Performance

Storage offloading

Large statevectors don't have to fit in DRAM. QXel can offload the statevector to CPU memory or secondary storage (NVMe/HDD), letting a single node reach higher qubit counts.

offload_type controls where the statevector lives: • 'none': keep it in GPU/CPU memory (default) • 'cpu': offload to CPU DRAM • 'storage': offload to secondary storage devices (requires path)

For storage offloading, pass the device paths. Multiple devices are striped RAID0-style for bandwidth, and both file names and device names work.

python

result = qxel.run(
    circuit,
    shots=1000,
    compute_type="cuda",
    offload_type="storage",
    path=["/dev/nvme0n1", "/dev/nvme1n1"],
).result()

Note You must pass path when offload_type='storage'. Faster drives (NVMe) and more of them improve throughput.

← GPU acceleration Distributed simulation →