Switching RAM mid-overclock?

[User5566]

Hi, I'm in a (hopefully) final stages of setting up my Ryzen 9 5950x overclock (with curve optimizer and pbo2). I had 12h of stable prime95 run and in general I felt pretty comfortable with my settings.

However, I've recently changed my memory from 3200MHz cl14 to double the size 4000mhz cl18 and my pc restarted when doing nothing... Unfortunately no errors were logged so I don't know which core was it.

And this leads me to writing this post. It is possible running faster memory (but with slightly worse latency) caused my previously stable OC to become unstable, or was it unstable before, but I was just lucky that it didn't crash before?

Is there a way to tell?

The details of the setup are:
Ryzen 9 5950x + corsair h150i aio
Before 32gb G. Skill 3200mhz cl14
Now 64gb g.skill 4000mhz cl18
MB Asus tuf x570 plus
Gpu - dual gigabyte rtx3090 24g with nvlink
PSU - 1000W be quiet pure power 12(or 13 I don't remember).

 

[EarthDog]

Is there a way to tell?

Sure. Disable your CPU overclock and see if it's still unstable with the memory at DDR4-4000. Is that memory kit on the QVL list for your board? Is your BIOS updated to the latest version?

That said, I'd guess it's the memory. 64GB of DDR4-4000 is a lot of stress on the memory controller. I'm not even sure that platform can run that much RAM at that speed... 3600.... sure. If I were you, I would add some SOC voltage (increase up to 1.3V max) and see if that helps with stability.

Your goal here is to isolate the problem. So, CPU back to stock, enable EXPO/XMP profiles on your RAM and test. If it's still not stable, then clearly it's your RAM and needs adjustments for stability (raising the SOC voltage).

 

[Woomack]

I recommend 3600-3800 1:1. Easier to stabilize and faster in ... pretty much everything. You can probably set tighter timings, too. I guess the 4k kit is 18-22-22, so you can set 16-18-18 or 16-19-19 at 3600-3800 1:1.
If it's a 2x32GB kit, then it's probably Hynix. If it's Hynix, then it can OC above 4000 ... but on high series motherboards with optimized BIOS. I have no idea if 4k will work without issues on the X570 TUF mobo, as it's considered not bad but a lower gaming series (like under Crosshair and Strix).

 

[User5566]

Thanks for the replies so far

No the ram is not on the QVL, but no 32gb modules are at OC speeds. However I always considered QVL more of a suggestion. The module uses samsung b dies and if possible I'd like to run two kits for the total of 128gb and a post above is telling me 64gb is too much

It is a g.skill f4-4000c18d-64gtzr. Yes indeed it is cl18-22-22-43

I was happy with 32gb of g.skill 3200mhz cl14-18-18-18-22, but I needed more ram, so I thought this particular kit will be fairly close in performance (slightly worse latency, but faster throughput).

I think it is good advice to try it at stock, no oc (other than ram) cpu settings. However being an optimist I am, I tried core cycler (as it crashed when not in use). And so far it crashed twice after 10min on core no 3. That is my "best" core. In curve optimizer it is the only core that required +1 to be stable and it still boosts to 5050mhz (4.9 effective). So I'll try upping it to maybe +5 and then I'll do what was suggested.

The motherboard QVL claims there are memories that should run at 4600mhz, but much smaller kits (16gb etc).

Edit: The ram is currently set at DOCP 1.4V. When it crashed in Windows last, it did log a WHEA Machine check exception for core 3, a "cache hierarchy error".

 

[Woomack]

I don't think it's Samsung B because Samsung B is lower density IC, and 16GB modules/double-sided are max. I also don't think it's Micron B (but possible), as G.Skill wasn't using Micron B in these kits. It's more likely Hynix, but in this capacity, Hynix is the best for OC/performance.

The best results are on Ryzen APU 4000/5000. On the best setup, I couldn't stabilize 128GB above 4200. The same with Hynix and Micron. 2x32GB could go up to 4600... but also only on Ryzen 5000 APU as the regular one gave up at 4133 or something near.
Ryzen 5000 (non-APU) has a worse memory controller. Also, your motherboard is not the best for RAM OC. I'm not saying it's not possible, but it may take you some time and manual tweaking to stabilize 32GB modules at a higher clock. I would start from 3600-3733, and once you stabilize it, then try one higher ratio and test again.
As EarthDog mentioned, you can check higher SOC voltage manually, as somewhere around 4000, the motherboard will probably limit SoC voltage. Also, check higher VDIMM.

