computerbaseAshes of the Singularity Beta1 DirectX 12 Benchmarks

[Hitman928]

I thought that is what I just said.

Yes, I was just backing up what you said with a quote from the release notes.

 

[TheELF]

I think you missed the part where he said "No difference at all."

And I think you missed the part where he said "The Async Compute mode is slightly faster (<5%). "
Yes, less then 5% is nothing that someone would notice if he wasn't looking for it but it is still a difference.

What "Extremes" are these?

The ones I bolded out?
"The CPU side threading has been removed. " The amd crowds mantra is that Dx12 will use more CPU cores so why disable it?
And "The sample is now queuing up to 4 frames, instead of 2 " pre-render of 4 frames is completely useless in reality, ask anybody who's tried it.

 

[Mahigan]

This article was from december 2014 after Maxwell v2 was launched and nVidia had announced FinFet for Pascal. Obviously nVidia will skip 20nm for their GPUs... :\
Where is the information that nVidia announced Maxwell will be using 20nm?



The line is very clear to me. nVidia gamers would be concerned about Pascal because of your fanfiction. Yet AMD's user dont care about missing features like CR, ROV...

You are creating stories over stories to hype AMD and you cant even back them up like this:


Tell us how we can ask the driver about it. With your insider information is should be easy to link to a article or programm.

I'm not creating stories..
http://www.techpowerup.com/mobile/2...s-amd-gcn-better-prepared-for-directx-12.html

And where did they get that quote? I invited Oxide over to overclock.net. Here's where the quote is from:
Second link page 122 of the thread:
https://www.google.com/search?q=oxi...#q=oxide+async+compute+disaster+overclock.net

Many other links reporting it there.

Now you want to know how I know all of these things? If you read the thread you'll find out.

 

[Mahigan]

https://forum.beyond3d.com/threads/intel-gen9-skylake.57204/page-6#post-1869935

https://forum.beyond3d.com/threads/intel-gen9-skylake.57204/page-7#post-1869983

Don't ever bother arguing with these ADFs, they are clueless as hell and don't know anything except to slander and FUD about things they don't understand.

No,

He said that at the beginning of Sept of 2015. Right after I theorized that nvidia wasn't compatible with Async Compute. Further testing has proven the hypothesis that nvidia don't support Async Compute under DX12. Same website and forums as well:
Read the whole thread... https://forum.beyond3d.com/threads/dx12-performance-discussion-and-analysis-thread.57188/page-36

Nvidia are able to execute the code synchronously but not asynchronously. This isn't a problem when you have a relatively small work load but as the work load becomes higher, nvidia cards begin to crash the Windows gpu driver.

I'm not making any of this up.

 

[ShintaiDK]

I'm not creating stories..
http://www.techpowerup.com/mobile/2...s-amd-gcn-better-prepared-for-directx-12.html

And where did they get that quote? I invited Oxide over to overclock.net. Here's where the quote is from:
Second link page 122 of the thread:
https://www.google.com/search?q=oxi...#q=oxide+async+compute+disaster+overclock.net

Many other links reporting it there.

Now you want to know how I know all of these things? If you read the thread you'll find out.

Do you have any other statements than from (AMD sponsored)Oxide that have rounded wccftech usually?

 

[Mahigan]

Do you have any other statements than from (AMD sponsored)Oxide that have rounded wccftech usually?

I already put up a Beyond3D test which confirmed it ans further testing has also confirmed it. Findings published here:
http://ext3h.makegames.de/DX12_Compute.html

As for Oxide being AMD sponsored, they worked far more with nvidia than AMD. Even implementing a separate path when it was discovered that Async compute did not work on nvidia cards. The separate path was at the request of nvidia and is fully nvidia optimized.

 

[ShintaiDK]

I already put up a Beyond3D test which confirmed it ans further testing has also confirmed it. Findings published here:
http://ext3h.makegames.de/DX12_Compute.html

So no. So what happens if AOTS isn't showing what will be the standard for DX12 games?

 

[Kenmitch]

Do you have any other statements than from (AMD sponsored)Oxide that have rounded wccftech usually?

Do you have a statement from NVidia that shows that they fully support it?

 

[Mahigan]

So no. So what happens if AOTS isn't showing what will be the standard for DX12 games?

Actually yes, Fable Legends.

Asynchronous compute testing showed the same thing:
http://wccftech.com/asynchronous-compute-investigated-in-fable-legends-dx12-benchmark/

Second page:
http://wccftech.com/asynchronous-compute-investigated-in-fable-legends-dx12-benchmark/2/

Fable Legends uses very little Async Compute, less than AotS. Hitman will truly show off the feature.

 

[Glo.]

Shintai, not that I am not agreeing with you, or anything, but...

Tell me what has Nvidia Maxwell on hardware level to handle workloads?
And second thing. Is MegaThread Engine capable of doing anything WITHOUT drivers that commands both hardware(Maxwell) and the application itself?

