Hi there guys, hope is all doing good.
After finding a good discount on 2x48GB kits (there was just 1 box left), and while having 2x32GB already, I said why not.
The kits are:
G.Skill F5-6400J3239F48G (A-Die) – 2024 Week 8
G.Skill F5-6400J3239G32G (M-Die) – 2023 Week 18
I had to test a lot, at 6000Mhz it crashed at any latency (CL50 for example), and also some timings/FCLK settings are kinda finicky to made it work mixing dies.
The final settings that for now I have found to be stable in 3 days (about just 1 entire day making sure there was no errors)
I followed a bit from this video https://www.youtube.com/watch?v=20Ka9nt1tYU, mostly for the resistance/impedance settings.
Max temps were found while doing TM5 Extreme/Absolute, on particular 1 48GB DIMM which maxed at 48°C. The rest is mostly between 35-40°C (I have a 120mm fan pointing at them). Pretty hot nonetheless for this particular one.
Latency is here on safe mode without internet.
And with safe mode and internet
You can notice the write speed is a bit slow, right?
Why I don't recommend this (4 slots and mixing ram sizes)
Because 2 main issues:
- As you can notice, latency is not that good and write speed is also a bit slow. I haven't tested much lower latencies but probably not much to do. Using 4 sticks of 2R will limit your speed and latencies.
- You will get a lot lower latencies by using 2×32, 2×48, etc. I think 2x16GB is the best scenario.
- Mixing 2x32GB and 2x48GB. In this case (correct me if I'm wrong), RAM runs on "asymmetric dual-channel mode", which means the matched part of the size will run at dual channel, and the rest, at single channel.
- This means that up to 128GB RAM, it runs at dual channel, and the rest at single channel. This hinders performance a lot in some applications.
I do have a use for this RAM (Machine Learning, specially some tasks with LLMs) where first I load the model into RAM (even if it's quantized!) which can use 120-140GB RAM before moving to GPU (2×4090+1×3090 for my case). For example, a 123B model (Mistral Large 2 123B) at 4BPW, uses about 140GB first, and then it loads into the GPUs (using about 68GB VRAM)
Also, if playing, you won't never (correct me if I'm wrong again) surpass 128GB RAM usage which can make sure you run at dual channel. Also, well even if using more than 128GB, system doesn't necessarily allocate memory in a strictly linear fashion from "bottom to top".
—–
So that's all! Now wondering, have you managed to run 4×32/4×48 on AM5/Z6XX-Z7XX? If yes, how do yours speed/latency go?
4 Comments
Nice
What’s your stick/slot arrangement? As long as you are straddling each kit across both channels, you won’t be operating in asymmetric/flex mode.
Ex:
| **32GB** | 48GB | **32GB** | 48GB | <- would result in each channel working with 80GB, no asymmetry.
| **32GB** | **32GB** | 48GB | 48GB | <- would result in a 64GB / 96GB split, with the bottom 128GB in dual channel and upper 32GB in single channel.
that is really cursed but are there any newer boards that let you set different timings per channel?
my giga x58a ud3r lets me set timings per channel so i can do really cursed high frequency mix match ic manufacturer and ranks and i have managed 2656c11 stable at 2.2v with a mix of psc in the inner channel with gdie in the middle and dual rank cfr on the outermost channel limited by the psc stick cause its some green pcb transcend 1333 but now that i have some elpida bbse and 2gbit d die i might be able to do 2800+ in mix match so ill probably get to work on that once i get home and maybe do some testing aswell see if it even has any real performance
boards that allow seperate timings per channel are extremely overpowered for overclocking mix match setups but i have yet to find a ddr4 or ddr5 board that allows it, would be hilarious to run mix match ddr4 at 5000+ stable just that i dont have a ddr4 or ddr5 platform atm so that sucks, maybe ill go cezzane for my entry into ddr4 or whatever other chip that has a really strong imc though id probably wait for zen5 apus for ddr5
that latency is still really good compared to normal expo/xmp timings most people use lol