Nvidia scandal with Async Compute support (DX12)

[nevcairiel]

Oddly, according to Microsoft (https://msdn.microsoft.com/en-us/library/windows/desktop/ff476876%28v=vs.85%29.aspx), DX11.3 supports feature level 12_1.


DX11.3 will get the new hardware features of DX12, but not the new API, and therefor not the CPU overhead reduction - but be available for older versions of Windows as well, while DX12 is not.

Well at least that was the plan. I haven't really heard if that ever came to be like that.

[pandy]

Your whole post is fundamentally wrong.


Well... this is your opinion, i've wrote about NVenc only to show that even if some aspects of 3D are not implemented there is sometimes no alternative solution to NVidia.
NVidia has long tradition to provide partially non working solutions and seem it is still better than other vendors.
Don't get me wrong - i am not NVidia fanboy.

[P.J]

Tie: http://anandtech.com/show/9659/fable-legends-directx-12-benchmark-analysis

[NikosD]

After reading the Anandtech's article regarding DX12 performance of Fable, from now on I will call that site "a disgusting piece of sh!t/ site" or ADPOS regarding the DX12 Nvidia scandal because of the two reasons that I copy here:

Reason 1

We tackled two synthetic tests earlier this year, Star Swarm and 3DMark, but due to timing and other industry events, we are waiting for a better time to test the Ashes of the Singularity benchmark as the game nears completion.


Reason 2

In fact, AMD sent us a note that there is a new driver available specifically for this benchmark which should improve the scores on the Fury X, although it arrived too late for this pre-release look at Fable Legends (Ryan did the testing but is covering Samsung’s 950 Pro launch in Korea at this time)


They are definitely in Nvidia's payroll.

[NikosD]

OK...Anandtech strikes again, this time with Ashes of the Singularity Revisited: A Beta Look at DirectX 12 & Asynchronous Shading here http://anandtech.com/print/10067/ashes-of-the-singularity-revisited-beta

Nvidia still looks ugly with that DX12 engine.

[P.J]

https://m.reddit.com/r/nvidia/comments/3j5e9b/analysis_async_compute_is_it_true_nvidia_cant_do/

[NikosD]

That's a funny article, actually an opinion, by a funny guy.

[P.J]

http://wccftech.com/nvidia-gtx-1080-asynchronous-compute/
http://www.overclock3d.net/articles/gpu_displays/gtx_1080_ashes_of_the_singularity_benchmarks/1

No idea tho' =/

[P.J]

http://www.eteknix.com/pascal-gtx-1080-async-compute-explored/

=/

[NikosD]

Pascal doesn't support hardware async shaders, it offers only an improvement on async computing over Maxwell.

It seems that Nvidia will still concentrate on how to make developers to not use async shaders, rather than implement it until next generation card appears.

[littleD]

Another voice in the dx12/async shaders and all time gpu vendors battle https://scalibq.wordpress.com/2016/05/17/nvidias-geforce-gtx-1080-and-the-enigma-that-is-directx-12/
This is comment from developer, knowledgable guy. He usually stands on intel and nvidia side (which means he usually criticises amd) but posts are informative. Do not enter if you dont want to see fanboism in comments.

[kuchikirukia]

The only scandals in graphics cards are the state of AMD's drivers and the power consumption of their cards.

[chummy]

Async Compute has too much noise from some AMD fanboys but only give 10% gains in performance in GPU bound scenario. In CPU bound side Async Compute do nothing(0) gains.

Dx12 without Async compute increase up to around 100% gains against Dx11 in CPU bound scene. At least with AMD GPU drivers, because you know how bad AMD perform in DX11.

AMD is so bad in DX11 than Nvidia card which do 50% of performance in DX12, is capable do same performance in DX11 CPU bound games.

