Unstable XMP After Upgrade from 16 to 32 Gb

[thetruck12]

I recently upgraded from 16 Gb (2x8) to 32 Gb (4x8) on my self-built PC. The sticks are G.Skill Ripjaw V Series DDR4-3000. The PC is about 7 months old and I upgraded about a month ago. I was stably running the 16 Gb on the XMP profile (3000 MHz 15-16-16-35 @ 1.35V). I tested stability after install with Prime95 using the 'blend' test overnight. When I upgraded, I initially kept the XMP profile on (without realizing it) and had memory instability issues playing Black Ops 4. Initially, I thought it was the game (which I got at the same time), but when I finally did a stress test with Prime95, it failed. To make sure the new sticks and slots were good, I swapped the new and old sticks around in various combinations (all with the XMP profile still on). Ultimately, I only found instability with all 4 sticks, which is when I finally thought to switch off XMP (base settings: 2128 MHz 15-15-15-36 @ 1.2V). After much testing with Prime95, I confirmed that all 4 sticks with the base speed are stable. I've been using my PC at those speeds for about 2 weeks now with no issues. However, I want to figure out why the XMP profile is unstable, since the sticks are rated to run at that speed. Unfortunately, when I now switch on the XMP profile to do additional testing, Windows does not properly load (get a blue screen and reset). Any ideas as to what may be the problem? Is there any further testing I can do without windows loading up?

Specs:
OS - Windows 10
MB - ASRock X470 Master SLI/AC ATX AM4
CPU - Ryzen 7 1800X (No overclocking)
RAM - G.Skill Ripjaw V Series DDR4-3000 (4x8Gb)
GPU - Zotac GeForce GTX 1050 Ti 4 Gb
PSU - SeaSonic EVO 620W 80+ Bronze

 

[jfriend00]

You can test just the RAM without booting into Windows with PassMark's MemTest86. On a Windows system that is stable, you will need to create a bootable USB stick using their install procedure. You can then interrupt the BIOS loading and tell it to boot from your USB stick. That will load up MemTest86 and let you test your RAM without booting Windows.

If your upgrade was adding a second set of 2x8, then that may be your issue. An XMP profile for a 4x8 set can easily be different than the same sticks configured as a 2x8, particularly when working with Ryzen (which is a lot pickier about memory configuration). So, you have an XMP profile for 2x8, not for 4x8. I don't know much about Ryzen memory configuration (I've always done my memory overclocking with Intel processors/motherboards), but the general things you can work on are manually setting a slower speed than 3000 or slowing the timings down from 15-16-16-35 to perhaps 16-17-17-37 or increasing the DRAM voltage. Generally, you wouldn't need to increase the DRAM voltage beyond 1.35V just for 3000MHz though. It's more likely a timing issue and it may be a timing issue related to having 4 sticks so it could even be a secondary or tertiary timing issue. There are literally dozens of memory timings, some of which are in the XMP memory profile (perhaps 7 or 8) and the rest of which your motherboard BIOS has to figure out based on your other configuration.

Here's what I would first try. Take your stable 2x8 and set it for 2133MHz (default profile, no XMP). Very that is stable with PassMark's MemTest86 (free version), HCI MemTest (free) or RAMTest (small purchase, but worth it). Then, still set for 2133MHz, add the second set of RAM. Test that to see if it's stable. If it's not, then reset your BIOS to all defaults (not by changing settings, but by actually resetting it to default internally). This will clear any prior training the motherboard may have done with the 2x8 set. Let it boot into the 4x8 set at the default speed of 2133MHz and test that for stability. If that's not stable, then you're either going to need to sell these sticks and buy a 4x8 set or find an expert in memory configuration for Ryzen (which isn't me).

If that is stable at 2133MHz, then you can try the XMP profile. If that's stable, you're good and it was probably a stuck memory training for the 2x8 and you now got it properly trained for the 4x8 set. If that's not stable, then you can try relaxing the timings as I mentioned above or will need that expert in Ryzen memory configuration to help you.

In the Intel world, there's also a CPU-related voltage called VCCIO which controls the voltage to the integrated memory controller in the CPU (IMC). Sometimes you have to increase that voltage too for 4x8 though a good motherboard would do that automatically for you. I don't know what it's called for a Ryzen motherboard.

 

[Zerileous]

Try the DRAM calculator.

You enter in your specs and the desired clock, and it gives you timings, including all the subs, as well as vsoc and vdram. It is somewhat tedious to enter them all into BIOS. You'll also need taiphoon burner to get all the specs on your memory. If they are the same chips on all four sticks you'll have better luck, as I don't think you can get settings for multiple chip setups from the calculator. That said you can see whichever one produces more relaxed settings and try that, if you don't have matching sticks.

