Curand documentation - contradictory statements about Mersenne twister

Reading about the documentation of curand library (cuRAND :: CUDA Toolkit Documentation) found some contradictory information:

“the widely-respected Mersenne twister [8] consistently fails two of the linear complexity tests.”

And a couple lines later there are the results of MT19937 “The MT19937 generator is, by far, the most widely used PRNG … All tests were passed”.

I might misunderstood something but can someone please help clarify these points?

Note that there are two different versions of Mersenne Twister PRNGs being discussed in the documentation: MTGP32 and MT19937.

The way I read it, the first quoted remark seems to apply to MTGP32, but I agree this is not entirely clear. I note that the CURAND documentation also says: The MTGP32 generator exhibits some marginal results on “BigCrush”.

Consider filing an enhancement request with NVIDIA to get the description clarified. Use the regular bug reporting form and prefix the synopsis with “RFE:” to mark it as an enhancement request.