RDRAND does have non-determinism. The problem is that you can't directly measure or monitor how much non-determinism. The values you get from RDRAND come from a DRBG that is periodically and automatically re-seeded from a NRBG, but if you call RDRAND many times in rapid succession, then the DRBG may not have been re-seeded between each invocation. RDSEED differs only in that it guarantees that the DRBG is re-seeded between invocations. If you are using the instructions infrequently enough, then the two are functionally identical, but "infrequently enough" has a value known only to Intel.
This comment in linux/arch/x86/kernel/cpu/rdrand.c claims:
/*
* Force a reseed cycle; we are architecturally guaranteed a reseed
* after no more than 512 128-bit chunks of random data. This also
* acts as a test of the CPU capability.
*/