 

[thetruck12]

I have gone through the suggested testing with MemTest86 using all available threads (16) in parallel; I ran all tests over 4 iterations. At the base speed (2128 MHz), the testing passed with the original sticks, but failed when I added in the new sticks. To isolate the suspected issue with the new sticks, I then tested the new sticks alone but they passed (ran it twice to be sure). I then tested both pairs of sticks in the alternate slots (A1/B1) to check if it is an issue with those slots. Oddly, the original sticks passed, but the new sticks failed (ran it twice to be sure). Finally, I tested all 4 in the opposite config as before (new sticks in A2/B2 and orig sticks in A1/B1) and it again failed. At this point, I am very confused. As a novice, the results don't seem to point to a specific issue.

For all the tests, I used the default BIOS settings (reset before each run). I considered a run a failure as soon as I saw errors (usually in test 3 or 4 during the 1st iteration) and then stopped the run. The errors generally came from only one of the threads during a given run. Could that indicate an issue with simultaneous multithreading (SMT)? All my previous testing with Prime95 had that turned off.

 

[jfriend00]

I have gone through the suggested testing with MemTest86 using all available threads (16) in parallel; I ran all tests over 4 iterations. At the base speed (2128 MHz), the testing passed with the original sticks, but failed when I added in the new sticks. To isolate the suspected issue with the new sticks, I then tested the new sticks alone but they passed (ran it twice to be sure). I then tested both pairs of sticks in the alternate slots (A1/B1) to check if it is an issue with those slots. Oddly, the original sticks passed, but the new sticks failed (ran it twice to be sure). Finally, I tested all 4 in the opposite config as before (new sticks in A2/B2 and orig sticks in A1/B1) and it again failed. At this point, I am very confused. As a novice, the results don't seem to point to a specific issue.

For all the tests, I used the default BIOS settings (reset before each run). I considered a run a failure as soon as I saw errors (usually in test 3 or 4 during the 1st iteration) and then stopped the run. The errors generally came from only one of the threads during a given run. Could that indicate an issue with simultaneous multithreading (SMT)? All my previous testing with Prime95 had that turned off.

Probably what you turned off in Prime95 was AVX instructions. There's no real point in testing with Prime95 if you turn off multi-threading as then you're only testing a single core which isn't likely what you really want to assess stability with.

There are unsubstantiated reports that the multi-threaded tests in MemTest86 relies on threading support in the BIOS and that it is sometimes buggy and can cause false positives - showing errors when the memory is just fine. Because of that, I only run it single threaded. Takes longer to run, but removes any doubt about false positives. I don't know if those rumors are true or not or perhaps it varies by motherboard BIOS.

If you can't get the 4 sticks to run at the base 2133MHz, then there's perhaps a compatibility issue with the four sticks where the second pair either has a problem on its own or simply isn't compatible with your first pair (the two sets of chips each want different memory timings).

I would first verify that you still see the same problem when running MemTest86 single threaded. I choose one of the options that cycles through the different cores so you're testing all of them. If you still see a problem even at 2133MHz, then there's probably no point in proceeding further. If you want 32GB, you will probably need to return the second set you bought, acquire a new matched 32GB set that is sold and tested and advertised as a 32GB set and then sell your first set.

There are voltages to play with (DRAM voltage and VCCIO), but if you can't even get easy stability at 2133MHz, there's really no point in messing with those. The eventual goal is presumably to get 32GB working at 3000 or higher. You want easy stability at 2133MHz before you start going higher. I'd return/sell what you have and buy a 32GB set that lists your motherboard on its QVL.

 

[thetruck12]

Probably what you turned off in Prime95 was AVX instructions. There's no real point in testing with Prime95 if you turn off multi-threading as then you're only testing a single core which isn't likely what you really want to assess stability with.

I had turned SMT off from the BIOS settings. I was still running Prime95 on all 8 cores (8 threads).

If you can't get the 4 sticks to run at the base 2133MHz, then there's perhaps a compatibility issue with the four sticks where the second pair either has a problem on its own or simply isn't compatible with your first pair (the two sets of chips each want different memory timings).

To clarify, the sticks are all the same model (G.Skill Ripjaw V Series DDR4-3000) from the same vendor, but were made 4 months apart (July 2018 vs Nov 2018). Could there still be incompatibilities?

I would first verify that you still see the same problem when running MemTest86 single threaded. I choose one of the options that cycles through the different cores so you're testing all of them. If you still see a problem even at 2133MHz, then there's probably no point in proceeding further. If you want 32GB, you will probably need to return the second set you bought, acquire a new matched 32GB set that is sold and tested and advertised as a 32GB set and then sell your first set.

There are voltages to play with (DRAM voltage and VCCIO), but if you can't even get easy stability at 2133MHz, there's really no point in messing with those. The eventual goal is presumably to get 32GB working at 3000 or higher. You want easy stability at 2133MHz before you start going higher. I'd return/sell what you have and buy a 32GB set that lists your motherboard on its QVL.

