0):
  if device_num ==0:
    sim = AerSimulator(method=&#39;statevector&#39;, device=&#39;CPU&#39;)
  elif device_num == 1:
    sim = AerSimulator(method=&#39;statevector&#39;, device=&#39;GPU&#39;)
  else:
    sim = AerSimulator(method=&#39;statevector&#39;, device=&#39;GPU&#39;, cuStateVec_enable=True)
  circuit = QuantumVolume(qubits, depth, seed=0)
  circuit.measure_all()
  circuit = transpile(circuit, sim)
   
  start = time.time()
  result = sim.run(circuit, shots=1, seed_simulator=12345).result()
  time_val = time.time() - start
  return time_val
</pre> First, we will compare the CPU, GPU, and cuQuantum&#39;s cuStateVec. We&#39;ll run the same problem on different configurations. 1. We will use the CPU. It&#39;s the Intel Core i9 13900KF, which is fairly high-performance.2. We will use the GPU, leveraging acceleration from IBM&#39;s program.3. We will use the GPU, but with acceleration using NVIDIA&#39;s program, cuStateVec. Only the first configuration uses the CPU, while both the second and third use GPUs. However, the third one is our main focus for this session: NVIDIA&#39;s cuQuantum. Both configurations use the same machine and program; the only difference is the SDK. To begin, <pre class="ql-syntax" spellcheck="false">#num of qubits
x = np.arange(10, 30)

#QV depth
dp = 10

#list for result
arr = [[],[],[]]

#label
label = [&#39;Qiskit + CPU&#39;, &#39;Qiskit + GPU&#39;, &#39;Qiskit + cuStateVec GPU&#39;]

#loop
for j in range(3):
  for i in x:
    r = exp_qv(i,dp,j)
    arr[j].append(r)
    print(i, r)
</pre> <pre class="ql-syntax" spellcheck="false">plt.xlabel(&#39;qubits&#39;)
plt.ylabel(&#39;sec&#39;)
for i in range(3):
  plt.plot(x, arr[i], label=label[i])
plt.title(&#39;Benchmark QV depth=&#39; + str(dp))
plt.legend()
plt.show()
</pre> <img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2023/10/10/1.png"/> Already, with 29 qubits, the situation is quite challenging for the CPU. Next, we&#39;ll set the QV depth to 30. <img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2023/10/10/2.png"/> The graph turned out to be exactly the same. Given this, using the CPU for QV seems to be quite challenging. From now on, I&#39;ll compare only between GPUs. I&#39;ll try setting the QV depth to 100. <img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2023/10/10/100.png"/> Essentially, there&#39;s consistently about a 1.4x speed increase. It felt the same even at QV depth=500. <img src="https://assets.blueqat.com/public/uploads/us-east-2:4805ff4b-c3cc-4344-b165-86544c34d0bf/2023/10/10/3.png"/> Overall, it&#39;s indeed quite challenging to compute large problems using a CPU. Additionally, even with a GPU, simply using cuQuantum results in about a 1.5x speed boost, even with the exact same configuration. While I&#39;d like to try various other examples, for this time, using NVIDIA cuQuantum proved to be optimal as a quantum computer simulator. That&#39;s all.

Benchmarking NVIDIA cuQuantum, a quantum computer simulator, using multiple tests on the consumer-grade GPU RTX4090 and Intel i9-13900KF.

Yuichiro Minato