Is Vulkan (Windows) actually stable for anyone?

I have a RX 580, and everything I've used Vulkan with usually causes a display driver crash, consistently, anywhere from 2-20 minutes maybe. On Dota 2 specifically, I also have a lot of random flickering.

• Age of Empires II: Definitive Edition (DXVK)
  
• FINAL FANTASY XIV (DXVK)
  
• Dota 2 (native with -vulkan launch flag)
  
• Guild Wars 2 (DXVK)

This seemingly only happens with Vulkan. If I keep the default renderer on those games (DX11 on all except GW2), or even use DX12 with GW2 (d912pxy), I have no issues. I had assumed DXVK was somehow the problem, but Dota 2 crashes as well.

Does Vulkan work well for anyone, specifically with games on Windows? One stability test I can think of would be to get through an average match (30-40-mins?) of Dota 2.

• Happens at stock GPU clocks
• Memory passed 24h memtest
• Happened on the last 4 driver releases, including 20.11.2
• Drivers always clean-installed (20.11.2 just installed on a fresh W10 install earlier today)
• DXVK on Linux (RADV) worked fine for me in the past

 

memtest is not a conclusive indicator of good or bad memory

 

If you use a AMD motherboard and the newer bios updates then far as I've heard the latest AGESA also with these A to C patches all have their little issues and can be sensitive even more so than usual so the first I would test is without the XMP profile to rule that possibility out entirely.
Secondary is using the latest Vulkan loader/runtime over AMD's included 1.2.135 one for additional fixes. - https://vulkan.lunarg.com/sdk/home#sdk/downloadConfirm/latest/windows/vulkan-runtime.exe

5700XT user here I pretty much use DXVK in everything and if available Vulkan or D3D12 natively since it's generally just better but some games can be very prone to instability but it does make for a great tester of stability.
(Watch_Dogs Legion doesn't like GPU memory overclocking at all, barely any headroom before it hits a driver crash for example.)

Generally I also use a five minute run of OCCT to confirm GPU Wattman stability for undervolting but the test is D3D11 based although any one single error and that's instability for the GPU test (I prefer max shader complexity but you can probably use less.) and on a poor setup that usually quickly leads to hundreds of errors detected so it's fairly good at giving a initial indicator here although I only use it for GPU stability testing.


EDIT: As to RAM and GPU instability and yeah driver crashes too best I've understood this as and well my poor explanation skills and how that's going to go.

Direct Memory Access or DMA and the PCI Express bridge and the chipset and how it ties into how the driver and GPU hardware is using these heavily so AMD cards can expose memory instability or borderline stability compared to how NVIDIA cards can seemingly be stable but then it's not a software issue it's a hardware problem at the end of it after all but testing and confirming this can be a bit of a process.

Makes it simple to just test with a lower RAM speed and then dialing in and confirming the XMP settings or just setting them all manually maybe also adding a little bit more voltage if supported and then you have Gear Down mode in addition to other systems like the load-line calibration and how some manufacturers use different settings and defaults and the RAM training and overall compatibility.

It sounds complicated and it kinda is, set it and forget it XMP profile #1 or whatever and just being done with it is nice but it's not guaranteed to be 100% stable and some kits and motherboards can be really finicky as a result or how to best explain this thing.

 

never had a problem with vulkan neither on my pascal nor turing cards.

 

Looks like that might be true. My RAM has XMP of 3266, and I was forcing some timings at that speed that I thought was stable (memtest ran fine and other games were fine). Took me a while to get to that point, but it looks like it still might not have been perfect.

I dropped the RAM speed and timings to auto (2133) and a quick match of Dota 2 with Vulkan had no random flickering (had a driver crash still, but I had a core clock OC). I'll give AoE2DE and a real Dota 2 match a go at some point soon at stock GPU clocks and see how it handles.

Would messing with the GPU power limit or memory timing affect stability? I noticed with poclmembench that using Memory timings 2 (the drop-down in AMD control panel) gave me higher speeds (190 to 205), and the power limit gave me more performance overall. I don't change anything outside of those usually (memory and core clocks and voltages I keep at auto).