[P.J]

http://www.anandtech.com/show/10486/futuremark-releases-3dmark-time-spy-directx12-benchmark

http://images.anandtech.com/graphs/graph10486/82854.png

[NikosD]

http://uploads.tapatalk-cdn.com/20160714/6e7036e5d778f414b75313f6c2d976cd.jpg

http://uploads.tapatalk-cdn.com/20160714/01dc5a482dfe56c5221c15f99a434c25.jpg

[nevcairiel]

Graphs entirely out of context, so lets add some context

DOOM Vulkan release notes:
https://community.bethesda.net/thread/54585?start=0&tstart=0

Currently asynchronous compute is only supported on AMD GPUs and requires DOOM Vulkan supported drivers to run. We are working with NVIDIA to enable asynchronous compute in Vulkan on NVIDIA GPUs. We hope to have an update soon.

The Time Spy Benchmark on the other hand supports async on both AMD and NVIDIA, which makes it the first to actually benchmark async compute on Pascal hardware, as far as I know.

[NikosD]

Graphs entirely out of context, so lets add some context



You meant graphs entirely in context, because they are all over Internet.

So, you misspelled in with out of.


DOOM Vulkan release notes:
https://community.bethesda.net/thread/54585?start=0&tstart=0



Hahaha

That is the same phrase we heard for the first time with "Aces" game when we first saw the huge difference between AMD and Nvidia on Async compute.

We are still waiting for the Nvidia patch...

[huhn]

never trust a diagram you haven't faked your self:
http://media.gamersnexus.net/images/media/2016/game-bench/doom/vulkan/vulkan-doom-1080p.png

[NikosD]

I can't believe that German sites like computerbase.de would use photoshop just to support AMD.

Look at the graph.

The performance gain of Vulkan's Async compute is huge especially on Fury cards and all AMD cards in general.

Nvidia gains from nothing to almost nothing, that's a fact.

http://uploads.tapatalk-cdn.com/20160716/cadf9259c13146b5acd91425baa6e119.jpg

[Atak_Snajpera]

fury x has 1.33 more tflops than 1070 so i'm not surprised that it wins here in low level api. In opencl we should notice similar difference.

[NikosD]

RX 480 is 25% faster than 1060 in Vulkan DOOM in 1440p resolution (faster than 980 too) and 32% faster than 1060 in 1080p resolution.

@nevcairiel
Notice the comment that 4GB 980 card can't play DOOM in 1080p using "nightmare" settings due to its 4GB only memory.
The game needs 5GB or more.

http://www.hardocp.com/article/2016/07/19/nvidia_geforce_gtx_1060_founders_edition_review/4#.V48he_l97cd

[huhn]

and the fury , fury x and nano too which are flag ship GPU's unlike the 980.

[NikosD]

The discussion with Nevcairiel was not about specific models.

It was more general about games requiring close or even more than 4GB memory, so cards like 1060 3GB are doomed from the beginning.

By the way, that new 1060 6GB card is doomed from the beginning too.

It's future proof 0% regarding DX12/Vulkan and for that reason it's not worth its money.

[huhn]

yeah you found out that doom 3 runs way better on AMD than nvidia. so what? nothing new.

now look at this nvidia is stopping amd into the ground in this game.

http://www.guru3d.com/articles-pages/geforce-gtx-1060-review,13.html

what did we learn here.
Rise Of The Tomb Raider DX 12 runs way better on nvidia than AMD. and again so what.

[NikosD]

Sorry to say that, but judging from your posts I have told you thousands times that you don't seem to understand the concept on various aspects and that's why you don't seem to learn anything.

Just a few posts earlier, you had even denied that AMD cards are leveraging a lot better the Vulkan API in DOOM and that the graphs are fake!

I don't remember exactly how many times I have told you that even the conversation with you is becoming difficult, unless you do like to troll and so i must not take you seriously...

[huhn]

yeah because we have one vulkan game.

everything is clear. it says everything!

Just a few posts earlier, you had even denied that AMD cards are leveraging a lot better the Vulkan API in DOOM and that the graphs are fake!

you should read it again...

[bidomo]

At the end of the day, is up to the end user's preference, I like both AMD and Nvidia, and I know I will get whatever I like the most or I feel I have a better benefit from.

AMD introduced VCE way before nvidia included the NVENC functionality in their cards.

