https://www.amd.com/system/files/TechDocs/amd-random-number-...
> "The RNG uses 16 separate ring oscillator chains as a noise source. Each chain consists of a different prime number of inverters in a free-running configuration which capture natural clock jitter"
So it's (deliberately) dependent on natural manufacturing variation in the silicon, and in any deterministic simulator won't produce random numbers. It also takes a while to start; it's possible that the startup conditions are bad such that "a while" turns out to be "forever".
The problem is as it always is with these kinds of things. They help assure that you can measure your model and your implementation of that model. They're much less capable of determining you've got the right model to begin with. That part is still up to the humans.
The state of commonplace capabilities for testing and verification of software is extremely far behind that of hardware. For a variety of economic, cultural, and use case incentives reasons.
I remember reading somewhere that, for a long time, Intel's "test cases" basically consisted of all the operating systems they could find, going back to MS-DOS 1.0 and including some more obscured ones, as well as tons of other third-party software.
Unfortunately it seems this new generation no longer follows the "NEVER break backwards compatibility" principle that made the PC such a great platform, and as a result we're seeing hardware/software/firmware that's increasingly buggy and divergent. It used to be the case that if a new processor couldn't run existing software (even that which depended on undocumented but constant behaviour) unmodified, it was a fatal flaw.