What I don't like about this "guide" is that they stick to the same numbers for latencies. They should stick to the same "time", not "cycle numbers".
CAS time is dependent only on memory voltage thus JEDEC speeds will have higher CAS times
than XMP which is overvolted.
- CL15@2133 15/(2133/2)ms = 14ns (in this test)
- CL15@4000 15/(4000/2)ms= 7.5ns (in this test)
- CL8@2133 8/(2133/2)ms = 7.5ns (should been tested considering CL15@4000 runs)
- CL28@4000 28/(4000/2)ms = 14ns (or this should been tested if CL15 was best for 2133)
So for a "speed-only" comparison, CL should have been adjusted for each speed to be just above 7.5ns.
Just searched for "kingston DDR4 4000", and here are timings of HX440C19PB3AK2/16:
•JEDEC: DDR4-2400 CL17-17-17 @1.2V == CAS 14.1ns
•XMP Profile #1: DDR4-4000 CL19-21-21 @1.35V == CAS 9.5ns
•XMP Profile #2: DDR4-3600 CL17-18-18 @1.35V == CAS 9.4ns
See how 4000 has +2CL? CL18@4000 is 9ns.
Since you need to test high-speeds, then all tests would be with higher voltage. Thus 2133 would not be penalized so much.
The only real-world "benefit" is that whoever stays at 2133/2400/2666 will use JEDEC 1.2V and anybody who goes above 2800 will be 1.35V or more. But then again, who stays @JEDEC will not be looking at these graphs and/or know how to get to 4000. But even then, slower speeds are artificially penalized as JEDEC entries will have different timing numbers for different frequencies.