Download MSI Afterburner, click on OC auto scan, wait till it's finished. Depending on the GPU it can last 15-45 minutes. Weirdly enough the 4090 took way longer than the 4070 but I don't have sample size to judge anything so it's whatever. When it's done save the profile to 1 and 5 so you have some redundancy at least. click on 1010 miliVolts, hold shift button down and select all the points after that in a big swoop, and drag them down. ( I know it's weird, look up a YouTube video if you must to figure it out ) after you drag it down in minus -100 or lower number you click on exit window button. And click on apply in the main window of MSI Afterburner. Go back to frequency curve editor and you'll notice it's flat lined in a single line. After the 1000 millivolts line or 1 volt. GOOD you undervolted the GPU. Why is this relevant? well for some reason Nvidia overclocked all their cards at 1.1volt but 1 volt is the sweet spot. MSI Afterburner does OC scan in the whole frequency range. But we just cut it off at 1 volt. I found out that when you do this instead of just maxing the frequency manually at 1 volt, it performs higher in 3DMARK. So obviously it's faster to just use auto OC (and easier also) only thing you have to do after this is manually OC memory. Any gpu does 1000 frequency. My 4070 does 2000 but my 4090 does 1500. So it depends on the memory modules silicon lottery. the 4090 has 24gb and the 4070 has 12gb, so the chance of getting a bad memory silicon lottery is 2 times as high on a 4090... so you get higher oc on average on lower vram gpu's. Have FUN.

I forgot to include this tool for testing vram stability : https://github.com/GpuZelenograd/memtest_vulkan

Be aware it runs your memory at high temp I got like 85 celcius or something after roughly 45-60 minutes on gaming x trio 4090 but most gpu's get way higher temps due to poorer cooling. I know 95-105 temp is max safe temp for gddr6x according to the internet. depending on sources 95 might be real number. 105 might be because of warranty issues but it's bad. 30 series had memory temp issues because they used lower capacity memory modules to populate the area and half of them were on the back side of the card which would get a lot hotter. 40 series doesn't have this problem. So the temps are much cooler.

I played F.E.A.R. briefly by pure coincidence with my 1500 MHz memory OC which is stable according to this software program after roughly 45 minutes to 60 minutes of stress testing. I don't recall exactly. I had this issue that my framerate would plummet after a bunch of minutes in game. Due to lower GPU usage, which one might ask how? but older games on this GPU trigger energy savings, running into less than 10% gpu usage. I fixed this issue somewhat by enabling max performance in nvidia contrl panel. After some trial and error. ( by the way this game will crash to desktop if you don't change resolution in public documents / timegate studio's cfg file on a high res monitor ) Long story short I got memory artifacts in F.E.A.R. even though it was stable before in testing at 1 volt. But I noticed it ran at 0.9 volt in this game due to the low load. I decreased the clock to 1400 MHz and now it's fine. ^^

Concluding my advice is going 100MHz lower on the memory clock after stable results for good measure.

(MSI Gaming X Trio, both cards)

Every GPU is different, the lower boost number in GPU-Z is normal when doing a undervolt curve.

I use 3dmark timespy extreme for verifying performance results.

​

https://i.redd.it/n9fpbktozpfc1.gif