repository·main·Indexed 20 days ago
https://github.com/quantumlib/qsimA high-performance state-vector (Schrödinger) quantum circuit simulator for C++ and Python (via the qsimcirq package). It is optimized for speed using gate fusion, vectorization (AVX/FMA), multithreading (OpenMP), and GPU support. qsim can be used as a simulation library for Cirq and supports various GPU backends, including native qsim, NVIDIA cuQuantum SDK/cuStateVec, and the cuQuantum Appliance for multi-GPU workloads.
Depending on the simulator chosen, you can perform the following operations:
| Operation | qsim | qsimh | MPS |
|---|---|---|---|
| Determine final state vector | Yes | No | No |
| Sample results | Yes | No | No |
| Calculate amplitudes for specific bitstrings | Yes | Yes | No |
| Simulate low-entanglement circuits (approximate) | No | No | Yes |
Before setting up GPU-based simulation, decide which of the three available backend options fits your needs:
The impact of noise on simulation performance depends on the error types (decoherence, depolarizing, coherent, or readout errors) and how they are represented using Kraus operator formalism:
Multinode simulation is effective when your workload can be parallelized:
The qsimh applications (e.g., qsimh_base.x, qsimh_amplitudes.x) use a hybrid Schrödinger-Feynman method. The lattice is split into two parts using a cut. The gates on the cut are divided into three segments:
-p): Gates applied before the first checkpoint. There is no summation over these paths.-r): Gates applied between the first and second checkpoint.Total gates on cut - p - r).A two-level checkpointing scheme is used to improve performance. The full summation over all paths for the root and suffix gates is performed.
Key Configuration Concepts:
-k): Defines how the lattice is split. For optimal performance, split the grid into roughly equal parts with the fewest cuts possible.-w): Specifies the path for the prefix gates (bit-shifted path indices). This is essential for distributed execution.The MPS simulator is a memory-efficient alternative to the standard state-vector simulator. Instead of storing an exponentially large vector, it represents the quantum state as a chain of tensors (one per qubit).
bond_dim): Controls the amount of entanglement the representation can capture. A higher bond dimension increases accuracy but uses more memory and computation time.bond_dim, the simulator uses Singular Value Decomposition (SVD) to truncate the state, keeping only the top bond_dim singular values.qsimcirq.The f parameter defines the maximum number of qubits allowed per fused gate. Users should experiment with different f values to optimize their specific simulation setup.
f=2 or f=3 may be optimal depending on circuit structure.f=6: A value of f=6 is rarely optimal.qsim is not specifically optimized. The performance is often dominated by the qsimcirq translation layer overhead. This overhead can be significant enough to mask the expected $2^N$ runtime growth.Simulation runtime in qsim scales based on the number of qubits ($N$) and circuit depth. Beyond 20 qubits, runtime grows linearly with circuit depth.
You can estimate the memory required for an $N$-qubit circuit using the following rule of thumb:
${\text{memory required}} = 8 \cdot 2^N \text{ bytes}$
Additionally, because qsim performs best when utilizing the maximum number of threads, it is recommended to use high-bandwidth memory (above 100GB/s) for multi-threaded simulations.
The qsim repository provides three distinct simulation engines, each suited for different hardware constraints and circuit types:
The qsimcirq interface provides several automated features for Cirq circuits:
qsim gate set using Cirq's decompose operation. If a gate cannot be decomposed, qsim attempts to parse it as a raw matrix if one is provided.QSimCircuit objects support parameterized gates. Values are assigned using Cirq's ParamResolver during the simulation process.