Why is DX12 so much slower than DX11?

Th

Thanks for the info mate. I will check it out get back to you.

 

That's good news but we can't upgrade to 2022, when is it going to be updated and out of experimental for 2021?

 

I'm rendering a large(!) amount of trees and terrain details on the GPU using the GPUInstancer asset. Like, a LOT.
I'm putting together a stripped down version of my project which I can share with you for testing purposes. I'll send you a DM.

 

Since Dynamic Resolution on Windows is only compatible with DX12, we decided to give our HDRP game a go on DX12. Unfortunately, what usually runs at 60 FPS on a RTX 3060 is now running below 15. It simply isn't playable. It's quite surprising considering the game runs at 30 FPS on Xbox One and PS4.

Is there any plan to fix DX12 on 2020 LTS or will it remain in this unusuable state forever? It's a bit sad that dynamic resolution cannot work on modern PCs while it can on Xbox One.

 

Hi, thanks a lot for the feedback.

So we can assume that DX12 will be on par with DX11 at some point in performance in all cases ?

 

I have read through most of this thread, and the take-away seems to be that DX12 is roughly on par with DX11 in terms of CPU overhead.

So I created a release build to test a scene that is quite punishing on my RTX 2060 at 2560x1440. The scene is rendering dense grass, trees and other scenery. A lot of what you see is in the form of combined meshes in "tiles" (grass and rocks, etc), but trees are individually rendered. So overall we have quite a high batch count, shadow caster count, etc., and I'm getting about 60 to 70 FPS in DX11.

I then created a DX12 build to see if we can gain some performance - but actually it completely tanked the frame rate down to 20FPS, with "Render Latency" showing a very high number.

Is there any practical advice or ideas about what might cause this unexpected outcome? Could it be that as it stands, DX12 in Unity hates high batch count? Could it be related to our use of custom passes? I'm just guessing here.

Release Build DX11 - 62FPS - 23ms Render Latency:


Release Build DX12 - 20 FPS, 80ms Render Latency:


Scene profile (in editor):


Project info:
Unity 2022.2.12
HDRP
Running on RTX 2060 @2560x1440

 

My game performance with unity 2022 (.16) build:

gtx 1660 URP
DX11
4k 25 fps AVG build (15-35)
1440p 50 fps AVG build (35-70)
1080p 87 fps AVG build (45-130)
DX12
4k 17 fps AVG build (15-35)
1440p 37 fps AVG build (20-45)
1080p 55 fps AVG build (35-60)

any news in this? DX12 is very slow

 

I created a build where DX11, and DX12 is in, default is DX11. First command is 105 FPs, second one is 30 fps 1080p Any idea?

 

This is very similar to my case - DX12 frame rate is approximately 30% of DX11 frame rate. Something seems very very wrong.

 

@mgeorgedeveloper @lacas8282

I'm pretty sure that if you post about DX12 being slower than DX11, you won't get the results you're after.

If you want to get this sorted out, your best bet is to submit a bug report through Unity's document here: https://unity3d.com/unity/qa/bug-reporting

It's super important to include a reproduction project alongside your report if you actually want your issue to be addressed. Otherwise, it could take ages for anyone else to come across it or for Unity Technologies to address it.

If you make it too difficult for Unity staff to recreate the problem you're experiencing, chances are slim to none that it'll ever get fixed. I know it's not really your job to become Unity's QA or whip up a top-notch bug report, but that's just the way things are.

Once you've submitted your report, you should receive an email confirmation complete with a Case number. Be sure to post that number (without the link) on this forum thread so that the Unity staff that watch this thread can track it down.

 

I've done some bug reports before, and some have worked out well and were indeed resolved. However, I feel a little bit reluctant to submit a bug report for something that is likely not a bug, but rather a general issue with DX12 performance in Unity. The evidence I've seen so far (for example this entire thread) suggests that it is not a bug isolated to some edge case.

For the time being, I will see if Unity responds to this thread again (as they have done a little while ago), and indeed if it turns out that I'm in the minority here, I will submit a bug report.

 

my game with a rtx 3070 (today I got)
DX11
4k 15 fps AVG ???? (WTF?)
1440p 140 fps AVG
1080p 231 fps AVG
DX12
4k 30 fps AVG
1440p 50 fps AVG
1080p 65 fps AVG

 

Is it only in urp? Or the hdrp build is also bad in dx12?

 

Everything I reported is HDRP.

 

So the dx12 is completely broken. What is released in 2017...nice 70% fps down..

