Whichever brings the lowest final latency in aida64 without losing significant bandwidth.
Just be aware that the test is sensitive to background processes so you will likely need to run it a few times for a rough average.
Whatever the highest frequency is that you can get stable with tRDRD_dg & tWRWR_dg set to 4.
I kind of stumbled into this same observation. As I went up in speed with looser timings (like from 4000MHz to 4133MHz), I didn't get faster AIDA64 memory benchmark results at the faster speed until I got tRDRD_dg & tWRWR_dg down to 4 again.
[email protected] (-1 AVX offset) on ASRock Z390 Taichi with Noctua NH-D15 air cooler
CPU offset voltage of -25mv, runs VRVout 1.240-1.313V on full AVX load, 1.225-1.275V on non-AVX load
2x8GB [email protected] at 1.45V, G.Skill F4-3733C17Q-32GTZKK (XMP rated [email protected])
EVGA GTX 1060 6GB OC with Corsair RMx 750W power supply
Samsung 970 EVO 500GB NVMe boot SSD and four other drives all in a Fractal Design R6 Case