Let's spill some beans. Here is all I know about geekbench3 feel free to join in.
1. The best scoring versions for intel CPU's with 16 cores or less is 3.1.5 get it here https://www.dropbox.com/s/4q0xuvioq8opx30/Geekbench-3.1.5-WindowsSetup.exe?dl= (New version 3.4.3B3 here has potential. Need more testing. https://geekbench.s3.amazonaws.com/Geekbench-3.4.3b3-WindowsSetup.exe )
2. Best scores are with no microcode bios on z490. You will see e-penis scores with microcode because its boosts the mem sub-score but the total score is around 300 points less at 5.3ghz on 10900k. 53600 with MC, 53900 without MC. 10k mem score doesn't matter when you are losing so much total score. Obviously not calling anyone out, have fun if you want to play mem score. It is fun for sure.
3. For intel HEDT make sure to disable Spectre and Meltdown when possible. (No need on AMD it actually scores worse when you mess with it for some reason.)
4. In most cases uses the newest version of windows 10 available. Every consecutive version of windows has incrementally increased my scores especially the AES tests scores.
5. Pros are changing their clocks during the run. DJ/LUA are difficult as well as DFFT on current gen intel, as well as raytrace on AMD so, dropping clocks during those sections can help you get much higher scores. Say for example 10900k I can run 7ghz for the entire run maxed out highest that will pass DJ/LUA, I can also run the first section of tests at 7.125ghz and the end even higher 7.2ghz+ For example start geek at 7.1ghz run to 13s drop to 7.0 for DJ/LUA wait for the 18s mark and clock back to 7.1ghz If you want even more performance the last 2 seconds are the memory tests which have no load on the cpu and pass very high 7.2 seems fine here. You will only be limited by your mem stability and the limit of bclk you can move in OS usually around 3 bclk. This also makes the "CPU mhz" field on ln2 subs irrelevant as they are changing during the test. Is it 7ghz because thats what I ran lua at? or is it 7.2ghz because I clocked the mem tests there for 2 seconds? Do I have special needs? Yes, I dont think so, and most definitely just to answer those for you.
6. 2x16gb at 4400 14-14 seems to be equivalent in score to 2x8gb 4850 14-14 and has less tendency to fail DJ/LUA. Will compare on ln2 this week all out and update.
7. Dont tune your mem for mem score, tune your mem for total score. You will find some timings that boost your mem score but drop your total. Keep that in mind when you are playing with subs. This is very apparent on Ryzen as its ability to train crazy timings is so easy but unfortunately they can be robbing from your total.
8. Most mainstream systems can get away with 1.7-2gb maxmem. Any less and you start to bug sub tests and IMO is a sign that the sub should be scrutinized using geek link comparison with a similar system.
9. On AMD Ryzen use the batfile from asus to boost your score. They say to start the bat file then run the test then run the default bat. I wont start an argument on this just sharing some secrets some people might not know. See Zen Perfboost : https://bit.ly/2kBs15c (After loading CineBench R15 or Geekbench3 bench then run the boost, just click defaults button after benching to prevent bsod)
3.4.3B3 seems to be the most consistent on Ryzen 3000 atm. Same for TRX40. Haven’t tried the rest yet. ~keeph8n