No DX12 Software is Suitable for Benchmarking *spawn*

The only game im aware of that shows consistent benefit to a large pool of both IHVs GPUs when using a low level API is that sniper game. Everything else is slower on nvidia. Some specific scenarios when CPU limited will show a benefit, like Doom for example but those are the exception and not the rule. Turing has helped some corner cases. AMD fares much better on avg with the transition. I think there may be rough times ahead for PC games.
 
For Turing, DX12 does nothing but hurts performance by as much as 12%, for Navi it doesn't improve fps at all, but it improves frame times in a significant way, still doesn't even matter because the frame times were atrociously bad on Navi @DX11, the only thing DX12 does is improve Navi's frame times to a level that is close to Turing's frame times.

Overall DX12 doesn't lead to any fps gains on any GPU architecture.

https://www.computerbase.de/2020-01.../#diagramm-directx-11-vs-directx-12-1920-1080

Why do you brush off the improvement of frame times so easily? Did the DX11 path suffered from some bug?
The comparison linked was done with high end CPU, I wouldn't expect much of a difference.
 
Why do you brush off the improvement of frame times so easily? Did the DX11 path suffered from some bug?
Yeah on AMD it did, frame times were unusually bad on AMD's Navi.

The only game im aware of that shows consistent benefit to a large pool of both IHVs GPUs when using a low level API is that sniper game.
Shadow of Tomb Raider and Hitman 2 are also among the best, they showed very large fps gains using DX12 on all hardware.
 
Yeah on AMD it did, frame times were unusually bad on AMD's Navi.


Shadow of Tomb Raider and Hitman 2 are also among the best, they showed very large fps gains using DX12 on all hardware.

The DX11 implementation in Shadow is extremely poor on the CPU performance front so its hard to expose but pre Turing is faster under DX11 when GPU limited. Hitman 2 i thought dropped DX12 support. I remember 1 being slower on Nvidia when GPU limited.
 
Everything else is slower on nvidia.
That's not really true. Most modern D3D12/VK renderers show comparable performance to that of D3D11 counterparts on Pascal and Turing h/w. Some are slower - which is usually a sign of badly programmed D3D12/VK renderer as they tend to have issues on all h/w, not just NV's. Some are faster, even in GPU limited scenarios.

But in theory why would a D3D12/VK renderer be faster than a D3D11 one when running in a GPU limited scenario and not actively using any performance features of D3D12/VK absent from D3D11? Right now if some D3D12/VK renderer is considerably faster than its D3D11 counterpart in a GPU limited load then it's more often than not a sign of D3D11 driver being bad at its job and less about such renderer being good.
 
but pre Turing is faster under DX11 when GPU limited
No it was faster on all GPUs, though I agree the DX11 appears to be broken.

For maybe the first time ever, Shadow of Tomb Raider appears to have a very solid DX12 implementation that is 230% faster than DX11 in one test!
https://www.purepc.pl/karty_graficz...pc_test_wydajnosci_kart_i_procesorow?page=0,3

Though this appears to be a result of a massively broken DX11 renderer. Exhibiting massive CPU limitations on all GPUs.

15138
15141

https://www.sweclockers.com/test/26...-och-forza-horizon-4-demo-pa-geforce-rtx-2000


The resulted are repeated in other sites as well.
PCGH:
DX12 is 200% faster than DX11 in their scene
http://www.pcgameshardware.de/Shado...Shadow-of-the-Tomb-Raider-Benchmarks-1264575/

PClab:
DX12 is 200% faster than DX11 in their scene
https://pclab.pl/art78818-5.html

ComputerBase:
DX12 is 12% faster than DX11 in their scene
https://www.computerbase.de/2018-09...ottr-mit-finalen-treibern-und-patch-1920-1080

That leaves the DX11 path broken, it can't produce fps beyond a certain point. My theory is that the studio had limited resources that are obvious in this title. They will also integrate RTX which will need DX12, so instead of developing two paths, they focused on DX12 for RTX. And DX11 got shafted, maybe it will improve with a later patch



Hitman 2 i thought dropped DX12 support.
Hitman 2 reintroduced DX12 back again through a patch, results were stellar.

DX12 is now added to Hitman 2, resulting in MASSIVE fps uplifts over DX11, I personally witnessed a 30fps increase on my 3770K in one of the sniping levels with huge draw distance and object count. DSOG experienced 20fps gains on their 4390K in crowded areas, others gained more on their 8700K!

https://www.dsogaming.com/news/hitm...-40fps-increase-in-cpu-ram-limited-scenarios/

This is quite possibly the best implementation of DX12 to date.

Also according to GameGPU, Control appears to be faster with DX12 on all GPUs as well.

https://gamegpu.com/action-/-fps-/-tps/control-test-gpu-cpu-2019
 
Last edited:
No it was faster on all GPUs, though I agree the DX11 appears to be broken.









Hitman 2 reintroduced DX12 back again through a patch, results were stellar.



