Enabling XMP Profile Causes GPU Driver to crash

[ressonantia]

Problem Statement:
System ran fine when first put together, even with XMP enabled. It started with random TDR crashes from the GPU drivers necessitating hard reboots, then progressed to the point where windows would not boot with XMP enabled. Keeps BSOD-ing with a thread_stuck_in_device_driver error.

Workaround:
Run RAM at 2133MHz (not ideal)

System specs:
CPU: Intel Core i5 6500 @ 3.2GHz (Skylake)
Mobo: Gigabyte GA-170N-WIFI (Z170/mITX)
RAM: 2 x 8GB DDR4 G.skill F4-3333C16D-16GVK (16-16-16-36-2N)
GPU: HIS AMD Radeon R9 290X 4GB
OS: Windows 10 Home (64 bit)
GPU Driver: Radeon Crimson 16.6.1

Background:
I recently upgraded my computer from an AMD A10-7850K to an Intel i5 6500, had to do a clean install of windows because of drivers. Since my motherboard only supports 3200MHz instead of the 3333MHz that the RAM modules are rated for, I turned on XMP and manually set the multiplier to x32 so that it runs at 3200MHz.

Things I have tried:

• Ran Memtest overnight with XMP (3200MHz) enabled - no errors detected
• Manually set timings for RAM so that it runs at 3200MHz - still bluescreens
• Unseated/Reseated both GPU and RAM modules - still bluescreens
• Reverted back to 16.5.3 version of Crimson drivers - still bluescreens

Any ideas? Is it hardware or software related?

 

[LTC8K6]

Try without the DGPU, use the IGP?

 

[ressonantia]

Is this to try and isolate if its the memory? Surely running Memtest for 6 hours with XMP enabled would be enough to exclude the RAM modules?

 

[UsandThem]

Was that ram on the motherboard's Qualified Vendor List, or did G Skill's ram configurator say that ram was fully certified with your motherboard?

I see that issue on memory reviews on places like Newegg, and while not all the time, enough times on here where it is at least somewhat common. If the XMP wasn't certified, you might have to manually configure it or it might not be 100% stable. You might have to update your BIOS, or possibly return it and select compatible ram.

I know on my Gigabyte GA-Z170X-UD5 board, neither Gigabyte nor G Skill say that brand of RAM is compatible in 8 GB modules or larger in DDR4 3200 and above. When the speed is DDR4 3200 and above, only certain 4 GB G-Skill modules are certified to work in XMP.

There are quite a few "Need new ram, which ones should I buy?" posts on this site. I always recommend people order ram QVL on the motherboard manufacturer's website or use the memory manufacturer's website. However, most of the time someone else will come in and say "Buy any ram you want, all memory at all speeds will work. The QVL is irrelevant" (like in this thread: http://forums.anandtech.com/showthread.php?t=2476669), which is 100% innacurate and misleading.

EDIT:

I just checked G-Skill's ram configurator now that I am not on mobile, and this is what I see:

You can manually set your timings and speed in your BIOS, update your BIOS to the newest version, and if that doesn't work, you will probably have to buy new RAM that is certified to work on your motherboard. Gigabyte has certified a limited number of modules on their website. It appears that board is very particular on the memory that will work on it:

http://www.gigabyte.com/products/product-page.aspx?pid=5518#dl

 

[LTC8K6]

Is this to try and isolate if its the memory? Surely running Memtest for 6 hours with XMP enabled would be enough to exclude the RAM modules?

Not sure why you think changing the graphics card is to test the ram?

It is to see if it's the video card/video card drivers.

Since you have an IGP, it's very easy for you to eliminate that one variable by removing the video card and using the IGP.

 

[UsandThem]

Since I am more awake now than when I first saw your post:

If your ram works fine at DDR4 2133 stock speeds and you are using a discreet graphic card, there is little to no real-world difference you will actually see. If you are a person who cares about synthetic benchmarks, then buy different ram if you are unable to manually set the timings and make it stable.

