[bitsandchips]: Pascal to not have improved Async Compute over Maxwell

[Despoiler]

https://forum.beyond3d.com/posts/1915300

I was actually wondering if this was the case. So NV launched a card without requesting changes to a game/benchmark that would showcase their new functionality. That is kind of a major blunder. While future reviews could change the results, people remember the launch reviews. For instance AMD was never able to shake the hot and loud reference cooler fiasco of the 290 series within the lifetime of that series.

 

[krumme]

Sweclockers dx12 bench suite is a nice overview
http://www.sweclockers.com/test/22087-nvidia-geforce-gtx-1080/20#content

 

[airfathaaaaa]

I was actually wondering if this was the case. So NV launched a card without requesting changes to a game/benchmark that would showcase their new functionality. That is kind of a major blunder. While future reviews could change the results, people remember the launch reviews. For instance AMD was never able to shake the hot and loud reference cooler fiasco of the 290 series within the lifetime of that series.

yeah but if the game forces async off on every nvidia card how come they did the test?

 

[sontin]

Pascal neither gains or loses performance with Async compute + graphics turned on.

Pascal is (slighty) faster with Async Compute. Sites like Computerbase are using multiple runs.

Fences synchronize work from various queues. They're needed for hardware which executes work from various contexts in a serial fashion. They're not required for GCN due to the ACEs being capable of handling synchronization on their own. Meaning that the ACEs can wait for work from a graphics task without needing a fence in place. GCNs hardware scheduler has this added flexibility. Case in point, GCN doesn't make use of fences under Vulkan.

Seriously, what? The hardware has no clue how to synchronize these queues.

I made a mistake in that first paragraph you quoted. Replace preemption with context switch. If you execute two different contexts one after the other in serial, a context switch is involved. Switching between contexts too often is one of NVIDIA's don't in their DX12 do's and don'ts.

Switching a context within a graphics queue is not Async Compute.

 

[antihelten]

yeah but if the game forces async off on every nvidia card how come they did the test?

Apparently the game has async set to off by default if it detects an Nvidia card, but it can still be forced on by messing with the .ini settings.

Also Ryan Smith (who has been talking with Dan Baker) had this to say on B3D:

There's still some work re-org that goes on when async is turned on in Ashes. Apparently this is benefiting the Pascal cards ever so slightly.

 

[airfathaaaaa]

Pascal is (slighty) faster with Async Compute. Sites like Computerbase are using multiple runs.



Seriously, what? The hardware has no clue how to synchronize these queues.



Switching a context within a graphics queue is not Async Compute.

he didnt say that switching a context is async
he says that switching the way maxwell was doing was coming with a very high cost
now on maxwell v3.0 that cost is gone with the pre emption but the card is still not doing it in paraller but in serial hence the brute forcing through high clocks we keep saying eventually after some point the card will be fed up with compute and graphics that will start to stall hence no2 the regression we see on 4k

 

[sontin]

There is no context switch involved when there is a dedicated graphics and a dedicated compute queue.

• Don’t toggle between compute and graphics on the same command queue more than absolutely necessary

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

 

[LTC8K6]

There is no context switch involved when there is a dedicated graphics and a dedicated compute queue.


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

That is pre-Pascal though.

 

[airfathaaaaa]

There is no context switch involved when there is a dedicated graphics and a dedicated compute queue.


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

yeah in maxwell v2.0...
in maxwell v3.0 they do have hence the no regression up untill 4k

 

[Good_fella]

60.000 for a space rts is damn good number
http://steamspy.com/app/228880
also


http://www.eteknix.com/nvidia-confirms-pascal-natively-supports-asynchronous-compute/

they are still downplaying it despite the fact that they do have"async support"

the problem is that on dx12 nvidia is only what 8-10% ahead of amd while bruteforcing

LOL overhyped space RTS (or should I say graphic/performance marvel) sold 60k since release. Another space RTS Starcraft 2 sold 1 million in 24 hours.

Nope. It's enabled in XBONE version, but disabled in DX12 PC version.. you know, nVidia

Is this part of AMD's master plan to consolize PC gaming?

Another one: "The Windows 10 version of Quantum Break uses the same reconstruction method as on Xbox One."

 

[Mopetar]

LOL overhyped space RTS (or should I say graphic/performance marvel) sold 60k since release. Another space RTS Starcraft 2 sold 1 million in 24 hours.

