The reason people are concerned about the
RDRAND instruction is that it is a proprietary hardware random generator and has never been formally analyzed. Without that analysis, there's no way to tell if it's a genuine hardware random number generator or, for example, a simple counter being passed through a cryptographic hash. In the latter case, the output will look random, but someone who knows the starting point of the counter knows exactly what numbers will be generated.
Anyone can look at the Linux source code for
/dev/random and confirm that the kernel is mixing
RDRAND output into its entropy pool in a way that doesn't reduce the amount of randomness available, even if
RDRAND is hopelessly biased.
Whether you can trust
RAND_bytes or not is a judgement call you'll have to make for yourself, but what I can say is that running it on a Celeron with
RDRAND is no worse than running it anywhere else.