The niche for fast HWRNGs actually exists in the stochastic simulation business (physics, statistics, machine learning, finance, to some extent CG); "real" randomness helps in scaling over tons of cores and specialised hardware must be in principle faster than general hardware executing software PRNG.
The bar is set high, though; the dream is about 1GB/s, RDRAND achieves 500-700MB/s.
>The niche for fast HWRNGs actually exists in the stochastic simulation business
>The bar is set high, though; the dream is about 1GB/s, RDRAND achieves 500-700MB/s.
Why are PRNGs not good enough? A hardware implementation of algorithms in the SHA series can achieve multiple GB/s of throughput. With a good seed, wouldn't the output of a simple counter-based or iterative hash-based PRNG be sufficiently random for simulations?
Yes, PRNGs should be good enough, and PRNGs that are supposed to be cryptographically secure should definitely be good enough. (A PRNG that is not cryptographically secure still could be good enough for simulations.) Any method to distinguish a PRNG from true randomness would be the basis of a cryptographic attack on the PRNG. In fact, PRNGs are better for simulations because it allows the results to be reproduced exactly from a small seed.
The bar is set high, though; the dream is about 1GB/s, RDRAND achieves 500-700MB/s.