You're comparing a new IP from a little known company to a sequel to one of the most massively popular RTS franchises in history from a well known publisher.

Normally we wouldn't spend much time talking about it, but it is one of the first DX12 titles which is what makes it noteworthy.

 

[sirmo]

You're comparing a new IP from a little known company to a sequel to one of the most massively popular RTS franchises in history from a well known publisher.

Normally we wouldn't spend much time talking about it, but it is one of the first DX12 titles which is what makes it noteworthy.

Exactly. It showcases the use of compute shaders with new APIs. It's a sign of things to come, which makes it a really interesting bench imo.

 

[BlitzWulf]

Can anyone confirm that this snippet from PCGames.de Total war:Warhammer review was translated correctly?


http://www.pcgames.de/Total-War-War...on-Creative-Assemblys-Strategiespiel-1195328/

A patch to a release is to fix these problems, also Creative Assembly then provides a mode for Direct X 12 after. The works, however, only with AMD graphics cards. The low-level interface Direct X 12 could bring significant advantages over Direct X 11 just in terms of CPU load. More's around

original text:

Ein Patch zu Release soll diese Probleme beheben, zudem liefert Creative Assembly dann einen Modus für Direct X 12 nach. Der funktioniert allerdings ausschließlich mit AMD-Grafikkarten. Die Low-Level-Schnittstelle Direct X 12 könnte gerade in Sachen CPU-Last deutliche Vorteile gegenüber Direct X 11 bringen. Mehr dazu gibt's rund um den Release

If this is true I wonder if DX12 mode is AMD exclusive because Nvidia cards can't run it or because AMD said so?
It doesnt really sound like something AMD would do but having GE games have DX12 mode exclusive to AMD would be a hell of a comeback to gameworks.
I'll take DX12 over HTFS any day of the week.


Edit: PCgames.de not PCGH

 

[Noctifer616]

Can anyone confirm that this snippet from PCGH's Total war:Warhammer review was translated correctly?


http://www.pcgames.de/Total-War-War...on-Creative-Assemblys-Strategiespiel-1195328/




original text:





If this is true I wonder if DX12 mode is AMD exclusive because Nvidia cards can't run it or because AMD said so?
It doesnt really sound like something AMD would do but having GE games have DX12 mode exclusive to AMD would be a hell of a comeback to gameworks.
I'll take DX12 over HTFS any day of the week.

The google translation isn't good. The part that says "zudem liefert Creative Assembly dann einen Modus für Direct X 12 nach" is wrong in the sense that "nach" is not meant as "after". "liefert nach" means to deliver.

From what I gather the article says that on release there will be a patch that fixes some issues and also adds DX12. And yes, it does say that it only works with AMD, which doesn't make sense.

Mind you, my German is quite rusty, so I could be misunderstanding it, however it does sound like that the DX12 is bundled with a A-sync fix patch on release.

EDIT: Yes, "nach" isn't a word used on it's own, just like "for" in English. So the translation to "after" doesn't take context into consideration.

 

[AtenRa]

Can anyone confirm that this snippet from PCGames.de Total war:Warhammer review was translated correctly?


http://www.pcgames.de/Total-War-War...on-Creative-Assemblys-Strategiespiel-1195328/




original text:





If this is true I wonder if DX12 mode is AMD exclusive because Nvidia cards can't run it or because AMD said so?
It doesnt really sound like something AMD would do but having GE games have DX12 mode exclusive to AMD would be a hell of a comeback to gameworks.
I'll take DX12 over HTFS any day of the week.


Edit: PCgames.de not PCGH

Here is the official blog

http://wiki.totalwar.com/w/Optimisation_Blog

We’re pleased to confirm that Total War: WARHAMMER will also be DX12 compatible, and our graphics team has been working in close concert with AMD’s engineers on the implementation. This will be patched in a little after the game launches, but we’re really happy with the DX12 performance we’re seeing so far, so watch this space!

and

In GPU terms, we’ve shifted our particle simulation pipeline from the pixel shader to the compute shader, which is a more efficient use of the GPU’s time. In fact we’ve done this with several parts of the rendering pipeline, further utilizing the GPU and letting the CPU focus on everything else it has to do.

Kepler will have a hard time in this one.

 

[airfathaaaaa]

You're comparing a new IP from a little known company to a sequel to one of the most massively popular RTS franchises in history from a well known publisher.