I know that an enthusiast will come in and say "That's not true. You have to use frame variance and run the CPU without the video card or run it at low resolutions so it's not GPU limited to see the real difference". And one might link to the Euro gaming review of the i3 that shows amazing differences by overclocking it and the ram.

But the other 99% of memory reviews, benchmarks, and DDR4 scaling articles shows otherwise. Take it up with the reviewers if you don't like their testing methods.

The big question is what are you getting for that money when it comes to performance gains? If you are an average joe, not much. If you're using a discrete graphics card, the gaming frame rate difference between DDR4-2133 and the XMP or even overclocked configurations is generally minimal to non-existent. Simply put, there are no games that need the 30GB/s+ of memory bandwidth that even our lowliest default memory speed configuration could muster. The same is true in most other applications as well, but in our most bandwidth-limited workloads like WinRAR extraction and LuxMark 3D rendering, there was actually a large performance improvement which only got bigger when we further tightened the timings or increased the memory frequency.

http://www.hardwarecanucks.com/foru...-ddr4-3600-16gb-low-latency-kit-review-7.html

The application-performance reality is often at odds with the sheer bandwidth on offer because the processor's on-chip cache mitigates most of the benefit. Fast DDR4 is perhaps useful in relatively niche scenarios where the enthusiast is pushing components to their limits and wants ultimate performance.

http://hexus.net/tech/reviews/ram/85847-gskill-trident-z-16gb-ddr4-3000-f4-3000c15d-16gtz/?page=7

 

[Ketchup]

Since I am more awake now than when I first saw your post:

If your ram works fine at DDR4 2133 stock speeds and you are using a discreet graphic card, there is little to no real-world difference you will actually see....

This. Put another way, you would see/feel no different going from single channel to dual channel when using a discreet video card with Skylake. The difference going to XMP would be even less than that (yeah, less than nothing).

The only thing that concerns me, as LTC8K6 pointed out, is the nature of the messages, which do hint at the video card, but I wouldn't worry too much about that yet. If a looser memory profile makes your system happy, then leave it at that.

 

[ressonantia]

Not sure why you think changing the graphics card is to test the ram?

It is to see if it's the video card/video card drivers.

Since you have an IGP, it's very easy for you to eliminate that one variable by removing the video card and using the IGP.

Right, I get it now. The thing is though, that the system runs fine when clocking the RAM at 2133MHz, actually it ran fine when I first put it together with the RAM running at 3200MHz. I've also tried the 16.5.3 drivers to see if it was driver related, but no dice.

Since I am more awake now than when I first saw your post:

If your ram works fine at DDR4 2133 stock speeds and you are using a discreet graphic card, there is little to no real-world difference you will actually see. If you are a person who cares about synthetic benchmarks, then buy different ram if you are unable to manually set the timings and make it stable.
-snip-

Yeah I realise that even in the real world, the performance difference might be 1-2 fps in games, if even that but just knowing that the RAM is not performing up to what I bought it as is just slightly unsettling, I might just have to bite the bullet and live with it.

This. Put another way, you would see/feel no different going from single channel to dual channel when using a discreet video card with Skylake. The difference going to XMP would be even less than that (yeah, less than nothing).

The only thing that concerns me, as LTC8K6 pointed out, is the nature of the messages, which do hint at the video card, but I wouldn't worry too much about that yet. If a looser memory profile makes your system happy, then leave it at that.

I've tried looser timings, and manually setting the timings as opposed to using XMP, but they still end up with the same error.

I think for me now, I can live with running the RAM at JEDEC specs but I am still very curious as to why this is happening.

 

[UsandThem]

Right, I get it now. The thing is though, that the system runs fine when clocking the RAM at 2133MHz, actually it ran fine when I first put it together with the RAM running at 3200MHz. I've also tried the 16.5.3 drivers to see if it was driver related, but no dice.

Yeah I realise that even in the real world, the performance difference might be 1-2 fps in games, if even that but just knowing that the RAM is not performing up to what I bought it as is just slightly unsettling, I might just have to bite the bullet and live with it.