 

Yeah my RAM settings were what was causing the problem. Interesting that it only seemed to affect Vulkan though, but overall glad to see it resolved. Feel a bit silly since OCs are usually the first thing to look at as a troubleshooting step

Was able to play multiple full matches of Dota 2 and AoE2DE (about 40-mins a match) no problem with Vulkan after using auto RAM settings.

 

Damn, this tells me I should avoid AMD no matter what.

edit
Just read that Cyberpunk 2077 wont have ray tracing with AMD GPUs at launch.
AMD need to step up their game!

 

Yeah there's a few issues, NVIDIA has some too but GPU wise the driver situation made me very doubtful about things and how long it took AMD to acknowledge and start fixing these issues although ones things got started it's gone pretty well though I had been hoping most of the majority of the issues would be resolved before the 6000 series came out and complicated matters.

CPU wise infinity fabric and the RAM and motherboard and XMP profiles can be hit & miss, good quality RAM can do pretty well but the constant tweaking in bios updates and both AMD AGESA and motherboard manufacturer changes you'd usually want to re-confirm and test after major updates here particularly when doing further tweaking although the gains can be really useful minimizing the resulting latency and now the 5000 series might allow even greater gains once the newer bios code and firmware to call it that (AGESA is a whole bunch of components.) is ready for deployment.

Ray tracing well it seems a few games kinda need NVIDIA device ID's but then AMD or the game developers of Godfall implemented AMD only supported tracing but just changing the config file enables it and results in...only player shadows being traced it's kinda underwhelming but the way AMD and NVIDIA approach ray tracing and then how it's looking like as it's implemented could also vary.

Seeing that NVIDIA has D3D extensions for NVAPI and ray tracing could be a blocker too for AMD (Watch_Dogs Legion which it seems works on AMD but with glitches.) and for Vulkan the earlier one or two titles using VK_NV_RTX_ type extensions instead of the standardized VK_KR_RTX_ ones will probably mean AMD does not support these. (Wolfenstein Youngblood and Quake 2 RTX I believe.)


Upgrade wise I've been looking at the 6800's from AMD and the 3080 from NVIDIA and now with the markup pricing and VAT both are almost the same price although there's no 3080's in stock although the 3070 situation has somewhat improved and the 3090's are slowly stocking up going to be interesting to see how things go and when it's hah well actually possible to upgrade because the situation with shipments and such limiting hardware availability.

AMD pulled some odd early launch OEM 6800 and 6800XT cards but outside of press and insiders who it seems got early access to purchase these some countries were barely supplied at all so the actual launch is looking to be the 25th.
6900's coming out on December 8th too so the 20.12.1 drivers to support that must land at that point too.

And then who knows what's going to break or change or get added with 20.12.2 when AMD launches that if I predict correctly and they want 20.12.1 as the 6900 support before major changes in the next driver.
AMD's take on GPU AI supported upscaling, maybe finalizing the GPU hardware scheduler implementation and SAM for well it does it's thing for this standard thing but eh it's free performance even if it's a small gain.
(With the hardware requirement caveat ha ha.)

 

Thanks, useful concise insight on a not particularly easy topic.

 

Yeah it's pretty tough I'm having a hard time just trying to track what's still problematic and what's been fixed either listed as a resolved issue or silently and then what's still problematic either listed or having been taken off the known issues listings but might still be a problem.

5000 series GPU's had some changes and a lot of issues with a variety of things software and potentially hardware side to some extent making for some complications.
Ampere's new but NVIDIA is doing a good job from what I am hearing although the situation with GPU availability makes it impossible to upgrade at the moment though this 5700XT now that the majority of the issues are resolved have worked pretty well for what it is.

6000 series is just launched and in more limited availability going to be a while before we see or hear more from what it might have problems with and how the driver situation will be this time around, I do expect shortages to be a thing as well and for how long is impossible to guess at.

I find it interesting and a very fascinating topic but I have a very hard time keeping up with just the known issues and what's been resolved and then the tech details as to how and why is well above my tech level for when AMD or NVIDIA even provides more details here or how some of this corresponds to general hardware stability like RAM and display driver issues.

I do learn quite a bit but I seem to be excellent at forgetting it too ha ha.
Generally it's not supposed to be that problematic buy the hardware plug it in and update the software/drivers but some of these developments have certainly been a experience.