Accelerating Quantum Computer Simulators using GPUs

Chaithanya Naik Mude, Vishnu Ramadas

{cmude, vramadas} @wisc.edu
University of Wisconsin-Madison

Introduction Tensor Networks QuEST Q-GPU

Quantum computers is an ever growing area with new research Tensor network methods are most widely used methods for quan- QuEST is first open-source quantum computer simulation toolkit Q-GPU is a new simulator from the University of Pittsburg that
changing the landscape regularly. Due to the dynamic nature of tum simulations and . Tensors capture the idea of multi-linear that can work on multiple platforms such as a single-node CPU, aims to be a high-performance and scalable QCS that uses GPUs
the field, and the fact that access to a quantum computer is not maps and tensor networks are collection of tensors connected by multi-node CPU, or GPU [11]. It was released by The Univer- to accelerate compute [12]. Existing simulators such as Qiskit [13]
easy, researchers rely on quantum computer simulators (QCS). contractions. They also utilize graphical representations to under- sity of Oxford in 2019 and can simulate generic quantum circuits statically assign all wavefunction amplitudes. This makes the GPU
Simulators significantly reduce the barrier for researchers to pro- stand the workflow intuitively. comprising of one and two-qubit and multi-qubit controlled gates compute inefficient for large qubit system simulations since data
totype and investigate new algorithms. It also enables researchers The main theme of these methods is approximate the quantum quantum circuits. It can be used with pure states (using state transfer to the GPU takes up a significant time of the overall run.
to probe into deeper circuits that NISQ era computers struggle state using tensor networks. They employ lossy data compres- vectors) and mixed states (using density matrices) under the pres- Q-GPU allocates state amplitudes dynamically and maximize the
with [1]. Although QCS can run any quantum program that a sion that preserves the important properties of the quantum state. ence of decoherence. data transfer overlap with the gate operation compute. This re-
real device can run, they quickly run into memory and computa- These methods enable simulation of large systems, use-cases in- QuEST optimizes Schrodinger-style simulations by replacing cer- sults in a much better GPU utilization rate since the time spent idle
tion time constraints [2]. clude domains quantum chemistry, variational optimization, un- tain gate operations with their effect. For instance, instead of waiting for data to arrive is lower than before. Q-GPU also uses
derstanding quantum devices. multiplying an X gate with a state vector, QuEST simply switches dynamic zero state amplitude pruning and lossless compression of
the probability amplitudes of the states. It also speeds up simula- non-zero amplitudes to reduce the data transfer time. Addition-
tion by distributing the workload across CPU cores using OpenMP, ally, it also performs dependency-aware quantum gate reordering
across different CPUs in a cluster using MPI, or on a GPU using to improve the chances of pruning zero state amplitudes.
CUDA.

