This design is an "Infinite Noise Multiplier", a hardware true random number generator. The name was inspired by my daughter, who is a never ending source of randomness.
The idea is simple: there is noise in every system. All you have to do is amplify it. This switched-cap circuit has a voltage V, which it multiplies by 2 every cycle. If V is greater than Vref/2, then it subtracts Vref/2, thus keeping V between 0 and Vref/2. The truly random output bit is just the output of the comparator between V and Vref/2.
This is best understood digitally. If you have a random value between 0 and 1, say .1001110110001, then this circuit would multiply by 2, getting 1.001110110001X, where X is some new random noise. It would subtract out 1 because the value is > 1, and if X is 0, we get: 0.0011101100010. Continuing this process, we shift out the bits one at a time, while noise creates random data in the least significant bits. The circiut is insensitive to non-random signal injection. If you add a non-random signal to the value, you get different, but still random bits shifting out.
There is a bit more to it than this. The voltage followers can only drive down to about 0.1V, which is good, because it creates a voltage to start multiplying. The multiplier only multiplies by about 1.8 instead of 2, which is also good, for keeping the voltage from running away. We subtract slightly less than Vref/2 so that V is always between about .2V and Vref/2V. The voltage followers introduce significant distortion and voltage offset. All this results in non-perfect random output. From SPICE simulations it seems that there may be a 25-ish% correlation between bits (this requires more analysis), and a 10-ish% bias for more 1's than 0's. To get high quality random data out, I have to XOR many of these partially random bits together. Here is the math to support this approach:
Let b represent the maximum amplitude of a bias/correlation function between bits so that the probability of an attacker guessing correctly an output bit knowing all the surrounding bits is no higher than 0.5 + b. As I XOR these bits together, the probability of being able to guess the bit drops. Call the max amplitude of the bias/correlation when XORing n bits together B(n). Let P(n) be the probability of an attacker correctly guessing the output bit with n bits XORed together. Then:
P(n+1) = (.5 + b)(.5 + B(n)) + (.5 - b)(.5 - B(n)
P(n+1) = .5 + 2*b*B(n)
B(n+1) = 2*b*B(n)
B(n) = (2*b)^n*b
I am XORing 80 bits together. With b = .25, B(80) = 2e-25, which should be more than adequate for any crypto purposes. However, I still recommend that the output of this RNG be further randomized with a cryptographically strong pseudo random number generators or hash function, just to be sure.
I am implementing this design on a tiny .35u mixed-signal IC. It occupies about 0.04 mm^2, and runs at 8MHz in SPICE simulations. 25 fit per mm^2, cnd each consumes about a milliwatt. Each outputs 8Mbit/sec, but XORing 80 bits together reduces it to 100Kbit/sec. Per mm^2, it should generate 2.5Mbit/sec of crypto ready random bits. This is a very low power process. In a high speed process, it should be possible to run at 100MHz instead of 8MHz, and the circuit should shrink dramatically.
Here are my awful hand-drawn schematics. The first schematic shows that on phase 1 we sample the voltage from the previous cycle, and store it on two hold capacitors in parallel. An NMOS transistor is used as a switch from the buffer of the previous stage. During phase 2, we use three more NMOS switches to stack the two capacitors, which in an ideal world would multiply the voltage by 2, but because of parasitics, it only does about 1.8. The output of the stacked capacitors feeds through a voltage-follower buffer into the next stage. Here there are two switchies: one to feed another hold capacitor during phase 2, and another to feed an almost identical hold capacitor also in phase 2. The first hold capacitor feeds the positive terminal of a comparator, while Vref/2 feeds the negative. This is a clocked comparator that compares its inputs on the start of phase 1. If the held voltage is greater than Vref/2, the comparator subtracts a charge from the other hold capacitor that causes its voltage to drop ideally by Vref/2. Due to parasitics it's somewhat less than this. This second hold capacitor where the subtraction occurs then feeds into another voltage follwer buffer, which loops back to where we started, so the process can repeat. The voltage follower schematic is also shown, and is simply 4 mosfets: two for current sources and a matched diff pair. It is a standard source-coupled voltage follower. Vref/2 is generated with a simple resistor divider. The two phases of the clock are generated with back-to-back NOR gates, which is also standard. The clocked comparator is a standard circuit as well, which is based on two inverters feeding back on each other, with some mosfets to reset the inverter outputs to 1 when the input clock is high, and when it goes low, the inverters are connected, and will flip one way or the other. Two mosfets driven by the input voltages cause one side or the other to win. The bias voltage generator for the buffers current sources is also shown, which is a resistor feeding into a mirror.
Here is a simulation of 80 clock cycles, showing typical random data that would be XORed together to generate one output bit. The random output is Q. The bottom signal is a 1MHz sine wave which is coupled into the hold cap after buffer1 through a capacitor 10% as large as the hold capacitor. The point is to show that non-random sources of noise do not cause the output to be less random. This circuit is highly insensitive to external signals.
The main sources of noise in this circuit are primarily shot noise and thermal noise. Shot noise is generated every time an electron is accelerated through the input transistors of the voltage followers. Thermal noise is generated in the bias current resistor. However, the source of noise in the SPICE simulation is the numerical errors in the simulator, which dominate over the quieter shot and thermal noise. The actual source of noise should not make any differece. Every analog signal has noise. It simply has to be amplified for us to use it.