I just tried running all 4 sticks at the base speed across all CPUs in serial with and without SMT turned on. Both showed errors during test 3 and 4. Any other suggestions?

Thanks a lot for your help.

 

[jfriend00]

To clarify, the sticks are all the same model (G.Skill Ripjaw V Series DDR4-3000) from the same vendor, but were made 4 months apart (July 2018 vs Nov 2018). Could there still be incompatibilities?

Yes, the only way to ensure you can run 4 sticks together is to buy them as a matched, tested set. Sometimes it works to mix sets purchased at different times and sometimes not. There's no requirement that G.Skill makes every DIMM of G.Skill Ripjaw V Series DDR4-3000 with the exact same memory chips. In fact, I'm quite sure from time to time a supplier will introduce new chips that offers some benefit to G.Skill (lower price, better features, better availability, etc...) and they will switch to that new chip. Even within the same chip, due to performance differences, you can't always say that any 4 sticks of the same model will all work together.

I just tried running all 4 sticks at the base speed across all CPUs in serial with and without SMT turned on. Both showed errors during test 3 and 4. Any other suggestions?
Thanks a lot for your help.

I can't tell exactly what test you ran here. If you ran the sticks in pairs at 2133MHz on default BIOS settings with no customization and still saw errors, then they're bad and need to be returned or exchanged under warranty. If they work at 2133MHz as pairs, but don't work together as 4 sticks, then they are probably incompatible as a set. As I said previously, there's just no reason to mess with them further. Get a set that works easily together at 2133MHz and then you can experiment with the XMP profile speed and higher. But, don't spend any time trying to tweak sticks that won't even work at 2133MHz. It's just not worth it. If your sticks don't do what the XMP profile says when run as the pair that you purchased them, then that is grounds to return them to where you purchased (if still within return time period) or to make a warranty claim with G.Skill if beyond the return time. That's what I would do.

 

[EarthDog]

Sorry if this was covered... but more ram tends to need additional system agent(imc) and io voltage...but that is intel

If the sticks are erroring out at jedec, even with additional SA and IO voltage, it may be the sticks.

 

[Zerileous]

I would still suggest you download Thaiphoon Burner so we can see exactly what chips you have and what their timings are.

There are also some timings that the 1000 series ryzens just do not like, which is where DRAM calculator can help. It can also help you with voltages (for AMD it would be called VSOC), however automatic settings "should" handle four sticks at the jdec speeds. However this is likely where your instability at XMP is coming from. It is also possible that the profiles call for different timings, and one is pushing timings to the board that the other stick doesn't like.

Else fails return them, however G.Skill is not necessarily going to accept an RMA if your motherboard wasn't listed on their DRAM configurator for those sticks.

 

[thetruck12]

I downloaded the free version, but I can't seem to get it to run (Win 10). However, based on the model number (F4-3000C15D-16GVKB), both sets of sticks are the same. After looking it up, I found a screenshot for the same model on a overclock.net forum

There are also some timings that the 1000 series ryzens just do not like, which is where DRAM calculator can help. It can also help you with voltages (for AMD it would be called VSOC), however automatic settings "should" handle four sticks at the jdec speeds. However this is likely where your instability at XMP is coming from. It is also possible that the profiles call for different timings, and one is pushing timings to the board that the other stick doesn't like.

Else fails return them, however G.Skill is not necessarily going to accept an RMA if your motherboard wasn't listed on their DRAM configurator for those sticks.

Looking at the DRAM calculator, it doesn't seem that I can calculate the timings for JDEC speeds. Is it OK if I just calculate the timings for 3000 MHz and go straight to that?

Also, I don't see this RAM model on the configurator or the QVL for my board (though I could've sworn it was on there when I bought the first set). Does that mean it's useless to even try to get the 4 sticks to work or just that the auto settings aren't likely to work?

 

[Johan45]

First getting 4 sticks of Hynix to run on AM4 is going to be a real PITA. You might have to start at 2666 which is the official supported speed for 4x8 sticks and work up from there. If you can't get them to stop erroring then I would just return the sticks, if you don't need 32GB then don't bother it's overkill for most users.

 

[trents]

What Johan45 said. You are likely expecting to much from Ryzen with that amount of RAM, that many sticks and at that frequency, not to mention it's not a matched kit and it's using Hynix chips. With Intel, it would likely not be a problem.

 

[Zerileous]

Yeah sorry, I didn't realize that both kits were the same model number (not just the same model and brand), not sure if it was mentioned and I missed it. Most of my theories pertained to having two kits with different chips. Still not being a matched set means it was never tested together, which could cause problems. Also being ripjaws it was likely never tested with AMD systems.

I would start with 2666 on the calculator and see how that does, as Johan45 advised. Also, it may caution you to gradually increase your SOC voltage, which is sound advice. It is a lot of settings, take your time and have a computer or phone handy in case your timings in the calculator don't match what you find in the bios. Good luck with the overclock!