Also according to GameGPU, Control appears to be faster with DX12 on all GPUs as well.

https://gamegpu.com/action-/-fps-/-tps/control-test-gpu-cpu-2019

https://www.overclockersclub.com/reviews/tomb_raider_shadow_performance/

The most recent Shadow testing i know of, the NV GPUs that werent CPU limited under DX11 both slow regressions when going to DX12. Your Computerbase result shows it as well when you get to 1440p and above. PCGH shows it at 4k.

That's not really true. Most modern D3D12/VK renderers show comparable performance to that of D3D11 counterparts on Pascal and Turing h/w. Some are slower - which is usually a sign of badly programmed D3D12/VK renderer as they tend to have issues on all h/w, not just NV's. Some are faster, even in GPU limited scenarios.

But in theory why would a D3D12/VK renderer be faster than a D3D11 one when running in a GPU limited scenario and not actively using any performance features of D3D12/VK absent from D3D11? Right now if some D3D12/VK renderer is considerably faster than its D3D11 counterpart in a GPU limited load then it's more often than not a sign of D3D11 driver being bad at its job and less about such renderer being good.

Comparable in that most times its not dramatically slower, but the gap is there the vast majority of the time in GPU limited scenarios. At least on pre Turing.
 
Last edited:
Games designed for DX12 were expected to take advantage of the massive reduction in draw call overhead to produce more varied and detailed environments. Those games were expected to choke on DX11 drivers.

We haven’t seen games like that yet because console CPUs are crap and there’s still a ton of DX11 PC hardware out there that devs can’t ignore.

I suspect we will only see DX11 really choke when devs start targeting next gen consoles.
 
The most recent Shadow testing i know of, the NV GPUs that werent CPU limited under DX11 both slow regressions when going to DX12. Your Computerbase result shows it as well when you get to 1440p and above. PCGH shows it at 4k.
Sigh, I guess you are right, so we can scratch Tomb Raider off the list.

We now only have Hitman 2, Control and Sniper Elite 4, maybe Metro Exodus to some degree.

Comparable in that most times its not dramatically slower, but the gap is there the vast majority of the time in GPU limited scenarios. At least on pre Turing.
Tracking most DX12 titles in this thread, I would say AMD hardware also tend to either show no gains from DX12, or show very limited gains. There are cases where DX12 gives AMD hardware huge gains, but that's because their DX11 performance where atrocious to begin with in these cases, so DX12 only acts as the buffer that propels AMD hardware to near NVIDIA DX11 performance.

Tomb Raider, Metro and Control all run better on AMD.
I will take the opportunity to revisit this subject again in the context of DX12, based on the latest benchmarks form Computerbase, it appears that Metro Exodus is still faster on Pascal than Vega.

https://www.computerbase.de/2019-12...cours-2020/2/#diagramm-metro-exodus-2560-1440
https://gamegpu.com/action-/-fps-/-tps/metro-exodus-test-gpu-cpu-2019

Shadow Of Tomb Raider is a tie.
https://www.computerbase.de/2019-12...#diagramm-shadow-of-the-tomb-raider-2560-1440

Control is a tie as well.
https://www.computerbase.de/2019-12/grafikkarten-testparcours-2020/2/#diagramm-control-2560-1440

New Games designed for DX12 were expected to take advantage of the massive reduction in draw call overhead to produce more varied and detailed environments. Those games were expected to choke on DX11 drivers.
Indeed, but we should have received some benefits on current games too, reducing draw calls bottleneck should prevent games from being CPU limited when using high draw distance settings, we don't see that at all though.
 
Indeed, but we should have received some benefits on current games too, reducing draw calls bottleneck should prevent games from being CPU limited when using high draw distance settings, we don't see that at all though.
On the contrary, we see this in pretty much every D3D12/VK title released to date, with only some exceptions. New renderers tend to be faster in CPU limited scenarios on all h/w, including even older Kepler GPUs.
 
Question is, are the gains worth the devs efforts ?
If you want to have any performance on the consoles; then the answer would be yes.
They should already have a path of DX12 available for most multiplatform releases.

Additional effort would be required to optimize it for PC; but it's not like they are re-writing their renderer. Drivers still need to improve (and will likely continue) as PC matures into low level APIs.
 
If you want to have any performance on the consoles; then the answer would be yes.
They should already have a path of DX12 available for most multiplatform releases.

Additional effort would be required to optimize it for PC; but it's not like they are re-writing their renderer. Drivers still need to improve (and will likely continue) as PC matures into low level APIs.

Why multiplatforms games (so, being on xbox series too) doesn't give us a dx12 path then ? Too hard to debug/ optimize for a open plateforme ?
 
Why multiplatforms games (so, being on xbox series too) doesn't give us a dx12 path then ? Too hard to debug/ optimize for a open plateforme ?
Well the drivers are going to be better console and there is only 1 hardware profile to cater for, and in this case the console setup is GCN based. Streamlining the number of configurations makes coding DX12 simpler I suspect.

But if you're going to use that as a code base and apply it to Intel and Nvidia, I don't know how performant it would be , or if you'd introduce instability into your game. I definitely believe that some work needs to be done.
 
Back
Top