I wanted to re-visit this thread after a rather lengthy learning experience about memory timings. I followed a memory OC guide which was provided to me by a user on the overclock.net forums. 
 
https://github.com/integralfx/MemTestHelper/blob/master/DDR4%20OC%20Guide.md
 
I'm currently on a Asus Maximus XI Gene and my 9900KS is running at 5.3Ghz core, 4.9Ghz caches at 1.480v (BIOS) at LLC 5. Ram is 4300 Mhz at 17-17-17-36 at 1.415v with 1.2500v for VCCIO and VCCSA. The CPU has been de-lidded, die and water-block have been lapped and I'm using a direct die kit with liquid metal TIM. This will pass 8 hours of RB 2.56 with a peak core temp in the high 80c range. 
 
Using my G.Skill's 4266Mhz memory XMP profile, LinX will hit about 330 GFLOPS. With the memory timings that I worked up, it will hit ~553 GFLOPS with a TON more power draw. Peak was 305 watts and 237 amps. Peak temp on my hottest core was 89c. If I didn't have all of the cooling hacks in place that I do, I'm quite certain it would have temp throttled. I can get matching residuals with XMP timings but not with my timings. So there is that. It seems LinX and LinpackExtreme are very sensitive to memory timings as was revealed previously in this thread. 
 
I think this also shows XMP timings aren't the best for performance. Simply turning on XMP, in my case for sure, did not get me the bandwidth that a 4266mhz dual channel kit should. This is very likely due to loose timings in the XMP profile to gain more compatibility. I had to manipulate my timings to get in that 90-95% of my max theoretical with good latency. 
 

Ambient was about 23c. It been getting warmer here so my open window days are numbered.
 
Does the bar in LinX keep flashing as if its progressing? Does LinpackExtreme do the same; only run once then freeze? 
 
It took me a while to work up those memory timings, I don't know if they are ideal for your system. I followed the guide I linked and used Super Pi 8M to verify performance was improving with changes. Then I used TM5 with the extreme profile to verify it was error free. I wish I would have kept my noes as I progressed. There were some timings that made big differences, I just can't remember which. On Asus boards you can change the training modes and when I switched to the one that tightens RTL and IO-L it made a surprising difference, at least in Super Pi 8M. 
 
What do your bandwidth and latency numbers look like? 
 

Nice. I feel good about the timings. Definitely near that 95% goal of max theoretical bandwidth. 
 
I may try to see if I can get it to run faster but I feel the way they bin these things now is that if it was going to be faster than 4266Mhz, they would have sold it that way.