I am asking this because it will be the reality what you will have to handle with DirectX 12. The application itself will handle hardware, without anything between them. Asynchronous Compute is pointless on Nvidia cards, from architectural point of view. It will always tank the performance on green team hardware, unfortunately. It would not if... there would be second Asynch Engine and Hardware Scheduler.

 

[sontin]

I'm not making any of this up.

Exactly, while I don't doubt nvidia's support for 12_1, it is true that "the nvidia driver exposed Async Compute as being present" but when Oxide went to use it it was "an unmitigated disaster". Oh and I didn't say what's in quotations, Oxide did.

I still wait for any proof that Microsoft has designed a way to get this information about multi engine support from the driver.

 

[Mahigan]

I still wait for any proof that Microsoft has designed a way to get this information about multi engine support from the driver.

Umm, Microsoft tools query the driver in order to determine feature support.

All GPU tools query the driver, see GPU Caps Viewer:
http://www.ozone3d.net/gpu_caps_viewer/

Oxide did the same, the nvidia driver exposed the feature, Async Compute, as being available but once they attempted to use it... Disaster.

Nvidia wanted Oxide to "shut it down" meaning remove all Async Compute from the game. Oxide refused and came to an agreement on creating a vendor ID specific path.

Again, drivers expose features and tools which look for feature support query the driver.

Now let's ask nvidia:
https://developer.nvidia.com/nvemulate

" By default, the NVIDIA OpenGL driver will expose all features supported by the NVIDIA GPU installed in your system."

Or Microsoft:
https://msdn.microsoft.com/en-us/library/jj134360(v=vs.85).aspx

" DirectX 9 drivers (which expose any of the D3DDEVCAPS2_* caps)"

Drivers expose features.

 

[ShintaiDK]

Actually yes, Fable Legends.

Asynchronous compute testing showed the same thing:
http://wccftech.com/asynchronous-compute-investigated-in-fable-legends-dx12-benchmark/

Second page:
http://wccftech.com/asynchronous-compute-investigated-in-fable-legends-dx12-benchmark/2/

Fable Legends uses very little Async Compute, less than AotS. Hitman will truly show off the feature.

I see you really like wccftech.

So Fable is DX12. And how is the GPU performance there?

http://techreport.com/review/29090/fable-legends-directx-12-performance-revealed

Or is there too much tessellation there?

 

[Azix]

The unfortunate thing is its hard to find software to record directx 12. Would record what was showing in the benches I ran. OBS failed. Afterburner nah.

 

[Mahigan]

I see you really like wccftech.

So Fable is DX12. And how is the GPU performance there?

http://techreport.com/review/29090/fable-legends-directx-12-performance-revealed

Or is there too much tessellation there?

Very little Async Compute as I already mentioned, and you can witness with your own eyes in those graphs from wccftech, but you missed something...

http://techreport.com/review/29357/amd-radeon-software-crimson-edition-an-overview

" Radeon owners may also see improved performance in general from Crimson. In a best-case scenario with an AMD Radeon R9 Fury X and Crimson, the company says it achieved a 20% improvement in average FPS with the*Fable Legends*DirectX 12 benchmark*running at 1080p"

Take that score of 75fps and add 20%.
You're looking at 90fps now with that CPU configuration in that test. Tech Report did their own testing on a weaker CPU and they did see a 20% increase and were able to confirm it. Again though, Fable Legends makes mild usage of Asynchronous compute. Nvidia cards can run the work loads synchronously.

The argument here is what will Hitman bring. According to AMD and IO Interactive, heavy use of Asynchronous compute. Async compute will be leveraged to a degree that has not yet been seen.

If we take that AotS used only a few mild post processing effects (as confirmed by Oxide) and it was an "unmitigated disaster" then imagine what Hitman DX12 will do to nvidia cards running the Async path?

That's the argument, you've all danced around it, trying to attack the argument from various angles (you and Sontin) but so far I've pretty much proven my point.

The real proof will come when Hitman DX11 and DX12 release. AMD will perform horribly under DX11 but very nicely under DX12. If nvidia can run the same DX12 path as AMD, not likely, their cards will perform poorly. Basically.. Roles reversed.

 

[ShintaiDK]

But the result matters doesn't it? What if the vast majority of games look like this?

Add some extra tessellation and AMD performance goes bonkers.

 

[Mahigan]

But the result matters doesn't it? What if the vast majority of games look like this?

Add some extra tessellation and AMD performance goes bonkers.

It's not "a little tessellation". AMD GCN can handle up to a factor of x15/x16 without suffering the small triangle problem in GCN1,2,3. It's only when you over tessellate that GCN1,2,3 tanks. Of course Tonga and Fiji fare better than Tahiti and Hawaii. Polaris will likely rectify this. Hence the brand new geometry processing units.

This was also part of my argument which you and Sontin objected too. Polaris is vastly different from GCN1,2,3 because almost ever key GPU component was redesigned (not tweaked but replaced hence why AMD calls them "new").

