I’ve got 4x32 running at 4800 MHz to work just fine without issues (boots just fine, no crashes or anything- haven’t run memtest). I haven’t attempted to get it to run higher, but probably will at some point. I’d rather have stability than eek out 5% more performance or something by trying to get it higher.
Specs:
CPU: Ryzen 7950X
Motherboard: Gigabyte X670E Aorus Master (BIOS version F6)
RAM: 4x32GB G.SKILL Ripjaws S5 6000 MHz ( F5-6000J3238G32GX2-RS5K)- two 2x32GB kits purchased a couple of weeks apart on Newegg.
CPU cooler: Noctua NH-D15
GPU: RTX 3060 Ti FE
Power supply: Corsair HX1000i
OS: Linux Mint
I kept everything at auto except turned frequency up to 4800 MHz and didn’t touch the voltage. I don’t really care to play around with timings as I don’t see it making much of a difference. I’m currently putting it through various stress tests to make sure it’s stable but I have not run into any issues. Maybe if after a week or so of stress testing no issues have shown up I will try and see how high it can go.
NOTE: With this RAM/motherboard combo I cannot get it to boot with XMP even with just two sticks. I can however, keep the timings on auto and bump the frequency to 6000 MHz without a problem with two sticks. I guess this motherboard just doesn’t like the XMP timings or something.
I have been working on this about a month. I can maybe do DDR5-4000 to 4400 but it isn’t perfectly stable in all scenarios and workloads. 4000 is “mostly” stable at 1.4v with an excellent gstkill kit.
Can everyone here doing stuff do (even the trial version is fine) AIDA64 memory and cache benchmark?
It seems as though 2x32gb @ 6000 is, you know, 71ish gigabytes/sec bandwidth… but 4x32 @ 4400 is only ~51 gigabytes/sec.
I had some really crazy stuff dialed on on the asrock taichi and I was getting 91 gigabytes/sec at 71ns latency with a fabric clock of 2133 and memory @ 4166 but it wasn’t stable. I didn’t care that it wasn’t stable because I have never seen 91 gigabytes/sec in aida64 on the am5 platform.
So everyone rocking 3x32 or 4x32 can you post your aida64 cache/memory benchmark screens? I’m really curious.
Good news is that whether it is a cheap pcie4 board or the $1000 asus board the memory support works about the same across all boards.
No board failed to post (@ 3800) with 128gb. But there are latency and, seemingly, bandwidth penalties.
64gb might be my recommendation for this platform, for now. I think I might be a bit worried/disappointed esp. as it seems 13th gen intel has really improved things in the ddr5 department in both speed and capacity support.
Thanks for posting this. Upgrading my work machine to a 7950X, and didn’t want to spring for 128GB if the performance hit was going to be big. Glad to see that’s an option if (when lol) I need to upgrade soon.
How are you finding the speed otherwise? I’ve made a point of ordering a TPM for my new Asrock board to avoid any potential stuttering problems caused by the fTPM, which I’m definitely seeing on my 5900X at the moment.
It’s great so far. I work on broadcast news so I’m doing a lot of transcoding video and multitasking and I haven’t noticed any stuttering at all except the windows 11 login screen is quite laggy for 10-15 seconds and then another 10-15 seconds when I first log in but after that it’s great. I should really play some games to see if there are any issues which just haven’t shown up in my normal work
I’ve had a couple of other issues unrelated to the 7950x where Topaz Video AI is very inconsistent in using the hardware to its fullest performance (4090 at 5-60% usage after a few days of it working at 100%) but I’m still posting what I think are the fastest times on their forum and there are so many moving parts with Topaz versions, Nvidia drivers, Windows updates…
I’ve been needing to double check exports from After Effects so they don’t have green flashes in them if they have an external video file on the timeline, and I purge the cache if it happens (I had this issue on my 12900k laptop so it’s probably an Adobe thing)
Overall very happy and without synthetic benchmarks I wouldn’t think there could possibly be any performance left on the table but apparently there is
Also have the 7950x with 128GB RAM (6000J3238G32GX2*, G-Skill DDR5 CL30 6000 32GB x 4). Right now I can only run it at 3800 CL30 unfortunately; higher and I’ll get random compilation errors and freezes.