 

any news in this? Why DX12 much slower than DX11? Unity 2022.3 LTS

I get 30% of DX11

if DX11 is 100 fps
DX12 is 30 FPS

Also DX11 has a lot of swapchain crash, any fix for that? (daily 1-3 crash, after the whole PC just die, I have to restart the win11 machine) /RTX 3070/ (on Vulkan it is OK, without crash)

 

I'm working on an arcade application that uses multiple displays. I recently updated to 2022 to be able to switch to directX12 to squeeze out some more performance. It absolutely tanked it! When trying to figure out why, I found something interesting.

In a completely brand new URP project with an empty scene, here are my frame times;
*note, in the editor, and 2 cameras each need an active game window .
1 camera/display directX 11 - ~2.5 ms
2 camera/displays directX 11 - ~5 ms
1 camera/display directX 12 - ~2.3 ms
2 camera/displays directX 12 - ~27 ms (not a mistype)

As soon as I add another camera and display, with directX 12, the frame time drops significantly. I can see in the profiler that a huge portion of the time is Semaphore.WaitForSignal (~12 ms). I have V-sync off. Maybe a 2nd display automatically causing v-sync to be enabled?

Interestingly, I don't see this issue in built in pipeline.
Submitted bug report - CASE IN-46509

 

Interesting case indeed. In our case, we just have one camera (HDRP) and the frame rate is always around 50% of the DX11 equivalent.

 

Similar results here with no way to circumvent it on 2022.3 LTS and URP, multiple GPUs. Producing a repro is literally impossible in projects where it appears, ignoring NDA and licensing, I don't think it's a sane approach to pack up a 150gb project into a bug report.
The last few versions of 2022.3 LTS have been throwing constant out-of-memory errors when running DX11 for me and my team-mates so we upgraded to DX12. It's fine on powerful machines running a 3080 but people with a 1060 or 2060 report a 40-50% decrease in framerate. We're talking drops from 60 to 25-30 just from DX12 being toggled on.
So we can either deal with constant editor crashes due to out-of-memory issues or lower the performance for most of the team. It's infuriating and was not an issue for anyone on 2021 LTS or even the first .1 or .2 releases of 2022 this year.

 

Do you get a swapchain issue when get the out-of-memory errors when running DX11 ?

 

Hi

You can use my Discord channel to chat on this, i only had a single case of swapchain crashes and found the way to circumvent the problem, but that can come from anything, so can see what could be if is something similar.

 

OOM could also come from this bug, which as it seems is being fixed soon and related to profiling checkpoints:
https://issuetracker.unity3d.com/is...eprofilerrecorder-when-editor-is-in-play-mode

 

Crashes my editor and throws this error. It happens in random scenes but only in play mode and only in DX11 mode.
I monitor stats on a separate monitor so I know I am nowhere near running out of memory.

 

What happens if you use a different API, such as OpenGL or Vulkan?

 

OpenGL runs like ass so I never used it since I tried a few years back.
Vulkan was a tad bit slower for us - not by much, a 2-5 fps difference for everyone who tried iirc. Didn't stick with it long enough to observe crashes though.

 

I never had a crash with the swapchain in Vulkan, with same shader that gave a 100% crash every time in DX11. But of course this was one specific code case, i think reasons for this issue can vary widly.

 

I think it may be worth checking if Vulkan solves your team's crashing problems. If it does, I imagine a 2-5 fps drop is better than 40%. And sure, this doesn't fix the underlying issue with Unity but hopefully it will be looked into, and this works as a stopgap solution.

 

In some cases i had Vulkan be exponentially faster also, so that 2-5fps drop does not seem bad, given that in some othert scenarios can be much faster than say OpenGL

 

UPDATE

(Please see my contributions further up - this is some follow-up information)

Yesterday I swapped out my RTX 2060 for a RTX 4070. With this swap, I'm no longer GPU bound and only utilising around 35% of the GPU when using DLSS and around 55% when rendering at native resolution of 1440p.

In this scenario, I'm CPU bound on the main thread.

So I checked again when switching between DX11 and DX12, and indeed it is now far closer to parity, at least in terms of the actual frame rate in the editor.

However...

There is still a clear difference in time spend on the render thread during "HDRenderPipeline:Render Main Camera".
DX11 takes 2.2ms, while DX12 takes 3.6ms.

So in conclusion... it seems on slower graphics cards where we become GPU bound, DX12 really falls to pieces compared to DX11. On a faster GPU like 4070, at least on the surface there is much better parity.

One other consideration, is that I went from 6GB to 12GB GPU RAM - I have noticed that DX12 always uses more GPU RAM compared to DX11, and maybe on my 6GB setup, I was running out of GPU RAM and system RAM was dragged into the situation, explaining the massive slow down. This is not confirmed, but it is one line of investigation I'm looking into.

