The usual circuit (found widely across the internet) for a true random number generator's (TRNG) entropy source is a reverse biased transistor like so:-
Q1 is typically a 2N3904 as it has been found to be quite noisy the wrong way round running in breakdown mode. +ve is typically in the order of 11V. This topology works well initially.
I came across someone who had build a TRNG and experienced problems whereby "after about 3 months of use the noise signal would drift". I've seen other references to long term degradation of such a topology but can't seem to locate them at present. I have the sense that continuously shoving electrons the wrong way up a transistor might be detrimental, but have not found any references to prove this one way or another. After all, this is not the principal use case for a transistor. Could the junction suffer gradual and irreversible damage thereby altering it's noise characteristics? Zener diodes are another situation altogether as breaking down is their modus operandi.
Three references on E.SE discuss reverse biasing, but not long term stability:-
Understanding reverse biased PN junction
Does a reverse-biased P-N junction create quantum noise?
Is my avalanche noise “random?”
Is a reverse biased transistor stable in the long term?
Best Answer
No they are not stable.
The current gain Ic/Ib of bipolar transistors strongly decreases when the oxide over the emitter-base junction is damaged. The results obtained with different stress conditions lead to point out two degradations processes: the main process consists in the increase of the recombination in the vicinity of the space-charge region in the emitter-base junction.
Ref DEGRADATION OF JUNCTION PARAMETERS OF AN ELECTRICALLY STRESSED NPN BIPOLAR TRANSISTOR
N. TOUFIK, F. PILANCHON and P. MIALHE*
The above report was based on the 2N2222 which has a lower reverse breakdown threshold than the 2N3904 and also more noise voltage, but perhaps faster degradation. I have yet to find a report on the 2N3904 so an uncontrolled avalanche current is unstable with a small resistor
Ref http://holdenc.altervista.org/avalanche/index.html
The above uses a constant current source unlike Paul Campbell's rng2.0 which relies on avalanche current in base * hFE limited by Collector R = 3.3k to 20V so if hFE is only 100, diode can be (20V-1Vsat)/3.3k/100(hfe)= 58uA which would be excessive.
Given the initial question design saturates the noise into a digital signal and noise can vary in current with temperature as well as spectral bandwidth it is not stable and the threshold above 10uA where failure rapidly degrades is unknown, limiting the reverse current to the Mfg's test criteria for Vr would make most sense.
The next question is how stable does it need to be? What is life span criteria and repeatability to NIST scores? What are the test results over ambient range?
Increasing the coupling capacitor size may increase the max number of consecutive 1's, a figure of merit by extending the number of decades of frequency range at the expense of startup delay or sensitivity to supply regulation drift.