I have a weird issue booting with the RAM kit Trident Z Neo G.Skill F4-3600C16Q-64GTZNC, it’s a quad 64GB kit. At first after installing the kit everthing worked fine for a few hours. Then at some point after quitting a memtest86 and doing a new cold boot I could not get it to POST anymore and it gets stuck in 3F error and reboot loop. This problem only happens when I try to boot on 4 sticks, if I remove one stick it boots without problems. I used the recommended slots for the sticks A2,B2,C2,D2
If I put back my previous ram kit everything works fine. It’s a corsair vengenace LPX running @2113MHz.
The weird part:
At first the whole kit worked fine, I enabled the XMP profile and I started it at 3500 and testing some real workload like compilation jobs which where just a bit unstable. I started lowering the frequencies until around 3400 where it felt stable under realistic load with multiple CPU heavy tasks. I then started a memtest86, I enabled SMP (multicore) mode on the memtest and I’m not sure if I did well.
I left the test to run and came back later to see a lot of errors so I stopped the test, which I think made the machine do a new cold boot. After this I could not anymore boot on the 4 sticks ever, except one weird case. Everytime it tries to boot on 4 sticks it is stuck on boot loop.
I know that this kit is not on the official QVL list abut I also found this post that has a very similar issue and hardware setup but I don’t have a big GPU, I am just running a small passive cooled NVIDIA card since this is a workstation/server setup.
So I tried many things (see below) and it didn’t make sense to me that it worked first then stopped working. I didn’t do anything funny with voltages, just applied XMP profiles and went on lowering frequencies. Why did it work first ? I can even confirm that it was running XMP profile at around 3500 since I did RAM benchmarks before and after and I can see a huge difference in results. I was thinking about the reasons that could have made it work first and the only thing that made sense to me is that I was running a different RAM kit before (32GB 8x4 corsair LPX 2133) and maybe the board rememberd some settings which it tried to apply. So I tried adding one stick from my previous kit with the 4 Trident GSKILL sticks to see if it boots. And you guessed it, it boots without errors but the ram is detected in single channel mode @2133 1.2v and memtest fails on some regions of the ram which is normal I guess from such a mix.
What I tried so far:
Resetting the CMOS
Switching the C2 D2 slots
Switching the whole kit to A1 B1 C1 D1
I can get it to boot on 1,2 or 3 sticks but not 4 sticks
I can get it to boot on 5 sticks (4 trident Z sticks ) + 1 stick from corsair LPX 2133
So I cannot boot the 4 sticks TridentZ but if I add an extra stick from my previous ram it boots and detects all the modules in single channel mode and the RAM does not pass the memtest.
Now I am back running on my previous RAM kit for which I rerun a memtest to make sure I don’t have a hardware issue on the board or the memory controller. The memtest passes without problem on my previous kit.
I am now thinking to sell the Trident Z it but as I understood it is supposed to be well optimized for Ryzen/Threadrippers so I want to be sure there is nothing else I can do to make this new kit boot “again”. Is it worth my effort to make this kit work or should I just order a new one that is listed on the QVL of the RAM websites ? I am fine to run it at lower frequenices I know that quad channel on Threadripper and x399 will be limited to max around 3200 or less.
Here is what happens with my 2950x and the ASUS x399 Gaming board.
I have the same 64GB matching sticks of Corsair Vengence LPX rated for 3200, but I found with all four sticks I can max out the speed to around 2666, it gets to be somewhat erratic at even 2900
So I end up running it at the lower speed but it does boot fine and runs all my apps and VM’s without any issues.
I’ve not tried any other RAM kits however so it could be that the speed needs to be tweaked.
How I did it was to load defaults so it runs the memory at the JDEC speed of 2133 and then slowly adjusted the speed manually in the BIOS until I had stability issues.
I have an x399 MSI board and got 8x16GB running at 2933Mhz
Check if the CPU is mounted correctly, socket SP3 is like any large server socket sensitive.
thread ripper reduces dims per channel once your ram is over a certain speed.
with the max for the 2950x being 2933 for quad channel support.
(after that i think it switches to dual channel support which means you can only use 2 dims over 2933 and even then the ram will switch over to 2t timings which slashes ram performance but allows you to increase ram bandwidth MT/s.)
2933 is the sweet spot for TR 2950x
so you have a choice…
you can check the jdec (cpu-z) and see what timings it offers for 2933.
if it supports 2933 then you can enter bios/uefi and set 2933 as the ram speed save and quit.
or
you can go full bore and manually enter all the timings from a tool like ryzen dram checker for optimized ram timings.
(addendum)
if your looking to maxamise your bandwidth and are willing to go to 2t timings.
install 3 sticks of ram.
boot to uefi and manually set the ram speed to 2t 3600.
leave everything else auto
save and see if you boot.
if it does.
shut down and install the 4th stick.
as you already switched to 2t the system might boot.
if it does boot check at what cas timings (cpu-z). and see if the trade off is worth it.(likely not)
I have 8x Samsung-B die and my system runs at CR1 14-16-16-16-34 2933Mhz SoC Voltage at 1.1625V, DRAM at 1.36V.
No idea what chips are on his F4-3600C16Q-64GTZNC, may be that his board does not get along with the RAM, but with TR it is often that the CPU was not evenly tightened
Try lowering ram speed to like 2933, even though you “can” get 3200 “working” you’ll likely get a shit load of invisible errors clogging up your memory bandwidth
It wasn’t until 3000 and 5000 that we really escaped the 3.2ghz ram barrier
for 8 channels, even with the gen 3 and gen 4 threadrippers where they support 8 channel memory, you have to lower the speed to 2666 for out-of-the-box support. this is because >4 dual-rank channels is quite hard on the memory controller. that said, you can stabilize 3200 or possibly even 3600 (I’ve never used a Zen+ chip so I’m not sure where the infinity fabric tops out) by giving enough SOC and VDDP voltage (I’d just give it 1.2V SOC and 1.1-1.15V VDDG IOD if you can specify the IOD voltage separately from the CCD voltage; if you can’t, you’re gonna have a harder time getting stability as the CCDs hate voltage >1.05-1.1V).
loosen out the timings completely while you’re testing this. and do test this very thoroughly - I used Google’s stressapptest over 8 hour windows to determine stability. though like when the timings were wrong it’d usually spit out errors within 5-10 minutes. when you start tightening timings, SCLs don’t go below 4/4 on TR, tWTR_s/tWTR_l/tRTP/tWR can all be left auto cause they really don’t go lower stably (and getting these wrong would produce single bit flips after THREE HOURS of stress testing so even if you can get them one or two ticks lower, it’s not worth your time to test as the performance difference is negligible). tRRD_s/tRRD_l/tFAW should do 4/4/16, tRCDWR goes all the way down to 8, tRAS goes down to 21 as long as tRC is suitably high (i.e. > tRP + tRAS). and for the rest of the primaries I’d just try to stabilize the memory’s XMP profile. I also had to give up tCL at anything resembling a reasonable value - anything below 22 at the higher speeds is too hard on the infinity fabric and shows errors if you stress the L3 cache (like with prime95 192k fft size), though memory stress tests will pass without issue.
you’re also gonna need to mess with ProcODT and the termination resistances + find the right AddrCmdSetup time for your board. literally every speed above 2666 is going to throw errors on a RAM stability test until you do.