Exhausted

This one:

I don't know the transistors, but have you tried a low voltage zener. (The problem with both is that you may not get much more than just the shot noise.)

George H.

Zener is OK but my experiments suggest noise is about an order of magnitude lower. On the plus side transistor junctions seem to 'age' and become quieter as time goes on. (both effects may be non-existent and just the result of my poor experiments)

Agreed, the zener noise mechanism comes from avalanche, which doesn't kick in until you get to voltages above about 8 volts. Low-current "microplasmas" form and quickly terminate after discharging the zener's self capacitance, creating a kind of popcorn noise. There are extensive threads on the subject 15-20 years ago on s.e.d., where I spent a month (haha, more) making and posting all kinds of nanosecond-scale measurements.

The zener physics effect (field emission) predominates at low-voltages, see the ON Semi handbook, HBD854, page 20.

Grin.. everyone should spend a month playing with avalanche breakdown.

Say Win, If you want to "waste" a few more hours with avalanches, I can send you some LED's that breakdown ~25 Volts and are single photon detectors. (Spad's) They make great little experiments for electronics courses. (perhaps Paul would be more interested?) These.. you can also get them from Newark.

A newsletter.

George H.

One suggestion: 1.2-volt Vbe voltage references are more noisy than resistors and transistors, etc., due to the 10x voltage step-up intrinsic to their design. If the reference's circuitry runs at low currents, it's even more noisy. The LMV431 with a 50uA core has 230nV/rt-Hz noise. The LM385-1.2, with its core under 10uA shows 400nV/rt-Hz. The REF1112 that runs at 1uA should be even more noisy. The lower-current types will suffer from limited frequency response.

Although the noise level is not specified, the basic physics behind BJT transistor noise at low currents (see AoE III, page 481 and especially Figures 8.12, and 8.57) means you may get much better part-to-part consistency than you could get with a zener diode.

David, if you try this, let us know what you learn.

Awesome!

We had a reverse-biased BE junction noise generator in an old product fail after several years in the field. The problem was that the statistical tests on the entropy source failed yet the DC measurements on the transistor seemed fine when a technician looked at it. The stats came good once the part was changed for a new one.

I guess one shouldn't expect too much when using a device outside its abs. max. ratings.

Allan

They would make a pretty good noise seed, probably even in the dark.

Well maybe, probably not in the dark. About 10-20% of them show a lot of after pulsing... breaking down almost all the time even in the dark. (After the channel breaks down, there are dangling charges left around and those can cause it to break down again.) So there is not a random distribution of pulses, but a large chance of having another pulse soon after one has happened. That shows up in the statistics.. you get a kinda 1/f distribution of times and not the expected exponential.

George H.

Hmm I wonder if an avalanche zener would have been better, at least they are made to break down all day. I guess if you needed speed (higher noise bandwidth) that the capacitance of the zener might have been a problem.

George H.

A random number seeder needs to be random but need not have good statistics. But once you have that much voltage available, you may as well use a zener.

An opamp with a lot of closed-loop gain will generate a ton of noise, with an awful 1/f type spectrum, but again that can be a fine random-number seed, used to scramble/hash some deterministic pseudo-random thing.

Good flat analog noise is best derived from a linear shift register, or from a zener, or both. It's easy to make a pseudo-random shift register thing that won't repeat for a thousand years.

Some uPs include a mess of ring oscillators to generate random numbers; we were just talking about that, specifically the temperature coefficient of ring oscillators inside an FPGA.

I used the shorted microphone input of a cheap sound-card (at full volume) to get random bits to stir into the 64kbit entropy pool in one of Europe's first Internet banking applications (SSL, but 1997, before web browsers were really usable for this).

Actually, for strong security, the goal isn't actually randomness, it's unpredictability. Doesn't matter much what data you use, as long as the bad guys can't guess it.

Using an online source would be terribly, terribly stupid.

I wonder if there is a public random data generator somewhere on the internet. You could use ping times or something as a source of randomness.

Any online webcam should provide a lot of noise.

I think you'd find that the entropy available from ping times is rather poor. There are more sources of correlation than you might expect.

One classic approach was to aim a webcam at a Lava Lamp, or one of the similar amusements that were filled with a glitter-loaded liquid. Taking two frames several seconds apart, and calculating the difference, produced some reasonably decent entropy.

Getting really *good* random entropy is surprisingly difficult, especially if you're concerned about having it *stay* random in the face of attacks, outside interference, "pulling", etc. which can make it more predictable.

Den torsdag den 17. marts 2016 kl. 01.14.48 UTC+1 skrev Dave Platt:

remember someone using a webcam and the Americium-241 source from a smoke detector

-Lasse

Well, you would certainly use it to seed some other, private random number generator, not just as-received. And a web site could certainly serve up a different block of bits to every request.

A busy city street scene should be pretty noisy. Any camera will have lots of per-pixel noise, and lots of pixels.

If you have a long-length (say, something like 256 bits) pseudo-random number generator, you can mush in some external randomness source (like a webcam, a sound card, a ring oscillator, whatever) and make it randomly jump all around its state sequence. Do that 64 times with different register lengths and noise sources and xor.