I've tried to diagnose when the RDRAND failures happen. From my experiments on Haswell, it only seems to happen when you have > N threads, where N is the number

I've tried to diagnose when the RDRAND failures happen. From my experiments on Haswell, it only seems to happen when you have > N threads, where N is the number of physical cores, simultaneously reading continuously from RDRAND. But when that happens, nearly every read from every thread will fail. This means a user process could easily deplete RDRAND's contribution to the kernel's entropy pool. So encouraging all kinds of application code to use RDRAND is really bad advice. It should at least be a privileged instruction.