AMD introduced Hardware Async compute WAAAAAAAAAAY before nvidia.

nvidia has had much more support in GPU video encoding than AMD for very long

Games run usually better in nvidia because they co-own DirectX with Micro$oft, I don't really know why DX12 is low level oriented, maybe nvidia didn't have anything to do with that, and M$ wanted a bit more power for the Xbone GNC architecture, I don't know.

What I know for sure is, nobody is getting paid for standing in the side of one company, so this is pointless, facts vs fanboys never gets anywhere.

[NikosD]

DX is a Microsoft technology that graphics card's giants like Nvidia and AMD try to make it fit better on their own HW.

Due to various reasons (only money) Nvidia in the past had managed to influence Microsoft more than AMD.

But regarding DX12 we could say that the only thing that Nvidia managed to force on MS is to put Fermi in the list of DX12 capable cards.

And Nvidia never released a driver to enable for Fermi! Fermi is just a step before going legacy and Nvidia is still struggling to release a DX12 driver for it, although it pushed MS a lot for that.

DX12 like Vulkan are based A LOT in AMD's own API called Mantle.

Vulkan even more than DX12, that's why Nvidia never told that Fermi will support Vulkan, because their influence in Khronos group which builds OpenGL and Vulkan is not like the influence on Microsoft.

DX12 and Vulkan are still AMD things and Nvidia needs another architecture after Pascal to come closer to AMD, maybe Volta will see.

Vega will destroy Pascal for sure.

And what Nvidia is doing right now in order to keep its fanboys and customers happy until a decent DX12/Vulkan architecture ?

Nvidia does what it know best, paying money for biased benchmarks/ games.

Read about the "Time Spy", so called "DX12" benchmark

http://www.overclock-and-game.com/news/pc-gaming/50-analyzing-futuremark-time-spy-fiasco

[Khanattila]

At the end of the day, is up to the end user's preference, I like both AMD and Nvidia, and I know I will get whatever I like the most or I feel I have a better benefit from.

AMD introduced VCE way before nvidia included the NVENC functionality in their cards.

AMD introduced Hardware Async compute WAAAAAAAAAAY before nvidia.

nvidia has had much more support in GPU video encoding than AMD for very long

Games run usually better in nvidia because they co-own DirectX with Micro$oft, I don't really know why DX12 is low level oriented, maybe nvidia didn't have anything to do with that, and M$ wanted a bit more power for the Xbone GNC architecture, I don't know.

What I know for sure is, nobody is getting paid for standing in the side of one company, so this is pointless, facts vs fanboys never gets anywhere.
Try to implement VCE and then let me know.

[NikosD]

Another AAA game goes to DX12 - Deus Ex: Mankind Divided - another HUGE WIN of Polaris and AMD hardware.

RX 480 exceeds the speed of Nvidia GTX 980 and shows similar performance with 1070 (!) at 1440p.

Take a look:
http://wccftech.com/nvidia-amd-deus-mankind-divided-dx12-benchmarks-performs-poorly-geforce/

[huhn]

you should read the patch it self.

Known DirectX 12 issue:
– There is a known bug that causes some very high end cards to regress relative to DirectX 11. This bug is being addressed by the development team.

Read more: http://wccftech.com/nvidia-amd-deus-mankind-divided-dx12-benchmarks-performs-poorly-geforce/#ixzz4JlFjECkC

the 1070 gets slowed a lot and the 1060 doesn't care in this test.

if you take a closer look even the RX 480 got not that much of a performance boost.

this is just a preview patch anyway.

[NikosD]

Always excuses when Nvidia is behind in performance

[Atak_Snajpera]

you should read the patch it self.


the 1070 gets slowed a lot and the 1060 doesn't care in this test.

if you take a closer look even the RX 480 got not that much of a performance boost.

this is just a preview patch anyway.

What is interesting that RX470 is again as fast as 1060. The same happens in Doom (Vulkan). Not bad for card with 4GiB of VRAM and costing 179$ vs 249$ (1060 6GiB).

[huhn]

in the games benchmark not in the game it self.

the dx 12 version is a simple disaster:
https://www.computerbase.de/2016-09/deus-ex-mankind-divided-dx12-benchmark/2/

[CruNcher]

What is interesting that RX470 is again as fast as 1060. The same happens in Doom (Vulkan). Not bad for card with 4GiB of VRAM and costing 179$ vs 249$ (1060 6GiB).