I think for me now, I can live with running the RAM at JEDEC specs but I am still very curious as to why this is happening.

http://www.intel.com/content/www/us/en/gaming/extreme-memory-profile-xmp.html

When DDR4 and z170 took off, there were a lot of 1-star reviews on motherboards from users who were having XMP issues or stability, or sometimes the ram wouldn't work at all. A lot of it was because people were buying whatever ram Newegg had $10 off on at the time. When you don't buy ram that is been certified by the motherboard or memory manufacturer, you take risks. Sometimes it works. Sometimes it doesn't.

If you have to see the possible 1-2 FPS in games using faster memory, that is your call. You have a locked CPU, so you might not even see that difference. If you want to benchmark your FPS in games, you will see a difference if you use integrated graphics or turn down the resolution to 800 x 600 or similar to take your video card out of the benchmark. That sounds absolutely boring, but that's just me. You can always sell your memory at a loss, and order ram that is on the QVL if you must have that tiny increase.

I run DDR4 2133, but I don't care about benchmarks. I just care that my computer is fast and stable when using it. I even thought at one time about getting DDR4 3200 because it is so inexpensive now. But I would get almost a zero return on that purchase. It will be much better spent on a video card or NVMe drive.

 

[UsandThem]

Also, and I might have missed it, but have you updated your BIOS version to the latest one for the board (maybe even a beta version)?

 

[ressonantia]

Yup, my bios is the F7 one. Initially I couldn't even enable XMP until I updated to this latest bios. But now its borked again

 

[ressonantia]

Update:

So I read through the QVL for the GA-Z170N-WIFI motherboard from Gigabyte again and this line caught my eye "When running XMP at DDR4 3200 MHz or higher, the system’s stability depends on the CPU’s capabilities" and a quick web search revealed that it might be related to the memory controller on the CPU (though why it manifests as GPU drivers crashing I haven't the faintest clue).

So it seems like I need to raise my VCCSA to achieve stability or something with XMP. Seeing as my bios does not have any VCCSA but it does have a VCCIO (and it seems like VCCSA is offset from VCCIO in my case) raising it from 0.96V to 0.965V has resulted in me being able to boot into Windows with XMP at 3200MHz. It seems stable so far, I'll need to play a few games to see if thats what it was.

 

[fastamdman]

I would recommend doing a real test instead of just gaming. Run a program like intel burn test, occt, prime 95, something along those lines that will actually push the processor at 100% load on all cores. Last thing you want is to be working on something important and have a random crash without being able to save your data before it happens.

Also use something like cpuid hardware monitor to check all of your voltages and temps to make sure they are in line. During Intel Burn Test, look for temps under 90C while overclocked. Personally I stay at 80C max, but depending on the chips, 85-90C is completely fine.

 

[ressonantia]

The latest version of Prime 95 makes the CPU overheat and shut off so thats not good. I've read somewhere that I should be using version 26.6 instead of 28.9. I am not overclocking the CPU/Bclk (though I guess running XMP bumps these up a bit) so these are standard clocks for the CPU. I've tried using both memtest86 (ran overnight) and HCI memtest (tested to 400% coverage I think) and both did not find errors when I did manage to run my RAM at 3200MHz.

 

[UsandThem]

Just use the Intel burn test (Intel Extreme Tuning Utility) if you are concerned about Prime.

It's safe and will stress your CPU and memory.

You can also use Intel Processor Diagnostic Tool.

If you can run both of those programs for an hour and not crash/have program stop because of error, you should be stable.

 

[ressonantia]

I'll give them a try tonight or something. Thanks for the suggestions. I've got my RAM clocked at the default 2133MHz now.

 

[ressonantia]

So I re-ran HCI memtest (200% coverage) and the Intel Processor Diagnostic Tool with my RAM clocked at 3200MHz and it all seems fine so far. I've run both the CPU and Memory stress tests from the Intel Extreme Tuning Utility (for 5 minutes each) until I can find some time to do a proper, longer test but even that passed.

I've bumped up VCCIO to 1.150V, I have no way of changing the VCCSA which stays at 1.068V