Because of IMC/IF ratios, it's still recommended to keep it at 1:1 and tighten memory timings for the lowest latency. To match 3600 1:1, you need at least 4800 1:2 ... and it's impossible with 32GB modules. With 16GB modules, it was possible only on Micron B, but it's a higher density IC with limited timings tweaking options, and to match 3600 1:1, it requires about 5000 1:2. It's good only for 4000/5000 APUs.

 

[EarthDog]

However I always considered QVL more of a suggestion.

While the QVL list isn't The Gospel, it's more than just suggestions. It tells you what was tested to work on that board. If it's not on the list, it means it didn't work (or wasn't tested).

The motherboard QVL claims there are memories that should run at 4600mhz, but much smaller kits (16gb etc).

Right. The increased capacity puts more stress on the IMC, which lowers the ceiling. The higher the capacity, the lower the speed it will support. If you're looking for 128GB, you'll likely require even slower speeds with the increased capacity.

 

[User5566]

I don't think it's Samsung B because Samsung B is lower density IC, and 16GB modules/double-sided are max. I also don't think it's Micron B (but possible), as G.Skill wasn't using Micron B in these kits. It's more likely Hynix, but in this capacity, Hynix is the best for OC/performance.

I re-checked. You are correct... It isn't a b die. I used the b-die finder site, I saw the part number I thought was this one (f4-4000c18q2-64gtzr), but my kit is not "q2", but "d". The other kit was 8*8gb so I couldn't use it anyway. I misread that type before.

I am running the latest bios. And 1.4V vdimm (as per xmp/docp).

The best results are on Ryzen APU 4000/5000. On the best setup, I couldn't stabilize 128GB above 4200. The same with Hynix and Micron. 2x32GB could go up to 4600... but also only on Ryzen 5000 APU as the regular one gave up at 4133 or something near.
Ryzen 5000 (non-APU) has a worse memory controller. Also, your motherboard is not the best for RAM OC. I'm not saying it's not possible, but it may take you some time and manual tweaking to stabilize 32GB modules at a higher clock. I would start from 3600-3733, and once you stabilize it, then try one higher ratio and test again.
As EarthDog mentioned, you can check higher SOC voltage manually, as somewhere around 4000, the motherboard will probably limit SoC voltage. Also, check higher VDIMM.

What kind of stability test would you suggest to use for memory testing on my Ryzen 9 5950x? Prime95 blend? On all cores? I find blend causes windows to just swap constantly... (I'd rather not stress test my nvme at this time) so I use custom setting for the amount of ram that leaves 8gb free and 2 free cores.

I've started it like this an hour and a half or so ago (no pbo2, no curve optimizer) at "auto" fclk that means 1800mhz on my bios and ram at 4000mhz.

Is this a good way to test ram stability at these particular settings or is there better software? Eventually I'd like to try up the fclk to 2000mhz (yes, I know how unlikely this is to work, but who knows, maybe I won the silicon lottery?)

Edit: Would it be better to use something like core cycler to test individual cores letting them boost a lot higher rather than a multicore workload that barely touches 3ghz(with no oc).

Then, upping the SOC voltage was mentioned. After watching some videos about self immolating ryzen cpus (although newer gen). Do you think it is a requirement to measure the actual so voltage delivered by the MB before messing with it? I have a good voltmeter. I just need to find out where are the test points to do it.

Because of IMC/IF ratios, it's still recommended to keep it at 1:1 and tighten memory timings for the lowest latency. To match 3600 1:1, you need at least 4800 1:2 ... and it's impossible with 32GB modules. With 16GB modules, it was possible only on Micron B, but it's a higher density IC with limited timings tweaking options, and to match 3600 1:1, it requires about 5000 1:2. It's good only for 4000/5000 APUs.

I guess the most likely scenario I would be happy with is to run it at 3600mhz cl16. But I'll still try squeezing more out of it first.

While the QVL list isn't The Gospel, it's more than just suggestions. It tells you what was tested to work on that board. If it's not on the list, it means it didn't work (or wasn't tested).

I'm hoping it's the latter. If not I can return these kits for free, but then what else can I get at this capacity? Probably nothing...

Right. The increased capacity puts more stress on the IMC, which lowers the ceiling. The higher the capacity, the lower the speed it will support. If you're looking for 128GB, you'll likely require even slower speeds with the increased capacity.

If that is the case I'll have to make some decisions after seeing some benchmarks. I don't know yet what real world performance hit will I have running it at slower speeds.

I wonder, what is the mechanism of bigger modules being more difficult to get to work at higher clocks for the MB. (assuming the manufacturer is right the module is good for let's say 4000mhz, what difference does it make for the MB if it addresses more RAM on the same number of modules or less?)