Yeah in 3d they slowly coming back and Polaris + Vega + Zen will be great Value with the Right Applications especially but Video wise which back in the Rage days ATI was the King with MPEG DCI Hardware Accelleration under Windows MPC that position has been lost since Nvidia came up with Purevideo and their fixed function VPX in Windows NT days :(

VCE/UVD aren't as far as Nvidias Asic they are already 8k capable while AMD is roughly 4K capable with lot of drawbacks since Tegra X1 and VPX7 Nvidia has widen the gap once more.

Though both are still far behind Intel that took the Hyper Speed Train to Perfection when starting Intel HD and their GMA GPU overhaul and Mobile Attack on ARM as well ;)

And still 4 First Generation Intel Cores alone can even compete with 1000s of Cuda Cores 13 sm and their first H.265 Hybrid Decoder in terms of Performance pretty well ;)


Though for most use cases UVD/VCE are OK it's not that it's overall bad but not as advanced yet but for the most consumer purposes absolutely ok, Hybrid VP9 is also not really a big issue even missing VP9 wouldn't be at all for a longer time span we wont see that big of adaption anyways it was a nice test but with AV1 they are all sitting in the Box directly at the Core implementing and improving from the Start for their future Product Launches.

Though especially on the Encoder side i doubt that either AMD nor Nvidia have made progress like Intel did because quality over everything isn't their target @ all there but the Speed and Latency variable of Game Streaming we all know that since the first H.264 Encoder :D

AMD now supports lookahead wow that was basically the only improvement you could read about in their Marketing papers for the Polaris Launch ;)


If someone says to me what would be your Desktop system decision for now i would say hands down Intel + AMD GPU ;)

Intel for all the Video Parts and AMDs GPU for the 3D Async execution path :)

A complete AMD System in the Future will be very exciting ;)

[aufkrawall]

in the games benchmark not in the game it self.

the dx 12 version is a simple disaster:
https://www.computerbase.de/2016-09/deus-ex-mankind-divided-dx12-benchmark/2/
The same was the case with the latest Total War game.
In the in-game benchmark, DX12 showed gains with AMD, but in real game, it was slower as well.
Imho really obvious how AMD tries to lead people astray, they are hardly any better morally than Nvidia.

[NikosD]

In pure close to metal API like Vulkan (DOOM3) and the upcoming Cryengine next month, AMD simply demolishes Nvidia.

Rx 470 can match 1060 6GB card.

But DX12 can be used in a way that can hide the real and pure lack of performance of Nvidia.

It is not a well and pure close to metal API like Vulkan.

Probably Microsoft is waiting for Nvidia to add suitable HW resources in next generation - Volta - or even beyond that to rewrite DX12 in a proper close to metal way like Vulkan.

[huhn]

But DX12 can be used in a way that can hide the real and pure lack of performance of Nvidia.


can you stop been paranoid?

[NikosD]

can you stop been paranoid?
Can you stop being so ignorant ?

http://www.overclock-and-game.com/news/pc-gaming/50-analyzing-futuremark-time-spy-fiasco

[huhn]

nothing to see in the link. just an "Analyze" of an worthless synthetic benchmark.

[aufkrawall]

In pure close to metal API like Vulkan (DOOM3) and the upcoming Cryengine next month, AMD simply demolishes Nvidia.

Those APIs offer a lot more control over the GPU than DX11/OGL for the developer, but they are not "pure close to metal".

I replaced my R9 390 with a GTX 1070. In Doom's multiplayer mode (it's not Doom 3 btw., it's a reboot) the 1070 often is ~50% faster. I don't see the problem for Nvidia...


Rx 470 can match 1060 6GB card.

There is not a big difference between 470 and 480, so what?


But DX12 can be used in a way that can hide the real and pure lack of performance of Nvidia.

Poorly optimized DX12 ports run also worse on AMD, see Deus Ex.
And Talos Principle btw. also uses Vulkan and runs worse than DX11 on both AMD & Nvidia.