Figure 2: Tensor network methods

Some of the best known applications of tensor networks are 1D

Matrix Product States (MPS), Tensor Trains (TT), Tree Tensor
Figure 8: Comparison of Q-GPU with Google Qsim-Cirq v0.8.0 and Microsoft
Networks (TTN), the Multi-scale Entanglement Renormalization (a) Performance on a Single Node (b) Normalized Performance on a QDK v0.15
Ansatz (MERA), Projected Entangled Pair States (PEPS) [8]. (OpenMP vs CUDA) Distributed System (MPI)
Figure 5: QuEST: Various parallelization approaches
Conclusions and Future work
Figure 1: Time to simulate various circuit sizes as a function of circuit QCLAB++
depth [3] Leveraging efficient representations and parallelism through hard-
QCLAB++ is a new simulator developed at Lawrence Berkeley ware can scale simulations to larger qubit systems. The hybrid
Modern CPUs offer various performance optimizations that soft- National Lab in early 2023 [1]. It is a simulator that tracks all the approach of using smaller sized QPUs alongside to speed up cer-
ware can make use of. For instance, state vector updates could amplitudes of the wavefunction and, as a result, scales exponen- tain tasks can improve performance of classical simulation which
use the vector processing unit, also called a SIMD unit, that allows tially with an increase in the number of qubits and linearly with in-turn, along with efficient tensor network representations can
an operation to work on multiple data elements at once [4]. Sim- gate count. The scaling, however, can be made more efficient scale for larger qubit systems.
ulators can extract additional performance in multi-node CPUs by using tensor networks or sparse representation of wavefunctions We focus mainly on simulation speedups for our work, but various
distributing computation across various nodes using libraries such at the cost of accuracy. QCLAB++ uses efficient quantum gate compression techniques that reduce the memory requirements are
as OpenMP or MPI [5] [6]. simulation algorithm from to calculate the effects of gates on the possible too. Work such as [14] and the more recent [15] pro-
GPUs, on the other hand, have been a very popular option for state vectors. pose techniques to compress the gate and state representations.
large scale scientific computing since its invention by Nvidia. GPUs Figure 3: Flops vs operations for different tasks The latter focuses on simulations that use tensor networks with
offer the ability to work with extremely large vectors/matrices by representations that are highly parallelizable and provide both sim-
parallelly performing operations on its elements. For example, they ulation speedups and reduced memory usage.
can multiply matrices in O(n) time since each element of a row of Figure 3 shows the advantage of using tensor methods. Figure 4
the resultant matrix can be computed independently of the others. highlights the speedup of using hybrid architectures that leverage References
GPUs are increasingly becoming a popular platform for QCS [7]. both CPU and GPU vs those that only use either one of them.
This is due to trade-off between cost of loading data into the [1] R. V. Beeumen, D. Camps, and N. Mehta, “Qclab++: Simulating quantum circuits on gpus,” 2023.
[2] F. Arute et al., “Quantum supremacy using a programmable superconducting processor,” Nature, vol. 574, pp. 505–510, Oct
Quantum Simulations GPUs and the parallelism offered by GPUs. 2019.
Figure 6: CPU versus GPU for QFT circuit [3] S. Boixo, S. V. Isakov, V. N. Smelyanskiy, and H. Neven, “Simulation of low-depth quantum circuits as complex undirected
graphical models,” 2018.
[4] C. Lomont, “Introduction to intel advanced vector extensions,” 2011.
Quantum simulations help explore different aspects of quantum [5] OpenMP Architecture Review Board, “OpenMP application program interface version 3.0,” May 2008.
[6] Message Passing Interface Forum, MPI: A Message-Passing Interface Standard Version 4.0, June 2021.
systems without the need for complex instruments. They are im- [7] E. Gutiérrez, S. Romero, M. A. Trenas, and E. L. Zapata, “Quantum computer simulation using the cuda programming
portant in evaluating the performance of algorithms like QAOA, model,” Computer Physics Communications, vol. 181, no. 2, pp. 283–300, 2010.
[8] J. Biamonte and V. Bergholm, “Tensor networks in a nutshell,” 2017.
VQE and also in understanding various quantum processes. [9] H. Manabe, Y. Suzuki, and A. S. Darmawan, “Efficient simulation of leakage errors in quantum error correcting codes using
Quantum computer simulators also help in designing efficient error Figure 4: Speedup comparison between Hybrid Devices vs CPU, GPU Devices
tensor network methods,” 2023.
[10] X. Yuan, J. Sun, J. Liu, Q. Zhao, and Y. Zhou, “Quantum simulation with hybrid tensor networks,” Physical Review Letters,
mitigation schemes by analyzing the designs using precise noise vol. 127, July 2021.
[11] T. Jones, A. Brown, I. Bush, and S. C. Benjamin, “Quest and high performance simulation of quantum computers,”
models through quantum simulations. They are used to tune The recent work [9] employed tensor networks to analyze effects of Scientific Reports, vol. 9, p. 10736, Jul 2019.

quantum machine learning models and debug quantum algorithms. leakage errors during quantum error correction in superconducting Figure 7: CPU versus GPU for Hamiltonian simulation circuit
[12] Y. Zhao, Y. Guo, Y. Yao, A. Dumi, D. M. Mulvey, S. Upadhyay, Y. Zhang, K. D. Jordan, J. Yang, and X. Tang, “Q-gpu: A
recipe of optimizations for quantum circuit simulation using gpus,” in 2022 IEEE International Symposium on
High-Performance Computer Architecture (HPCA), pp. 726–740, 2022.
Qiskit Aer, Cirq, Qulacs, QuilSim, CuQuantum are a few popular systems. The work [10] on hybrid tensor networks helps to simu- [13] Qiskit contributors, “Qiskit: An open-source framework for quantum computing,” 2023.
and easily available quantum computer simulation frameworks that late larger systems using relatively smaller processors. [14] X.-C. Wu, S. Di, E. M. Dasgupta, F. Cappello, H. Finkel, Y. Alexeev, and F. T. Chong, “Full-state quantum circuit
simulation by using data compression,” in Proceedings of the International Conference for High Performance Computing,
can be leveraged for research. Most simulators use a Schrodinger- Networking, Storage and Analysis, SC ’19, (New York, NY, USA), Association for Computing Machinery, 2019.

style simulation that models the interaction between gates and [15] R.-Y. Sun, T. Shirakawa, and S. Yunoki, “Improved real-space parallelizable matrix-product state compression and its
application to unitary quantum dynamics simulation,” 2023.
state vectors and is the focus of our work.