QXel

The quantum simulator built for HPC.

QXel is QubiStack's high-performance, GPU-accelerated quantum circuit simulator. It pushes a single node past the usual memory wall, runs on the SDKs you already use, and scales from a laptop to a multi-node GPU cluster.

Browse the guides

42

Qubits on a single node

3

Simulators (state vector, density matrix, stabilizer)

3

SDKs (Braket, Qiskit, PennyLane)

What QXel gives you.

42-qubit single node

Conventional simulators max out at 30 to 35 qubits per node. Secondary-storage offloading pushes QXel to 42, at a fraction of the memory cost.

Multi-framework

Bring your existing code. qiskit, pennylane, and the AWS Braket SDK all run on QXel through a single Local backend.

GPU & distributed

CUDA and cuQuantum kernels for GPU acceleration. MPI-based distributed simulation across multiple nodes for the largest circuits.

Zero-setup cloud

Submit a circuit to a managed GPU worker and get the result back. Job history and results are all handled for you.

How it works.

Storage offloading
Use NVMe or HDD instead of expensive DRAM to scale qubit count without scaling memory cost.
GPU kernels
Hand-tuned CUDA gate kernels and cuQuantum integration for maximum throughput per device.
Distributed simulation
MPI-based multi-node execution, in single-process-per-node and single-process-per-device modes.
Gate fusion
Fewer kernel launches, less memory traffic, faster runs.
ghz.py
from braket.devices import LocalSimulator
from braket.circuits import Circuit

# 3-qubit GHZ state
circuit = Circuit().h(0).cnot(0, 1).cnot(1, 2)
circuit.probability()

# GPU kernels + statevector offloaded to two NVMe drives
qxel = LocalSimulator(backend="QXel-sv")
result = qxel.run(
    circuit,
    shots=4096,
    compute_type="cuda",
    offload_type="storage",
    max_fusion=2,
    path=["/dev/nvme0n1", "/dev/nvme1n1"],
).result()

print(result.values)

Tune every run.

Pass keyword arguments to run() to choose kernels, offloading, and gate fusion.

compute_typecpu · cuda · cuquantum

Pick the kernel set per run.

offload_typenone · cpu · storage

Offload the statevector to DRAM or secondary storage.

max_fusionint (default 1)

Max gate width for fusion; fewer kernel launches.

pathList<string>

Storage devices for offloading, striped RAID0-style.

Result types

AmplitudeExpectationVarianceProbabilityStateVectorDensityMatrixSample

Ready to run on QXel?

Read the docs to install QXel and run your first circuit, or explore the guides and references.

QXel guidesQXel SaaS guides