Nice preview video in the tweet!
Putting the tweet text here so it shows up in forum search:
GSkill f5-6000j3644d64gx4 -- first ddr5 6000 kit plug and play on this motherboard from MSI! video soon.
I noticed from another thread there’s 3 variants of that kit:
Some users have also posted some observations, starting with: Where are the DDR5 Unbuffered 256GB 64GBx4 Kits? - #127 by dmayle
I registered here just to report the same finding. I bought 2 sets of the same kit you did from v-color, and while a single pair boots up right away with DDR5-5600, nothing I did – including setting the UEFI to the exact same settings as auto – would allow the system to POST. I’ve waited for up to 1h for it; I didn’t have a speaker plugged in then when I was testing it on a separate test rig and always needed a CMOS reset afterward to get it back to the UEFI.
what agesa version? a lot of mobos are dropping bioses the last few days/weeks with 1.2.0.3f and thats a good starting point… its absolutely bonkers how big of a difference it makes.
and the board to board variation is huge.
About 2 mins to POST with 256gb though! which is crazy awesome!
However older ddr5 kits may be more trouble than before/longer post time.
This was using the latest v3205 which claims to be 1.2.0.3f. Which is why your Twitter post was surprising to me.
Ack, I keep trying to find an edit action and I can’t – this is a Asus Proart x670E, Ryzen 9950X, two kits of v-color TE564G64D852K (64gb x2, 6400MT, ECC UDIMM)
i dont see any 64gb or 2x64gb support in the QVL of that motherboard.
mentest86 is not a heavy load. It is not enough to stress test memory. You want something like ycruncher with VT3 to stress test. Also, single CCD Zen 5 cannot saturate memory bus because of the limitation of the single inifinity fabric.
Right now, I am still on AGESA 1.2.0.3e. Admittedly, it is not the best bios for 2DPC in the history. I went through my archive. I found the best results I got was from AGESA 1.1.0.1. It happened to be the first bios that supports 256GB.
Here was my 128GB at 6800MT/s back in 2024. It was only benchmark stable.
https://www.overclock.net/posts/29304038/
1.2.0.3e was a huge regression for me, leaving me unable to boot with anything higher than 5000, and sending training times through the roof. Fortunately the patch notes stating that downgrading was impossible were false.
Curious what mobo/proc you are running.
After that post I decided to give the new agesa a try. Looks like I finally managed to reach 5600MHz stable with my 4x64GB setup!
I’ve also toyed a little bit with TRFC and TREFI values and achieved 60GB/s memory bandwith with mlc. Currently running some extra tests to guarantee that things are stable, but so far everything is looking good!
cool, can you post your full zen timings? and mobo? it could help others.
I’ve found huge variance across boards and some CPU silicon lottery factors in testing. video soon.
msi is far out in front followed by Asus for compatibility… then ASRock then gigabyte.
asrok just posted a bunch of bioses on Friday tho ulfh
Don’t have Windows, so monitor_cpu it is:
sudo monitor_cpu -m | tail -n +4
GDM: Enabled
CR: 1T
Tcl: 42
Tras: 82
Trcdrd: 42
Trcdwr: 42
Trc: 135
Trp: 42
Trrds: 8
Trrdl: 14
Trtp: 21
Tfaw: 32
Tcwl: 40
Twtrs: 7
Twtrl: 28
Twr: 84
Trdrddd: 12
Trdrdsd: 11
Trdrdsc: 1
Trdrdscl: 7
Twrwrdd: 11
Twrwrsd: 10
Twrwrsc: 1
Twrwrscl: 21
Twrrd: 8
Trdwr: 20
Tcke: 0
Trfc: 204
Trfc2: 224
Trfc4: 9
sudo dmidecode -t memory | grep -i 'configured memory' | head -n1
Configured Memory Speed: 5600 MT/s
sudo agesafetch
:: Searching Reserved region #1 (5123 KiB)...
