I’m new to RAM overclocking and I have 2 (3) different sets of RAM and I’m wondering which would give me the best gaming performance? I’m GPU “limited” most of the times so I understand that the (if there’s any) performance improvement would be in something like the “1%” lows and not average FPS.
I’m not sure if it’s possible to determine that just off of these screenshots though, so my apologies if it’s a stupid question.
The 6200MHz 1.4V 64GB is 4x16GB from 2 sets with EXPO profiles of 6400MHz CL30 but I can’t get them stable at 6400MHz.
The 8000MHz 1.4V 48GB is a set of 24×2 with EXPO profiles of 8000MHz CL40.
The latency according the AIDA64 test, is lower on the 8000MHz but the read/ writhes speeds are also be lower. That’s what made me unsure what would be the better choice.
I just installed the 8000MHz today so I’m still validating if they are stable. The 6200MHz passed 12h of AIDA64 and Prime95 with PBO and undervolt on the CPU.
Bonus question, how did I do with the timings and voltages and is there anything else I should be aware of/ do differently?
7 Comments
You should try if you can run fclk 2200 – its better then the suggested ratio if its like 3 steps above
On a Single CCD, you aren’t seeing the throughput benefit running 8000 MHz.
You can tighten your tRAS and tRC significantly. I’m not familiar with the 64 GB dual rank Hynix A die, but for 48 GB single rank M die, you can set tRAS to 48 and tRC to 96 at 8000MHz. The same 48GB set would also be able to easily run at 6200MHz at stock voltages with 30-38-36-30 tRC: 66 or even attempt CL28.
As a blasphemous aside, on my Dual CCD Ryzen and 96GB dual rank M die, I’m currently seeing CL38 7400MHz outperform CL30 6400Mhz 1:1. But still in the process of fine tuning it.
Timings look fine. Probably still some optimization to be had, but nothing from a quick glance looked significantly bad.
Single CCD CPUs (like a 9800x3D) are limited in terms of read/write outright speed by the infinity fabric, and therefore your FCLK speed. Your 1:1 6200 MT/s setup had a slightly higher FCLK, therefore higher read, etc. You can technically run higher FCLKs on either, but for the 2:1 8000 MT/s setup, if gaming is your goal, then 2000 FCLK is preferred for the latency advantage. If you were doing bandwidth heavy stuff, then perhaps higher (ideally 2200) could make sense.
Your Vsoc on the 2:1 8000 MT/s setup could probably go lower. I’m at 1.05v Vsoc on my setup. This is one nice upside to a 2:1 setup, since UCLK is so low for 8000 MT/s (only 2000), you don’t need a lot of Vsoc, and lower Vsoc helps with FCLK stability, lower idle power consumption and temperatures, etc.
Subtimings need quite a bit of work in either configuration. But my advice would be to max out FCLK first, then work on RAM; since both mem and FCLK are dependent on SOC voltage but scale in the opposite directions in terms of stability. You could probably argue that you’ll see more of a real world performance benefit from FCLK on X3D as well (compared to RAM), but I don’t have the data to back that up.
Why is ram that slow on AMD? Genuine question, I just switched from Intel and I feel like a noob in the AMD world and trying to learn the overclocking side of AMD.
This is my stock XMP 6800CL32 on my 13900KS:
[https://i.imgur.com/Mnf9ykM.png](https://i.imgur.com/Mnf9ykM.png)
Stock timings of the 8000mhz kits don’t look so nice 24gb sticks are m-die I think so there should be much more possible
Single ccd chips don’t benefit as much as dual ccd chips with 8000. Buildzoid recommended going. 6200 and 2200 if possible since the x3D chiplet compensates for latency and it’s already bandwidth starved. Yoh could also try going for 8400/2100 but I doubt it’ll happen. Worth a shot tho. Assuming the expo is stable you can try raising vdd and vddq to 1.5v, and raise VPP to 1.82 or so and giving that a shot. I doubt it’ll work but that’s part of the game.
The 48gb kit will be easier on the memory controller so 6400 may be possible on that kit with 1.255-1.3v soc. If you can’t get that try 6200 at 1.2-1.25v and push fclk high as it can go.