Kepler, Maxwell, Pascal - Performance comparison from new drivers (353 -> 376)

[AtenRa]

I was fully expecting all the Kepler slides to look like this:

That is exactly what it looks like today if each line represents the performance in new games each year relatively to Hawaii.

First line from bottom is a 2012 game, second line is a 2015 game and then third line is a 2016 Game.

When we say we have a performance regression from the drivers we mean the drivers doesnt make Kepler to perform at 100% in new games because they are not optimized for the Kepler architecture. We dont mean they regress the performance from driver to driver in the same game.

 

[tviceman]

When we say we have a performance regression from the drivers we mean the drivers doesnt make Kepler to perform at 100% in new games because they are not optimized for the Kepler architecture. We dont mean they regress the performance from driver to driver in the same game.

http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...om-new-drivers-353-376.2497059/#post-38683694

 

[beginner99]

How many more such games do we need to see before people accept low level APIs are a failure?

It will take time till we get the full benefit of this but as long as games and engines support both dx11 and dx12 the whole architecture of the engine and game must take dx11 shortcomings into account. Is windows 7 still this big that no game yet is dx12 only?

 

[Ancalagon44]

Huh? Are looking at the graphs? Ashes, Tomb Raider and Deus Ex DX12 are all slower across the board compared to DX11.

Likewise, Doom Vulkan is slower across the board than OpenGL.

How many more such games do we need to see before people accept low level APIs are a failure?

You could see it another way.

Nvidia's driver development team is way, way ahead of AMD's driver development team.

Nvidia's drivers are so good that, when they let the game developers write DX12 code to interact with the GPU, they don't do as good a job as the Nvidia driver developers themselves. With AMD, it is the other way around - game developers do a better job at getting performance out of AMD GPUs than even AMD can.

 

[zlatan]

Nvidia's drivers are so good that, when they let the game developers write DX12 code to interact with the GPU, they don't do as good a job as the Nvidia driver developers themselves. With AMD, it is the other way around - game developers do a better job at getting performance out of AMD GPUs than even AMD can.

Nope, the problem is different. NV built some fast path in the architecture to speed up some relatively useful and mostly used workloads in D3D11. But D3D12 don't allow them to use these fast paths, and this is a huge problem, because the normal paths are relatively weak on their GPUs.
It is not really hard now to outperform D3D11 with D3D12, but only possible with well designed hardwares.

A lot of GPU architectures designed differently. AMD don't really believes in fast paths, so they do a more robust and flexible design, which is the primary reason why they can get better performance in D3D12 compared to D3D11. This is a disadvantage in static binding, constant buffers, etc, but will be an advantage with the bindless binding in the future games.
For now GCN get some advantage in todays games by accessing the structured and constant buffer with the same performance. For the other hand NV access the structured buffer way slower than the constant buffer, and there is nothing we can do to help the hardware.

So there is nothing wrong with the APIs and the vendor implementations, the main difference is the architecture design.

 

[Innokentij]

So there is nothing wrong with the APIs and the vendor implementations, the main difference is the architecture design.

Say what now, not what all the DX12 games been released so far shows, concidering how many patches they need before performance DX12vsDX11 regression is gone. I totaly buy the fact that nvidia always been better at drivers for DX11 then AMD with huge overhead lower performance drivers.

 

[Carfax83]

Nope, the problem is different. NV built some fast path in the architecture to speed up some relatively useful and mostly used workloads in D3D11. But D3D12 don't allow them to use these fast paths, and this is a huge problem, because the normal paths are relatively weak on their GPUs.

This really doesn't make much sense at all to me. DX12 is a low level API, it's supposed to expose more of the GPU, not less. Anything available in DX11 should be available in DX12, and then some. Plus, there are games where NVidia definitely gains in DX12, like Ashes of the Singularity, a very highly optimized DX12 title. So to me it seems like it's much more a matter of optimization than anything else. It took MONTHS before Oxide studios could patch the game and bring NVidia's DX12 performance up to par as the game was heavily AMD biased from it's inception (plus NVidia definitely tweaked their DX12 drivers), but now NVidia is very competitive in that game.

It is not really hard now to outperform D3D11 with D3D12, but only possible with well designed hardwares.

You know there are times when I really want to believe that you are an unbiased source, but when you make comments like these, it's hard for me to drink the Kool-Aid. The insinuation being of course, that NVidia's architecture is inherently DX12 unfriendly. No matter how many times we see NVidia outperform AMD in DX12 titles, this myth still persists and it seems to be driven by the same people.

 

[Arachnotronic]

You know there are times when I really want to believe that you are an unbiased source, but when you make comments like these, it's hard for me to drink the Kool-Aid. The insinuation being of course, that NVidia's architecture is inherently DX12 unfriendly. No matter how many times we see NVidia outperform AMD in DX12 titles, this myth still persists and it seems to be driven by the same people.

Yeah, the meme is getting old.

 

[Bacon1]

It took MONTHS before Oxide studios could patch the game and bring NVidia's DX12 performance up to par as the game was heavily AMD biased from it's inception (plus NVidia definitely tweaked their DX12 drivers), but now NVidia is very competitive in that game.