-> Found AGESA version: ComboAm5PI 1.2.0.3f
9950x on a x670 ProArt with BIOS 3205, using 4x of Kingston’s KVR64A52BD8-64.
I’ve set the refresh interval do 32000 and TRFC1 to 1228 (copy pasted those values from another place here). Also left FCLK as 2000MHz, “Power Down Enable” disabled, “UCLK DIV1 MODE” as “UCLK=MEMCLK”.
mlc results:
Intel(R) Memory Latency Checker - v3.11b
*** Unable to modify prefetchers (try executing 'modprobe msr')
*** So, enabling random access for latency measurements
Measuring idle latencies for random access (in ns)...
Numa node
Numa node 0
0 85.7
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 62219.8
3:1 Reads-Writes : 59474.9
2:1 Reads-Writes : 60049.0
1:1 Reads-Writes : 61289.9
Stream-triad like: 59533.5
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 62269.8
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 196.22 62060.0
00002 196.45 62080.3
00008 185.82 61491.2
00015 174.93 61057.4
00050 199.50 61816.5
00100 197.10 61857.1
00200 194.89 61727.7
00300 197.73 61789.0
00400 201.65 61925.4
00500 195.77 61777.6
00700 117.37 51857.8
01000 103.91 37797.1
01300 99.10 29851.4
01700 96.36 23324.6
02500 90.57 16364.5
03500 89.27 12009.1
05000 88.34 8686.8
09000 87.37 5188.5
20000 86.65 2754.3
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 15.4
Local Socket L2->L2 HITM latency 15.4
I am on the AGESA 1.2.0.3f now. I don’t see any improvement. Maybe the improvement is not meant for x670e Carbon. Maybe my double 2x32GB Kingston kits have already got patched before that.
I actually don’t mind being stuck at 5600 at all. There is a optimized fclk for 5600. That is 2100 fclk. It runs pyprime under 9 seconds with Ryzen 9950x, with CL30 @5600 1.25v. Very efficient.
Plus 7950x undervolted to negative 27 on all cores. These settings all together are rock solid. I tried many overnight Y cruncher sessions, gaming, editing, EU5 editor with demanding sample scenes and also let the PC to idle on desktop for long periods of time. Works perfectly.
Thanks to help of people in this thread, as I have very little idea what I’m doing. I think the RAM OC could be pushed even more in my case, but I just lack the knowledge and patience. This is good enough for me.
Behold my push pull 420 AIO and RAM cooling solution (3x 6k fans, capped at reasonably quiet 3k). Both AIO and RAM coolers had to be attached by questionable DIY methods, but hey. It works. Besides 3x140 fans in the front, There are also 2x140 hidden fans in the bottom compartment, close to the front ones. There’s way too many fans inside this case. And I like it.
GPU is 7900 XTX, PSU is EVGA 750 GQ, UPS Green Cell 2000VA 1400W
This was build for heavy procedural modelling, a quick demonstration in this thread, for those interested.
I had my best results with the first or second BIOS that added 48GB DIMM support.
The BIOS that stated “Significantly enhanced memory compatibility, with a focus on configurations utilizing all four DIMM slots” was the worst for me.
now in the other thread 
currently undecided, whether to get a 4x64gb kit from g.skill. which is 170$ cheaper than my other option that are a couple of vcolor 2x64gb kits with sk hynix m die and 400mhz faster.
pros for g.skill: i get a full kit of paired tested ram, and its cheaper.
cons for g.skill: samsung m die. not as good a sk.hynix.
pros for vcolor: 6400mhz, Sk.hynix chips. best of all from what ive read.
cons for vcolor: would have to buy 2 separate 128gb kits. and more expensive.
Interesting. I’m looking at this setup now.
9950x with Asrock B650E Taichi Lite
I got the 2x64gb Gskill Trident 6000 CL34 kits and i was able to push it to 6000 CL32, but no further