 

That's what we see in our project (fully custom SRP rendering, customized engine based on 2022.3.8f1).
DX12 is using significantly more VRAM than DX11 and once it reaches certain amount of dedicated GPU memory usage, it's not just 5-10% slower than DX11 but 50% slower, utilizing much more shared GPU memory (based on Target Manager -> Performance tab -> GPU activity details).
Unity is looking into both situations and the first one, "normally" slower rendering could be potentially solved by 2023.x split jobs. I'm merging 2023.2 beta now to verify that.

 

It's true, and I think it's just poor memory management by Unity. I can run Unity editor, start my game. Get the same FPS as DX11, and after like a minute it goes downhill.

 

i am in HDRP unity 2022.3.0
Is it normal?



So on a 8 GB VRAM card I get 10 GB usage?

Code (CSharp):

1. gpuMemoryInMB= Profiler.GetAllocatedMemoryForGraphicsDriver() / 1024 / 1024;
2.  
3. "\nGPU VRAM " + SysInfo.GetGraphicsMemorySize() + " MB /" + gpuMemoryInMB + " MB, max was " + _maxGpuMemoryInMB + " MB"

Or how can I get back the correct used memory in the editor vs runtime?

 

Btw, we did a close investigation of the slow D3D12 rendering with Unity team on Unity 2023.2 beta, where D3D12 should be further improved (split jobs etc.). There are two kinds of slowdowns in our game when comparing D3D11 vs D3D12. A standard one (happening all the time, about 10-15%) and a special one (after a while of a game play and world discovery, something happens and rendering perf. is suddenly halved even on places where it used to be fine before, while GPU utilization is suddenly close to 100% instead of 70-80% in 1080p). Unity helped us identify the first one. We are using a lot of shuriken particle systems and that is allocating about 40-60MB per frame just for meshes. D3D12 implementation is using a scratch buffer memory for that of size 1MB, so it was spending the most of the time in reallocation. Modifying engine source and having the D3D12 scratch memory pre-allocated to 100MB is fixing that. The other slowdown is probably about running out of dedicated VRAM (reproduced on RTX3070 with 8GB VRAM, but not on Radeon RX6900XT with 16GB VRAM), after which engine starts to using shared GPU memory (based on Task Manager -> Performance -> GPU stats). That hasn't been solved yet

 

Regarding this secondary type of slowdown... the 50% or more drop in frame rate. Since I talked about this above a few months ago, we've been steadily reducing GPU memory consumption and our frame rate difference is also now around the -10 to -20% range, instead of the insane drops we had. Part of this is no doubt due to memory consumption and running out of dedicated vram - but it's also hard to tell because I have changed from a RTX 2060 to a 4070, and maybe the 2060 was hitting other kinds of bottlenecks in DX12 vs. DX11 that wasn't purely related to dedicated vram depletion.

BTW we're on 2022.3 LTS, and probably not jumping to 2023 until it becomes LTS - but in the meantime I'm just reporting that after doing a lot of optimization of memory usage and also getting batch counts down (another area where DX12 was failing to keep up as well as DX11), things are looking OK now.

 

"after doing a lot of optimization of memory usage and also getting batch counts down"


any help is this area? maybe a blog?

 

(@TomasBlaho tagging you because you mentioned "running out of GPU RAM" as your second slowdown cause.)

I just want to refresh this post with a bit of updated information. As our game world complexity is going up, I noticed we are using more and more GPU RAM. We do a whole lot of mesh combining at play-time resulting in quite a lot of unique mesh data for our main town area. There is also HDRP allocating a large amount of RT memory for all sorts of buffers and effects and stuff. We also have quite a lot of unique textures, etc... there are many reasons why the memory consumption is a little on the high side for us right now...

But the interesting thing is, about two weeks ago I noticed that despite huge effort to try to reduce memory consumption, I was running into crashes or severe slowdowns, as my 12GB RTX 4070 was running out of RAM! I haven't dug into it a lot yet, for example with the memory profiler, because it's just to hard to do with the constant crashing etc...

So - I switched to DX11 and it immediately dropped the GPU RAM usage by HALF. This seems really weird and pretty serious. I will stick to DX11 for a few weeks until I get a chance to dig into this situation a bit deeper. It seems that something is severely broken with DX12 memory allocation, at least in the context of what our game is going.

BTW this is on Unity 2022.3.12. We will move to 2022.3.16 (or 17) soon.

EDIT: the vastly increased GPU RAM usage on DX12, happens in the Editor, but also in Dev and Release builds. It's not just related to the Editor using more GPU RAM.