Electronics is one of my hobbies. I'm very interested in hearing it, if anyone has a technical explanation why this happens.

 

[EarthDog]

but then what else can I get at this capacity? Probably nothing...

2x32GB DDR4-3600?

G.SKILL TridentZ RGB Series 64GB DDR4 3600 RAM Memory - Newegg.com

Buy G.SKILL TridentZ RGB Series 64GB (2 x 32GB) 288-Pin PC RAM DDR4 3600 (PC4 28800) Desktop Memory Model F4-3600C18D-64GTZR with fast shipping and top-rated customer service. Once you know, you Newegg!

www.newegg.com

Electronics is one of my hobbies. I'm very interested in hearing it, if anyone has a technical explanation why this happens.

Oh @Woomack !!!!!!!!!!!!!!!

 

[Woomack]

Hynix is scaling with voltages almost only on CL. At under 4000, you can leave SOC at auto, as it will probably be enough of what the motherboard sets. However, 3600-3733 will require 1.40-1.45V VDIMM at CL16. Hynix (I assume it's Hynix, but Micron will be acting almost the same) is scaling with VDIMM almost only on CL, so it's like:
3600 CL16-19-19 ~1.40-1.45V
3600 CL14-19-19 ~1.50-1.55V
3733, with some luck, will require similar voltages. Since it's hard for IMC to run stable with high-density memory, you probably won't be able to set it at 3800+ 1:1. You can try 4000-4133 1:2, but you may need to set a higher voltage or something like CL18-24-24. In theory, Hynix won't need more than 18-26-26 up to 4400-4600, so you can set these timings with other timings at auto, set 1.4V+ SOC, and check how high it will go. Too high SOC issues are related almost only to Ryzen 7000 X3D CPUs. It's not an issue even on regular Ryzen 7000 CPUs.

I won't say it in a technical way, but in an easy way, it's like there is a limit of ranks that can be supported/addressed. More ranks can be translated into more "devices". It generally looks like 2x dual-rank modules are as hard for the IMC as 4x single-rank modules. Dual-rank (or quad) has "wider access" to stored data in the same cycle, so it's also more demanding on the IMC side. It also looks like 2/4 memory slots always have better signals and more optimized traces. Here, you also have to consider all the differences between RAM ranks, but also motherboard topology, IMC capabilities or even how the motherboard manages all the signals, timings, and everything else. There are too many variables to give a quite short and straight answer.
Optimized traces and shorter traces between RAM slots and CPU improve RAM OC. You can also browse the web for DDR topology as there are differences, and there is some more to read. When you compare and gather up all the info, you will see why the most expensive OC motherboards are so good and why two memory slot motherboards overclock RAM the best.

 

[User5566]

2x32GB DDR4-3600?

G.SKILL TridentZ RGB Series 64GB DDR4 3600 RAM Memory - Newegg.com

Buy G.SKILL TridentZ RGB Series 64GB (2 x 32GB) 288-Pin PC RAM DDR4 3600 (PC4 28800) Desktop Memory Model F4-3600C18D-64GTZR with fast shipping and top-rated customer service. Once you know, you Newegg!

www.newegg.com

But that one is also not on the QVL (despite the board supporting 128gb ram there is no 32gb stick at any speed on the QVL). This particular kit is almost the same price as the 4000mhz kit here in my part of the world (Central EU).

I guess if my 4000mgz kit doesn't even run at 3600 I can try this one instead.

Hynix is scaling with voltages almost only on CL. At under 4000, you can leave SOC at auto, as it will probably be enough of what the motherboard sets. However, 3600-3733 will require 1.40-1.45V VDIMM at CL16. Hynix (I assume it's Hynix, but Micron will be acting almost the same) is scaling with VDIMM almost only on CL, so it's like:
3600 CL16-19-19 ~1.40-1.45V
3600 CL14-19-19 ~1.50-1.55V
3733, with some luck, will require similar voltages. Since it's hard for IMC to run stable with high-density memory, you probably won't be able to set it at 3800+ 1:1. You can try 4000-4133 1:2, but you may need to set a higher voltage or something like CL18-24-24. In theory, Hynix won't need more than 18-26-26 up to 4400-4600, so you can set these timings with other timings at auto, set 1.4V+ SOC, and check how high it will go. Too high SOC issues are related almost only to Ryzen 7000 X3D CPUs. It's not an issue even on regular Ryzen 7000 CPUs.

Thank you, this is very helpful.

