>> Yes :-) It's not hard to sample the thermal noise, the problem is how to
>> sample it billions of times per second with enough resolution to be
>> useful.
>
> Well, no, you typically sample it a few times a second and use that to
> seed a normal PRNG.

But will that not cause the same problems seen already today with 
unrandomness? If the pixel shader for every pixel gets exactly the same 
random seed value, and just uses a PRNG to modify it, won't you see 
patterns? I would have thought every pixel needs its own unique random 
seed value every frame to really get rid of those sort of PRNG patterns. 
If you want a *truly* unbiased ray tracer for example.

> In other news, Intel added an op-code to generate truly-random numbers
> using exactly this technique. The Linux kernel team added it to
> /dev/random... and then took it out again, because they were apparently
> concerned that the NSA would modify the die of your CPU to make the
> random numbers non-random. (!!)

Apparently the on-chip thermal noise sampler outputs a stream of bits at 
3 GHz. I don't know how big the hardware is for that circuit, but it's 
not *that* far away from what you'd need to get a true random number for 
every pixel processed on the GPU.

> Jesus, that's next-level. Surely if you don't trust the hardware your OS
> is running on, it's already game over. (?)

Yes, something like a true RNG in every home home PC seems like 
something the NSA would be extremely interested in.

Post a reply to this message