Vulkan/OpenGL Next Generation Initiative: unified API for mobile and non-mobile devices.

  • Thread starter Deleted member 13524
  • Start date
One API to rule them all...
Anyway supporting fewer API => more time spent on other things.
 
Seems NV's Vulkan path has eventually caught up with GCN's in Doom. The 1070 is slightly ahead of the FuryX, while the 980Ti is close by. (Reviews using the latest 378 NVIDIA driver)

Doom_Vulkan_Average_FPS.png

http://techreport.com/review/31562/nvidia-geforce-gtx-1080-ti-graphics-card-reviewed/5
http://www.techspot.com/review/1352-nvidia-geforce-gtx-1080-ti/page2.html
https://www.purepc.pl/karty_graficz..._ti_gaming_x_pascal_bardzo_wypasiony?page=0,8
http://www.tomshardware.com/reviews/nvidia-geforce-gtx-1080-ti,4972-3.html
 
Not that I'm saying that NV drivers are perfect and always were... But is it just the driver though? I mean Doom now may be quite different then Doom half a year ago. There has been quite a few patches and it's not like Doom is retested every time a new patch or a new driver gets released.
 
BRiT said:
So it wasnt because Vulkan was an AMD Low Level Construct but because Nvidia Vulkan Drivers just weren't up to par.
Click to expand...
There is only so much you can extract using drivers, Shader Intrinsics is one of the key advantages AMD had with Vulkan in Doom. And it is a GCN only feature.
 
I thought Shader Intrinsics was found in both AMD and Nvidia. Implimentation might be different but to the same end.

Reading Between The Threads: Shader Intrinsics
When writing compute shaders, it’s often necessary to communicate values between threads. This is typically done via shared memory. Kepler GPUs introduced “shuffle” intrinsics, which allow threads of a warp to directly read each other's registers avoiding memory access and synchronization. Shared memory is relatively fast but instructions that operate without using memory of any kind are significantly faster still.

This article discusses those “warp shuffle” and “warp vote” intrinsics and how you can take advantage of them in your DirectX, OpenGL, and Vulkan applications in addition to CUDA. We also provide the ShuffleIntrinsicsVk sample which illustrates basic use cases of those intrinsics.
Click to expand...
https://developer.nvidia.com/reading-between-threads-shader-intrinsics
 
pharma said:
I thought Shader Intrinsics was found in both AMD and Nvidia.
Click to expand...
They are, I just meant that when Doom was first introduced, it only supported AMD's Intrinsics. At that time that wasn't clear enough. AMD was first to introduce it as that was ported straight from consoles. It remains to be seen whether Doom added support for NV's intrinsics or not. That wasn't officially announced by either NV or the developer.
 
MDolenc said:
Not that I'm saying that NV drivers are perfect and always were... But is it just the driver though? I mean Doom now may be quite different then Doom half a year ago. There has been quite a few patches and it's not like Doom is retested every time a new patch or a new driver gets released.
Click to expand...
What Doom got since last August are just paid DLC content packages. No patch to the actual game rendering code has been added and no rendering issues have been addressed since [1].... and I know I am beating a dead horse ;)

[1]
 
pharma said:
I thought Shader Intrinsics was found in both AMD and Nvidia. Implimentation might be different but to the same end.

Reading Between The Threads: Shader Intrinsics

https://developer.nvidia.com/reading-between-threads-shader-intrinsics
Click to expand...

Nvidia is still to fully update their extensions and added to them in Nov/Dec 2016 for Vulkan.
And that ties nicely into the article you linked where in one paragraph the Nvidia engineer says:
Nvidia said:
In parallel, we are also working with the respective Khronos working groups in order to find the best way to bring cross-vendor shader intrinsics already standardized in OpenGL over to Vulkan and SPIR-V. We are furthermore working on Vulkan and SPIR-V extensions to expose our native intrinsics, but prioritized the cross-vendor functionality higher, especially since there is notable overlap in functionality.
Click to expand...


But the improvements in context of those reviews would be down to drivers but still not entirely convinced those links use best setting for AMD (could also depend upon map).
Here is an example of how it has been improving from HardOCP, probably would be more without Nightmare Settings.

1489379901fkk6C8hkMC_9_1.png


But I am not convinced about those review results linked earlier when compared to say PCGamesHardware that has current drivers and TSSAA enabled; 980ti is about equal to the 580, FuryX around 8% faster than the reference 1070.
http://www.pcgameshardware.de/Radeo...939/Tests/RX-570-Review-Benchmarks-1225896/2/
Maybe map is also influencing their results but this hints maybe Async Compute (that provides roughly 8% in Doom) may not be active in those other reviews *shrug*.

Anyway Nvidia will only make a true jump once their own additional Vulkan extensions are complete (still ongoing though some more are now available as mentioned beginning of post) and used by the dev just like AMD's currently are in Doom.
Cheers
 
Last edited:
Nvidia is still to fully update their extensions and added to them in Nov/Dec 2016 for Vulkan.
Click to expand...
If those reviews are using the most current drivers, I'm not sure whether they include Vulkan extensions found in the current "Beta" drivers. Since they are still developmental drivers I don't know the timing when changes are integrated. But below is a list of changes included since January in Vulkan "Beta" developer's drivers, but most likely not included in WHQL or Hotfix drivers.
Developer Beta Driver Release Updates
March 27th, 2017 - Windows 377.14, Linux 375.27.15

  • SPIR-V compiler bug fixes
  • Updated Vulkan loader to version 1.0.42.1
March 15th, 2017 - Windows 377.07

  • Bug fixes
March 8th, 2017 - Linux 375.27.14, Windows 377.06

  • Bug fixes
February 28th, 2017 - Linux 375.27.13, Windows 377.01

  • Fix issue with SteamVR shaders
February 27th, 2017 - Linux 375.27.12, Windows 376.98

February 9th, 2017 - Linux 375.27.10

  • Bug fixes
February 1st, 2017 - Linux 375.27.08, Windows 376.80

  • Bug fixes
January 23rd, 2017 - Linux 375.27.07, Windows 376.71

January 10th, 2017 - Linux 375.27.03, Windows 376.66

  • Vulkan beta drivers with experimental API interop features
Click to expand...
 