I won't say it in a technical way, but in an easy way, it's like there is a limit of ranks that can be supported/addressed. More ranks can be translated into more "devices". It generally looks like 2x dual-rank modules are as hard for the IMC as 4x single-rank modules. Dual-rank (or quad) has "wider access" to stored data in the same cycle, so it's also more demanding on the IMC side. It also looks like 2/4 memory slots always have better signals and more optimized traces. Here, you also have to consider all the differences between RAM ranks, but also motherboard topology, IMC capabilities or even how the motherboard manages all the signals, timings, and everything else. There are too many variables to give a quite short and straight answer.
Optimized traces and shorter traces between RAM slots and CPU improve RAM OC. You can also browse the web for DDR topology as there are differences, and there is some more to read. When you compare and gather up all the info, you will see why the most expensive OC motherboards are so good and why two memory slot motherboards overclock RAM the best.

Ok.

So far I've been running Prime95 for 6h. I'm going to try single core core cycler with ycruncher next (no oc on the CPU, ram at 4000mhz cl18).

 

[EarthDog]

So far I've been running Prime95 for 6h.

What test in P95? Blend tests RAM, the others (large is useless, lol) not so much.

 

[User5566]

What test in P95? Blend tests RAM, the others (large is useless, lol) not so much.

Custom with 56gb ram limit(everything else on default) and on 14 out of 16 cores.

I believe the only difference between this and blend, is that this doesn't fill entire ram to cause swapping all the time.

However, after 6h of prime95 I just had a crash after 10 minutes of a single core workload with core cycler... So no sadly, it's not stable.

But I'm not 100% convinced it is the ram's fault. When I was tweaking my oc before I had to set core 3 to +1 to be stable. Now I'm not running any oc, but I'm going to repeat this test with +2 set on core 3 just out of curiosity.

Edit: it's not the new ram problem. I've put the old ram back in, and it crashed the same way during the same test on the same core with memory at lowest settings.

Now I'm trying to, clear CMOS, disable all the "auto oc" features of the board, and try again. If it does it again I'll try under Linux. If it happens there too I might need to overvoltage that core a bit or RMA the cpu. Shame. As its a nice unit. I never had this problem on my previous cpu a ryzen 7 3700x.

 

[User5566]

Well, it looks like I can't replicate the crash on Linux because ycruncher software for some reason has different tests available on Windows and Linux. On Windows it crashes within seconds after starting n64 test. On Linux there is no n64, there is n63. It's a different test so of course it doesn't crash.

 

[freeagent]

G.Skill has their own QVL. The QVL is not a suggestion as noted by ED. Also, mems that are QVL'd on say Zen 2 might not be on Zen 3..

Very few Zen 3 parts can do 2000 1:1, they are unicorns.. I have a 5600X that can scale to 2100, but is only stable up to 2000.

All of my Zen 3 parts can drive my 3200C14 sticks to 4400+, but actual performance is not that great, even though my timings are much tighter than yours.

TM5, Anta777 Extreme, or Absolut are the profiles you want to run. They will hammer the memory pretty hard, might want to cool them.

Linpack Xtreme is pretty good for checking core and subsystem stability, but it will beat your CPU with no fecks given about mercy.

 

[User5566]

G.Skill has their own QVL.

I didn't know. I only ever checked Asus's one. Now looking at G.SKILL QVL they seem to have tested up to 3800mhz on Ryzen 5xxx on this MB.

Also, mems that are QVL'd on say Zen 2 might not be on Zen 3..

Yes, Asus has separate qvls for ryzen 3xxx and 5xxxx.

Very few Zen 3 parts can do 2000 1:1, they are unicorns.. I have a 5600X that can scale to 2100, but is only stable up to 2000.

All of my Zen 3 parts can drive my 3200C14 sticks to 4400+, but actual performance is not that great, even though my timings are much tighter than yours.

TM5, Anta777 Extreme, or Absolut are the profiles you want to run. They will hammer the memory pretty hard, might want to cool them.

Linpack Xtreme is pretty good for checking core and subsystem stability, but it will beat your CPU with no fecks given about mercy.

Thanks.

I use Linux 99% of the time, but I use windows for stability testing because of tools like furmark for GPUs, cpu-z etc.


This time I seem to have found a weird bug. My "old" ram is on the MB QVL at 3200 mhz.

It seems older versions of y-cruncher had n64 check even on Linux so I downloaded it and I've been running it for last half an hour. Considering on Windows same check crashes in seconds it definitely looks weird.

Also I noticed core 3 boosts about 75MHz higher in windows. On Linux (while the test is running, with no oc) the highest it gets is 4975mhz. On Windows it boosts to 5050 a lot. Perhaps that's why it behaves differently.