Normally we wouldn't spend much time talking about it, but it is one of the first DX12 titles which is what makes it noteworthy.

i wouldnt go and say they are little given that they gave us soase to people that knows about rts they are pretty well known
problem is (as we all know) they are showcasing what can be done when dx12 can be fully utilised and certain key areas of it dont really work well on nvidia

 

[thesmokingman]

i wouldnt go and say they are little given that they gave us soase to people that knows about rts they are pretty well known
problem is (as we all know) they are showcasing what can be done when dx12 can be fully utilised and certain key areas of it dont really work well on nvidia

This, so much this. They are one of the big dogs in rts. This engine will underpin remakes of their better known franchises. They even wrote about needing a next gen level engine to achieve their corporate goals.

 

[Despoiler]

So how is the app gaining 15-20% performance if not by concurrently running tasks? Moreover NVIDIA showed diagrams in the same session with the press to
Illustrate concurrent execution of graphics and compute. The demo is just a follow up to that. The whole video is on YouTube.

Regarding perf improvements imagine AMD new architecture is much better at filling their cores with gfx tasks (since parts like Fury X demonstrate they are not very great or at that on large configurations). This means the opportunities for concurrently scheduling compute and gfx workloads are reduced and therefore you might see a lower perf improvement from async compute on their newer GPUs! If that happens are you suddenly going to say their implementation suck or will you congratulate them with fixing a major issue of their architecture?

Nvidia confirms they have to do async, aka dynamic load balancing, in software in the drivers. So no they did not fix their architecture. Then they demo the functionality in DX11 not DX12. It's odd. DX11 and async don't have any relevance together. What in the actual eff? 1080 benchmarks in AoTS that show async on/off, Pascal is not regressing in performance compared to DX11 as we saw in Maxwell. Real world gain in the relevant API is 5%. Generally speaking, no one congratulates anyone when they only match the previous benchmark performance. It's the lowest bar there is.

 

[sontin]

Oh look, Async Compute on a GTX1080:

Software was the Microsoft n-Body demo.

 

[Silverforce11]

Oh look, Async Compute on a GTX1080:

Software was the Microsoft n-Body demo.

Is that why it gains almost no performance in that very benchmark?

Or in Async On/Off Ashes of the Singularity? <- Still the only game that allows testing AC on/off to compare.

 

[guskline]

I will be receiving my GTX 1080 FE on Thursday and by the weekend I should have the waterblock on and have it installed on the 5960x rig. I'll run my licensed version of AOTS and report my findings. I intend on moving my GTX 980 TI SC with the EK block to my 4790k rig so I'll report those findings also.

 

[sontin]

Is that why it gains almost no performance in that very benchmark?

Or in Async On/Off Ashes of the Singularity? <- Still the only game that allows testing AC on/off to compare.

Having the ability to do proper Async Compute doesnt say anything about performance improvements.

Ashes is tailored towards AMD. So why would anyone expect the same or any gain on nVidia hardware?

 

[Abwx]

Having the ability to do proper Async Compute doesnt say anything about performance improvements.

Ashes is tailored towards AMD. So why would anyone expect the same or any gain on nVidia hardware?

It s tailored toward async compute, either your hardware support it and there s some gain or it doesnt support it and at best there s no losses...

I guess that you didnt get that async compute is about gaining something...

 

[sirmo]

Having the ability to do proper Async Compute doesnt say anything about performance improvements.

Ashes is tailored towards AMD. So why would anyone expect the same or any gain on nVidia hardware?

It is not. Repeating it over and over again does not make it true. Oxide developers shared the sources of the game with both Nvidia and AMD. Read their blog about this issue.

 

[ThatBuzzkiller]

Having the ability to do proper Async Compute doesnt say anything about performance improvements.

Ashes is tailored towards AMD. So why would anyone expect the same or any gain on nVidia hardware?

Well for one thing the gains of async compute are dependent on how geometry limited you are ...

If the Xbox One was able to see moderate gains with async compute with a ratio of 768 ALU ops per 1 rasterized triangle then a 1080 should see similar gains like a 290 would since they both have a ratio of 1280 ALU ops per 1 rasterized triangle ...

The gains with async compute on a Fury X would be even better since it has an even more skewed compute to graphics ratio capabilities ...

The fact that the 1080 is having trouble seeing definitive gains means that Nvidia still doesn't have proper hardware support for async compute ...