The new Intel random number generator, code name "Bull Mountain", starts from thermal noise at 3 GHz, conditions the bits, and then uses these numbers as seeds to generate 511 pseusorandom numbers per true random number. It is this last step that I dislike and don't trust.
--Aha. However, this last step probably will be good enough to assure Intel bits pass all the usual test packages like TestU01 (unless they've been total idiots... which they might have been). It likely will destroy cryptographic security, but you could just only take 1 number per 512 to get that back. If Intel is doing this I don't see why they needed 3 GHz, why couldn't (3/512) GHz have sufficed?
Thermal noise has power 4kT per unit bandwidth. If you're building your own circuit, 10 MHz is a reasonable bandwidth, and at temperature 300 K, the noise is 1.6e-13 W. With a 1 Mohm input impedance, and using P = V*V/R, the RMS noise voltage is 400 microVolt.
--You already blew it because the input capacitance of typical op-amp pin is 4 pF, and R*C for R=1megaOhm and C=4pF is 4 microsec, so you just cut off your "10 MHz" bandwidth right there.
An amplifier gain of 2500 brings this up to a convenient 1 V, and this can be done in two or three stages using op-amps, with AC coupling between stages to avoid amplifying the op-amp bias voltages.
--Hi-gain amps are vulnerable to oscillation from tiny parasitic capacitances coupling output and input. You can try to avoid using fancy shielding, but, why even go there if you don't have to? This is especially worrying when your amp is high frequency (10 MHz is pretty high and with several cascaded amps you maybe need 3X more speed). Most commercial op-amps are designed for lower frequencies, or at least this used to be true. By the way, if you really want to take the "big amp" approach I disparage, a better source of noise is a neon light. Lots more than thermal noise in a resistor. --Intel and your schemes both seem vulnerable to crosstalk from nonrandom noise, i.e. from other signals within your computer (if phsyically located within your computer, e.g. on the processor chip) Your "random" bits would then be highly "predictable" to the right predictor. Again you can try to shield, but why go there if you do not have to?