In what circumstances did qsim return incorrect results in 22.11?

[eliottrosenberg]

I saw that the release notes for cuQuantum Appliance 23.03 say "Fixed an issue that qsim can return incorrect results when multiple GPUs are used for simulation."

Under what circumstances were the incorrect results seen? I used cuQuantum Appliance 22.11 with qsim, using multiple GPUs, to take simulation data for an experiment and want to make sure that I don't need to go back and check the quality of the data.

Thanks!

[leofang]

Hi @eliottrosenberg, it refers to a delicate synchronization bug that would not be easily triggered unless a frequent data exchange between GPUs happens, which is problem dependent.

The bug was fixed in part with cuStateVec v1.3.0 (included as part of the 23.03 release). The relevant entry is:

Fix for an issue of custatevecMultiDeviceSwapIndexBits() that CUDA calls are not correctly ordered on streams allocated on multiple GPUs.

[eliottrosenberg]

The results matched what we expected from studying earlier cycles without using multiple GPUs.

[eliottrosenberg]

Do you happen to know whether this is a circuit that would be affected?

[eliottrosenberg]

I just re-ran some of my 36-qubit circuits in cuQuantum Appliance v23.03, and the results are consistent with what I found using v22.11.

[mtjrider]

Sorry for not responding sooner and thanks for rerunning your workload!

I would have been surprised if your results were affected given the circuit's local structure.

This is an excellent datapoint for us. Very much appreciated.

[eliottrosenberg]

@leofang @mtjrider Thanks for your help with this! The project that I was referring to is now available as an arXiv preprint: https://arxiv.org/abs/2306.09333. I'm happy to answer questions about this work or present it to your group if you're interested.