Gigabyte X670 Aorus Elite AX; managed to get 4200Mhz stable. 4400/4600 would sometimes give me errors, 4800+ would just have full errors/fail boot or even retrain the memory each time.
with 2x 64GB Corsair Vengence 5200Mhz at 1.25v (no voltage increased), VRAM set in bios at 1.25v it wouldn’t even boot if that wasn’t set. I also noticed that IGPU would fail to work with drivers if the CSM was enabled.
On a side note: I did a proxmox hypervisor on this one and I can confirm the passthrough gpu works very nicely, but the iommu groups on all other slots suck, either a bios issue or the platform itself.
Thanks for telling me. I hadn’t noticed that release. I just installed it now.
The memory settings look slightly different now and I believe there’s a new item, ADDR_CMD_MODE with the options Buf or UnBuf. I hope that this controls 1/2T mode or geardown? Haven’t been able to immediately find the answer online yet. Sometimes where ADDR_CMD_MODE is mentioned the options are 1T or 2T, so does that mean Buf and Unbuf will relate to those?
Not sure what else has been added I just fixated on that one option.
It trained to 3600CL30-29-29 this time. I added ram voltage 1.25v and -10 curve quickly and I’m in windows now with this 3600CL30. It’s interesting it likes to train to CL30 but anyway I’m going to try out the higher rates at CL40 again on this new bios.
At the moment it still says 1T on my aida result so I’m going to play around a bit and see if I can get to 2T mode or if I’m already on 2T and it’s just misreporting.
edit: Ok maybe aida was misreporting 1T because ZenTimings reports 2T mode
edit again The Plot Thickens, (1T rate reported now in ZenTimings):
1.25v 4400 is 2+ hours memtest64 stable now. Wasn’t stable the last time I tried 4400CL40-40-40 on 1.25. It’s also reporting 1T command rate so maybe we could be on for some higher rates if that’s not just a random number and we can get it to 2T. Will do more testing tomorrow
I’ve got 4400CL40 1.25v stable now* so that’s a step forward compared to the previous BIOS version.
(*stable so far)
I’ll try tighten CAS at the very end but I don’t think it’s important compared to trying for higher speeds.
The reason I want 2T mode (if it is actually working not just being misreported in software), it’ll allow something higher like 4800 maybe to be stable compared to 1T
I will try 4600/4800 etc tonight but I’m happy to have 4400cl40 1.25 and 2200 fabric now on this new BIOS
mega transfers give you bandwidth, cas latency gives you speed.
look at the table Memory timing examples.
notice first word return. thats the relative speed of reading and writing the first word (2 bytes of data) from the ram array.
the idea being the lower that cas latency number the faster more performative the ram.
the higher MT/s values means it can be performative across more bandwidth.
if you notice in the wiki table, ddr4 3600 has a first word return of 7.78ns
then look at ddr5 and the closest you can get is 6400 giving a 10ns first word return.
its effectively 25% slower in cycling data. but can push x2 the amount of data
in gaming, you want lowest latency, as bandwidth is less of a priority.
with ryzen 7000 you have a choice.
you can run with 2x48-64000
or
you can run 4x up to 4800 with 4x4800 giving about 1/3rd more bandwidth than the 2 6400 sticks.
and after seeing all the issues running ddr5. and i was building a ryzen 7000 build.
i would be building it with 4 sticks of ddr4 3600.
as that way i’d get optimum bandwidth with quad channel support and crazy low latency…
i wonder if @wendell would be willing to test this out (assuming part availability and time) and see if there’s any real world rather than benchmark difference.
Thanks to everyone who has helped so far. I think to answer my original question for the fine people using google who come across this thread; as someone who wants 128GB unless it sucks, using it now I would say it doesn’t suck. No stability issues, wasn’t hard to get first boot with 4 dims, all 128GB working as you’d expect. Downside is that there is performance loss vs 2 dimms but at least for my video/3d workload I haven’t felt a downside yet.
With that said: I’m pressing on trying to get all the performance I can, currently memtesting 4600cl40-40-40 1.25v. Stable so far (only 10 minutes into memtest 64.) 0805 bios had an effect on memory stability, this wasn’t always posting and wasn’t stable before without 1.4v
Edit:
Passed over an hour of memtest64, will leave it running 4600 1.25v tomorrow and see if its still stable
