On Sep 5, 2021, at 3:23 PM, John Rose <john.r.r...@oracle.com<mailto:john.r.r...@oracle.com>> wrote:
To increase throughput use vectors or generate more than one random sample per crank turn. But back to back aes steps are probably always twice the latency of a single wide multiply. So I think there might be some more room for cleverly using single aes rounds, say two in parallel with 1-cycle input transforms. Put simply, engineers and academics approach novelty differently. (Yikes! I typed the mail on my phone and somehow that last sentence went to the wrong place. It should have been at the end of the PP that speculated about differing motivations surrounding publication and engineering.)