Last edited by a moderator:
CSI PC said:
But I am not convinced about those review results linked earlier when compared to say PCGamesHardware that has current drivers and TSSAA enabled
Click to expand...
TechReport and PurePC specifically mention they enabled TSSAA 8X in their testing. And even without the AA, FuryX was always faster than 1070 until recently. Also apparently PCgameshardwrae don't use the 378 driver yet.

Only for the Geforce results, we sometimes get results from the 378 branch on time - since the newer drivers only improve the performance in the most recent game releases, we do not expect any impact on our results.
Click to expand...
 
DavidGraham said:
TechReport and PurePC specifically mention they enabled TSSAA 8X in their testing. And even without the AA, FuryX was always faster than 1070 until recently. Also apparently PCgameshardwrae don't use the 378 driver yet.
Click to expand...
PCGameshardware do, the review I linked used (378.xx), and they also used 378.78 for the 1080ti review but importantly the results pretty much align to the Nvidia performance in the earlier linked 580 review.
So maybe does come down to map benchmarked on, as you rightly mention TechReport do use TSSAA (did not realise mentioned in the OpenGL benchmark rather than Vulkan).

Separately worth noting that AMD drivers were still improving performance as well in Doom Vulkan, albeit a bit behind the relative gains seen by Nvidia.
Cheers
 
Last edited:
pharma said:
If those reviews are using the most current drivers, I'm not sure whether they include Vulkan extensions found in the current "Beta" drivers. Since they are still developmental drivers I don't know the timing when changes are integrated. But below is a list of changes included since January in Vulkan "Beta" developer's drivers, but most likely not included in WHQL or Hotfix drivers.
Click to expand...
Seems Doom has not been updated itself to use the new Nvidia extensions, so stuck with the smaller driver optimisation gains relative to those.

Cheers
 
DavidGraham said:
Maybe we can chart out the different Doom maps with their IHV preference.

I will add ComputerBase tests as well, with the same result: 1070 > FuryX.
https://www.computerbase.de/2017-04/radeon-rx-580-570-test/2/#diagramm-doom-2560-1440
Hardware.fr results (though with Nightmare quality and TSSAA 8X)
http://www.hardware.fr/articles/957-13/benchmark-doom.html
Click to expand...
Bear in mind that FuryX may have issues with Doom at 1440p depending upon review/something else; while the Computerbase results at 1440p quite poor notice FuryX is near the top of the benchmark and just 6% behind 1070FE when set to 1080p.
The other reviews has the Fury X around 6-10% behind (except PCGamesHardware) the 1070FE at 1440p and higher, which makes one wonder what happened in the Computerbase.de benchmark as they use same map as a couple of others (see below).

The Challenge is identifying the map for each site.
Seem to be same map:
Computerbase.de: https://www.computerbase.de/2016-05/grafikkarten-testsystem-2016/10/
hardware.fr: http://www.hardware.fr/articles/957-13/benchmark-doom.html
Techreport seems to be same map as well but started earlier in the section :

Different map:
PCGameshardware uses a different map and with no monsters on screen to keep consistency (maybe this could be critical to performance behaviour):

We need more results with maps and how ran to make more of an informed decision, but interesting couple of differences are map and PCGamesHardware without enemies for consistency, so either of these two variables.

Sorry, seems it wants to auto change a couple of those links to media instead of keeping as url.
Cheers
 
Last edited:
I was wondering why the pcgameshardware results were off compared to the others, naturally testing empty maps is not indicative of real performance, nor does it stress the GPUs. A better benchmark is to record the average of multiple runs of combat scenes.
 
DavidGraham said:
I was wondering why the pcgameshardware results were off compared to the others, naturally testing empty maps is not indicative of real performance, nor does it stress the GPUs. A better benchmark is to record the average of multiple runs of combat scenes.
Click to expand...
Yeah and like I mentioned possibly that or map.
The problem is that having spawned enemies removes consistency, but it needed to be validated if the enemies impact performance between manufacturers with said option settings and how much real world differs to the empty zone.
But possibly more important.
PCGameshardware is the only one using that specific map/zone, but the other 3 I managed to identify and have a reverse of performance all seem to use identical map/zone to each other.

So ideally needs PCGameshardware to test the map/zone used by those other 3 publications, or one of those publications using the zone that PCGameshardware does.

It could be map/zone as I notice in some recent benchmark reviews that Hitman DX12 ironically is better on Nvidia equivalent tier GPUs these days since the Hitman driver update while other reviews using right driver still have same tier behind.
Cheers
 
Last edited:
You must log in or register to reply here.

Similar threads

S
good websites for devices & SoC reviews
Replies
0
Views
1K
sanjeev2.k
S
B
Why isn't programmable blending more widely supported? (History and discussion)
Replies
13
Views
11K
Ethatron
Ethatron
J
Will OpenGL ES 3.0 prove to be a short-lived standard?
2 3
Replies
53
Views
18K
jedibeeftrix
J
M
AMD Mantle API [updating]
102 103 104
Replies
2K
Views
442K
lanek
L
G
  • Locked
AMD's Unified Gaming Strategy and the JHH Titanic
Replies
0
Views
3K
grokzilla
G
Back
Top