The effort involved in the implementation of RdRand instruction to promote its use in security-related applications, in particular the self-testing functionality, is very impressive!
I'm interested in cryptographic applications of the RdRand instruction. Sometimes I might want an entropy source which is unaffected by deterministic processing which could potentially compromise my cryptographic applications. In particular, I am concerned that the Online Health Tests (OHTs) will affect the output of the RdRand instruction. Suppose the entropy source is in fact working properly. The ES will sometimes fail the OHT even though the entropy source is fine. That means that ES bit streams that fail the OHT will never be seen. Hence certain output sequences which would be generated by a perfectly random entropy source will never appear in an output sequence generated by the RdRand instruction. If I understand the documentation, there's a 65536 bit sliding window which is used for the OHT. 256-bit chunks output from the OHT are then used to create up to 1022 64-bit RdRands. It then seems if I only use one out of every 1022*65536/256 = 261,632 values returned by RdRand, each 64-bit sample selected will be outside the sliding window of all the other 64-bit samples selected.
Is that in fact the case? If I use only one out of every 261,632 calls to RdRand, will I in fact get 64-bit samples from non-overlapping sliding windows used by the OHT?
Even when the RdRand outputs I use are from non-overlapping sliding windows, it would still be helpful to know the exact algorithm used by the OHT to see if that might affect my cryptographic applications. Is documentation available for the algorithm used by the OHT? If not, I would like to request that.