Quantum as a Service

Access to boosted quantum emulators

• 

  Quantum volume helps evaluate a platform’s capacity by running increasingly complex circuits until hitting hardware limits.

  We use square circuits (e.g., 24x24, 30x30) where the number of qubits equals the circuit depth. Gates are randomly chosen from Pauli X, Y, Z, Hadamard, and two-qubit controls like CX and CZ. These circuits, built with Qiskit and Cirq, were executed on both local setups and our Quantum as a Service (QaaS) platforms in state vector mode with double-precision floats.
  As shown in Figure 2, our CPU-based QaaS ran 2–10x faster than local setups for small qubit counts. Local setups hit memory limits beyond 30 qubits, while our 512GB QaaS platform reached 34 qubits in less time than 30 qubits locally.

• 

  To go further, we also benchmarked our GPU-accelerated platforms. In Figure 3, it completely outperforms CPU-based platforms, a 34-qubits circuit passed from 150s on our C64512M to 12s on our 8-L80S configuration offering a 12x boost.

  Moreover, these multi-GPUs platforms allow us to reach up to 36 qubits on squared-circuits, which is more than convenient to push forward quantum computing exploration. As our benchmark is double floating precision, it is possible to reach one additional qubit by switching to single floating precision.

• 

  We also ran some benchmarks on Qsim emulation. Qsim is a full wave function simulator written in C++. It uses gate fusion, vectorized instructions and OpenMP multi-threading to achieve state of the art state vector simulations of quantum circuits. Figure 4 shows us really top performance on CPU setups. The average performance is twice time faster than Aer emulation and offers an additional qubit for the same hardware configuration.

  For even faster Qsim execution, we provide GPU platforms powered by Nvidia cuQuantum.

• 

  When we run Quandela’s Exqalibur, which is dedicated to photonic emulation, we can notice that our GPU-accelerated platforms exhibit a substantial computation speedup for equivalent circuit size, taking less than a second compared to 241 seconds for Apple M2 or 695 seconds for an Intel i7.

  Moreover, in local setups we encounter limitations by running squared-circuits with more than 11 photons due to excessive memory requirement. In contrast our H100 GPU-accelerated platform, enables us to extend up to 31 photons in 2h.