 

[ShintaiDK]

Can we agree that the only thing that matters is the result and what people will buy and play? Tessellation, Async or not. Its all secondary.

 

[Mahigan]

Can we agree that the only thing that matters is the result and what people will buy and play? Tessellation, Async or not. Its all secondary.

Of course, but we're theorising future performance based on our collective knowledge of GPUs and API features.

Theory is not practice, but informed theory is often reality. Think Einstein and gravitational waves.

People are hyped up about Pascal, I was hyped up about Maxwell (never cared for fiji because I knew before hand that it would simply be a full GCN product which would entail a weak front end due to 28nm limitations and AMD already struggling on 28nm power usage with Hawaii).

This time around, I'm hyped about Polaris moreso than Pascal.

 

[PhonakV30]

I see you really like wccftech.

So Fable is DX12. And how is the GPU performance there?

http://techreport.com/review/29090/fable-legends-directx-12-performance-revealed

Or is there too much tessellation there?

You mean Yesterday Fable wasn't DX12?

 

[Silverforce11]

Re: Fable Benchmark

Devs told journalists there's ~5% Async Compute usage for the benchmark.

We don't know how the final game shapes out yet, but they were talking about using Async Compute for spell effects and and shadows (more than just global illum in the benchmark, which was a scenery run through) in the demo of actual gameplay.

For an Unreal Engine (Engine sponsored by NV, with PhysX integration) game where AMD normally performs 25-50% worse than the relative NV GPU, the performance is so far, outstanding.

It bodes well for AMD in DX12 if they perform so good in an engine that is normally so crippling.

Compare the above with ANY game made with Unreal Engine 4 and you will be astounded at how much AMD performance has jumped up.

We'll see how much Async Compute Hitman uses, but basically there is potential for a performance speed up for GPUs that support this feature.

@sontin
The DX12 AC test, I hope you understand now, will run perfectly fine on any DX12 hardware. The analysis is to see a performance increase, just a proof of concept that some compute tasks are running in parallel, thus the GPU increases in performance. How much performance increase depends on the load.

 

[xthetenth]

I still wait for any proof that Microsoft has designed a way to get this information about multi engine support from the driver.

Just as a heads up, querying for support is one of the fundamental parts of Windows interacting with hardware. DX queries, SCVMM queries, so on and so forth through the entire Windows world. If it's a component designed to interact with hardware that has support for some but not necessarily all of a spec, it's getting queried. So as a casual example, I point to the entire Windows enterprise storage space and the existence of Azure as some of the many proofs that MS has developed the esoteric technology of querying hardware. MS gives you a list of what you need to support to be compliant with a spec, and you implement it. Among other things are methods that expose what relevant functionality you support. And that's how we get support for the wonderful variety of stuff Windows is compatible with.

 

[airfathaaaaa]

https://developer.nvidia.com/dx12-dos-and-donts

this is the closest of what we will ever get from nvidia about async compute

its funny that people still thinks that they support it while even they said they cant

 

[sontin]

Umm, Microsoft tools query the driver in order to determine feature support.

All GPU tools query the driver, see GPU Caps Viewer:
http://www.ozone3d.net/gpu_caps_viewer/

Oxide did the same, the nvidia driver exposed the feature, Async Compute, as being available but once they attempted to use it... Disaster.


Nvidia wanted Oxide to "shut it down" meaning remove all Async Compute from the game. Oxide refused and came to an agreement on creating a vendor ID specific path.

Again, drivers expose features and tools which look for feature support query the driver.

Now let's ask nvidia:
https://developer.nvidia.com/nvemulate

" By default, the NVIDIA OpenGL driver will expose all features supported by the NVIDIA GPU installed in your system."

Or Microsoft:
https://msdn.microsoft.com/en-us/library/jj134360(v=vs.85).aspx

" DirectX 9 drivers (which expose any of the D3DDEVCAPS2_* caps)"

Drivers expose features.

You havent answered the question. What is the cap which describes the multi engine support of the underlaying hardware.

Seriously, what you just wrote is nothing else than:
nVidia has provided a DX12 driver so they are supporting everything AMD supports. :\
I still wait to see how a "DX12 driver" can or cannot support Async Compute.

 

[Glo.]

You havent answered the question. What is the cap which describes the multi engine support of the underlaying hardware.

Seriously, what you just wrote is nothing else than:
nVidia has provided a DX12 driver so they are supporting everything AMD supports. :\
I still wait to see how a "DX12 driver" can or cannot support Async Compute.

It is not the driver. It is the hardware. You have to factor Context switching between Compute and Graphics here. Driver from Nvidia exposed it for the application, but on hardware - well there wasn't anything that was able to switch context between Compute and Graphics, therefore there was no asynchronous scheduling. It is simple as it can be. Thats why you have to add specific code for Nvidia hardware into the application, that is not using Asynchronous Compute.