I think any of you guys running SK Hynix RAM at 3200 should run through Memtest86 overnight. I have CMU16GX4M2C3200C16s (ver. 5.39), and it POSTs and boots up fine at 3200 on 1.40 stable; can even bench it, play games, etc. (1700x @ 3.8GHz, 1.26v vCore, 1.0v NB/SoC)
However, when I ran it through Memtest, there were some 1-bit errors in test 7 (81 of them, in fact). Test 7 is: [Moving inversions, 32-bit pattern]. This occurs at any RAM voltage up to 1.45v and any NB voltage up to 1.2v. I wasn't willing to go any higher than that. 1-bit doesn't seem like much, but when you're doing millions of operations, it adds up.
So, while it does work, you may eventually corrupt some data in the long run and you might see random application crashes from time to time if you have similar memory errors. Some don't care as long as it boots. That's fine too, but I don't consider that stable. Many of the timings still aren't being set correctly too, so some RAM ICs might be more tolerant of that. Mine aren't, at least at 3200.
tRC is higher than XMP spec at 75 (3200 spec is 54)
tRFC is set extremely aggressively at 312T (3200 spec for my modules is 559T)
I re-ran Memtest at 14-14-14-28 at 2933 (0.850v NB/SoC, 1.35v RAM), and there were 0 errors in all tests (interesting, as the XMP timings for 2900 are 15-17-17-34). My 1700x's IMC seems tolerant of higher speeds without much of a voltage bump.
Here's an example of the memory errors in my Memtest logfile:
2017-04-14 11:01:30 - [MEM ERROR - Data] Test: 7, CPU: 15, Address: 4AF94EC4, Expected: FFDFFFFF, Actual: FFDFFF7F
2017-04-14 11:01:30 - [MEM ERROR - Data] Test: 7, CPU: 3, Address: 1A503744, Expected: 00000100, Actual: 00000180
2017-04-14 11:01:30 - [MEM ERROR - Data] Test: 7, CPU: 12, Address: 3EAC5344, Expected: 00400000, Actual: 00400080
It's always just 1-bit out of place. So, it can go undetected in normal usage until things start acting weird. Same behavior as the 1.41 beta.
YMMV, as always with silicon.