Running 4X32Gb on AM5 you might be better off with a very tight 4400MTs CL24 profile, running MCLK 2200, FCLK 2200, UCLK 2200.
As mentioned before, 2800; 2100; 2800 should work as well on 8-layer PCBs, and would be much faster than a 4400MT/s setup.
Right now, I'm on the road, but once I'm back next week, I'll try to test 128GB with a fresh 7950X3D and a new B650E Aorus Master board.
Sorry, what is this RAM training?
Finding the ideal timing values with finetuning?
No, it doesn't fine tune or auto-detect any of your timings. It's a way more low level process that is required to ensure proper communications between your memory controller and the memory chips on an electrical level.
I'm not good enough at electrical engineering to explain it properly or in depth, but to put it simply, the memory controller inside your CPU writes and reads data-patterns to the memory chips repeatedly and adjusts the electrical signals accordingly to ensure signal integrity. Therefore, if you change the memory setup by adding or removing sticks, or change the frequency of the memory, you have to retrain. If you run very aggressive overclocks, you can even adjust training parameters to ensure better stability. However, training will take ages then.
It's something the newest generations of AMD & Intel CPU's do
It's also done for DDR4, but most folks won't notice it, because the training requirements aren't as stringent as DDR5. For example, after clear CMOS, training 128GB with one of my 5950X's takes about 10 seconds, 128GB with one of my 7950X's takes between 150 and 300 seconds depending on the motherboard and AGESA version.
Most DDR4 boards also default to memory context restore automatically, which makes consecutive reboots almost instantly. If your motherboard has a post-code display, it's easy to follow.
Mabye it's harder to get 4 ram sticks to run at 6000 with a 7950X3D
X3D has lower voltage caps.
There is a rumor, or more a suspicion, that at least early X3D chips have a lower binned I/O die. I don't know if it's true or not, but I've seen multiple claims on forums that X3D's seem to be weaker in the memory department. I think, even that Buildzoid guy mentioned the suspicion in one of his rambling videos. Unlike Mr. $0.02, he seems to know a thing or two about memory overclocking.
My 1st 7950X3D has a manufacturing date around May '23 and has a worse memory controller than my 7950X from Aug. '23 and my 7700 from Nov. '22 (not a typo, it was manufactured really that early). I got another 7950X3D last Friday, that was manufactured Jan. '24. I'll probably have enough time to test the CPU a bit with 4x 32GB and maybe even 4x 48GB before building the new system. Personally, I'm really curious about the memory OC possibilities of that one.
Generally speaking, the most important voltage for memory overclocks of the AM5 memory controllers is
VSOC. Both regular and X3D CPUs are capped at 1.3V to my knowledge. Most motherboard BIOSes with AGESA 1.1 and later seem to default to 1.2V or 1.25V. Gigabyte is also quite sneaky and set the cap twice in their BIOS GUIs to make it harder for people to accidentally increase it.
VDDIO, VDDQ and VDD aren't capped, as far as I know, and you can happily fry your memory sticks, if you chose to do so. I think you have to stay below 1.8V for VDD or the system won't POST, but I'm sure that some BIOSes allow values above that.
VDDIO and VDDQ have a very pronounced sweet spot effect for quad-rank setups, and vastly differ between board manufacturers. On my Gigabyte boards it seems to be 1.25V for both, but for my MSI one I gave up and just switched back to 96GB, because 128GB at 4800MT/s was the only speed that was stable and easy to tune. My only Asus is a 2 DIMM board.
Adjusting VDDP and VDDGs is also possible, but I wouldn't touch them for daily setups, mainly because it's extremely annoying to test them for stability unless you connect an oscilloscope to your board.