Potential GPU and CUDA support?

[zazabap]

Feature details

If I am understanding the number of qubits available for emulation correctly, the qubits number largely depend on the RAM size available. Since more and more people may have access to GPU computing cluster (some Nvidia GPUs, RTX 3090TI, A6000, H100, B200), is it possible to shift some computation unit to GPU or providing GPU version of computation?

Implementation

I found possible solution from cupy to replace the numpy wrapper. Also the previous issue mentioned pytorch for tensor computation. For more details, overall review of the code structure might be necessary.

How important would you say this feature is?

2: Somewhat important. Needed this quarter.

Additional information

No response

[jzaia18]

Hi @zazabap thank you for opening an issue! Have you looked into PennyLane-Lightning yet? I believe it has the features you're looking for.

These links may be of interest:

• https://pennylane.ai/performance#gpu-enabled
• https://docs.pennylane.ai/projects/lightning/en/stable/index.html

Though, the default.qubit jax and pytorch interfaces also work with GPU data and (depending on the type of problem) may be a better fit.