You have some utter misunderstandings.

[NikosD]

Those APIs offer a lot more control over the GPU than DX11/OGL for the developer, but they are not "pure close to metal".


You don't seem to understand the concept and know what are you talking about.

Both DX12 and Vulkan are basically different implementation and general-purpose port of AMD's proprietary Mantle close to metal API.

Both DX12 and Vulkan are close to metal APIs that moved the weight of optimizations and writing good code from the drivers to the developer of the game.

What I'm saying is that Vulkan is closer to Mantle and closer to the metal than DX12 and a good game developer of DX12/Vulkan can take advantage of the HW resources of AMD's cards which were built for Mantle.

So DX11 has nothing to do with DX12 and OpenGL with Vulkan.

[aufkrawall]

Both DX12 and Vulkan are basically different implementation and general-purpose port of AMD's proprietary Mantle close to metal API.

Where do I find the term "close to metal" by Khronos or Microsoft?


What I'm saying is that Vulkan is closer to Mantle and closer to the metal than DX12 and a good game developer of DX12/Vulkan can take advantage of the HW resources of AMD's cards which were built for Mantle.

Just because a lot of API calls have similar names, it doesn't mean that Mantle and Vulkan share most of the DNA.


So DX11 has nothing to do with DX12 and OpenGL with Vulkan.
I don't think you read my post correctly.

[CruNcher]

It's not close to metal it's explicit you can't allow close to metal access on the PC for some simple reasons on the side of security.

The Console Ecosystem works differently first you have shitloads of NDA Second you have very advanced Hypervisor that got hardened more and more after attack over attack, and are used these days even for timeable hardware resource access (Xbox One Indie Dev Mode).

On the PC side the same has been tried to establish with UEFI but it's not sure if it's gonna succeed and Microsoft isn't sure either even on Windows 10 the xs for 3rd party applications on Windows has been getting narrowed down from version to version because mostly of content protection reasons a big part now plays the driver that has to ensure certain protections and protected paths.

Low Level xs without correctly functioning Hypervisor and protected Zones (Trusted Zones,PMPs) would be a big problem of keeping the knowledge secret.

And yes it's about Content Protection not so much about AVG Users Content Protection that is a second goal commercial objective but most off all it's about System Control a System inside a System that the User has nio xs no control at all over (silently sleeping) and can any time used against him if wished..

[NikosD]

Close to metal = low level API

[NikosD]

Fastest performance ever for DX12 and AMD playing a super AAA game - Battlefield 1

http://wccftech.com/battlefield-1-directx-12-benchmarks-amd-nvidia/

[huhn]

here the first "better" still not 100% proper test to BF 1.
it is not really possible to do a proper test with his game.

https://www.youtube.com/watch?v=ZxsIOV2AjMc

and the same image as "always".
AMD fails at DX11 and nvidia lose some performance with DX12.

a good driver is still better than DX12...

[CruNcher]

Fast Motion Streaming Scenarios are the Hardest for the System and Nvidia is excellent at them even in DX11 :)

https://www.youtube.com/watch?v=N_ISj656Zkw

Nvidia gains a + 10 FPS improvement on the same lower shader load scenario here at High Motion.

That's enough overhead to get 60 FPS stable.

[huhn]

one of the major problem of AMD is the problem the GPU are not fully utilized.

nvidia doesn't have this problem the DX 11 driver is utilizing the GPU really good and the CPU overhead is a lot lower too. so they fixed the problem without Async Compute. so i don't understand why Async Compute is important or generally needed.
the biggest problem of AMD is there driver.

[CruNcher]

That's not the whole story the DCE efficiency plays into this as well ;)

Nvidia did not fixed the Multi-threading efficiency of Windows they worked around it on the Driver level and that is generally a bad idea as it has drawbacks then todo it right.

AMD was working behind closed doors on a better solution as a whole, while Nvidia mainly uses this workaround as a Weapon and Product feature and to save time for bigger improvements.


Though time is running out ;)

The release of Pascal (Paxwell) was the most hysterical,panical and fear driven release i ever saw from Nvidia since some time.

It made me to invest some more into AMD, because of it.