Why do people keep saying that AOTS is AMD biased? The developers have stated they worked closer with Nvidia than AMD. It was Nvidia's drivers not being up to date that was the issue. Just look at the OP. 7-12% improvement from DX12 Drivers. Same reason that Vulkan in Doom was slower at first. Nvidia's drivers weren't good. They fixed that for a 12% gain once again.

The difference is once the initial DX12 / Vulkan drivers get worked out, there won't be the need for game specific drivers anymore to the extent there was for DX11, which is basically every game rendering improperly w/o drivers.

 

[MajinCry]

Zlatan is a proper dev you eejits. Might want to listen to the guy that works with this stuff on a day to day basis, rather than going with what feels good in light of your hardware purchases.

If said member is who or what you are claiming them to be, then they are also perfectly capable of addressing directly any concerns or comments regarding their posts here. They don't get any special treatment over other forum participants, and don't need you derailing the thread to white knight for them.
-- stahlhart

 

[nathanddrews]

What we have here is a failure to communicate.
https://youtu.be/MdSH0l9NMek

 

[Despoiler]

Am I missing something here? 353.63 was released July 2015. 368.39 June 2016. The latest drivers are from Dec 2016. In terms of the data and the way the article presents it, Doom took Nvidia a year and half to support Vulkan and it's still behind OGL. Hitman a full year for DX12 support. AoTs DX12 performance regressed in the first year, but now is the highest. DX11 performance regressed and the latest drivers are the slowest. Star Wars Battlefront. Crash on oldest driver. It's exactly what Nvidia owners and non-owners have noticed. Not only that, the data doesn't take into account the time it took for any quality of experience issues to be fixed.

 

[Ansau]

And it's so silly. Instead of praising a particular company for having the position and foresight to win big contracts (like the PS4 and XB1), people instead want to seek out negativity because it fits their narrative that one company is evil and the other is good.

Look at AoTS, Shadow Warrior 2, Quantum Break Steam version (which performs better on ALL GPUs, not just Nvidia), No Man's Sky, ARK, XCOM 2, Civ 6, Rise of the Tomb Raider (which was XB1 exclusive for 10-11 months before PS4), Vanishing of Ethan Carter, SOMA, and EVERY.SINGLE.PC.EXCLUSIVE.MMO. Those are all games that were only released on 1 console (except ROTTR which still had a 11 month exclusive XB1) or were PC only and the 780 TI either ties or beats the R9 290X.

AMD had great foresight to win the console contracts and scored a headwind of performance upticks for games that are released on all 3 platforms because of it. Kepler AND Maxwell suffered more because of this than anything else.

Sorry, but that's not true. Checking from here: https://www.computerbase.de/thema/grafikkarte/rangliste/
290(non X) vs 780ti.
AoTS: 44.1 vs 42.5.
Rise Tomb Raider: 43.6 vs 37.6.
Vanishing Ethan Carter: 47.0 vs 43.3.
Talos Principle: 52.6 vs 45.4.
XCOM2: 46.0 vs 30.0.
And the average of all games is 50.0 vs 46.9. That's Nvidia aging badly, where the cut down version of Hawaii ends up beating the full GK110 by a 6.6%.

And the problem with Nvidia has nothing to do about consoles or AMD marketing. It's their politic to only make 1 or 2 iterations from every major architecture, and from there designing and optimizing games only with the latest in mind. This approach induces their products to have a rather short lifespan. Kepler users had the bad luck they bought gpus sentenced to die sooner that later.
AMD takes a more conservative path by having a base design they improve it through several iterations, same with their cpu lines like Bulldozer and now Zen, and like Intel is doing with their cpus since Sandy Bridge. This approach allows products to be relevant and get support for more time.

That is the true meaning of the lack of support of Nvidia on older architectures, not that newer drivers yield worse performance. But of course Internet if full of uneducated people that misunderstand most of the things they read.

 

[raghu78]

You know there are times when I really want to believe that you are an unbiased source, but when you make comments like these, it's hard for me to drink the Kool-Aid. The insinuation being of course, that NVidia's architecture is inherently DX12 unfriendly. No matter how many times we see NVidia outperform AMD in DX12 titles, this myth still persists and it seems to be driven by the same people.

Of the roughly dozen odd DX12 games AMD wins 75% atleast if not more. Rx 480 is on avg atleast 5% (if not more )faster in DX12 vs GTX 1060.

https://en.wikipedia.org/wiki/List_of_games_with_DirectX_12_support

Whats even more revealing is Nvidia's DX11 performance is on par or superior in every DX12 title. You could say better DX11 drivers but thats exactly why Nvidia does better as they have a lot more driver software engineers. Raja Koduri stated in the recent PC World interview that the DX11 drivers are massive and have actual graphics shader code optimized for various game engines / games. In contrast the DX12 driver is lightweight and gives a lot more control to the developer.

 

[crisium]

Odd that one or two see this as complete exoneration of Kepler since fall 2014.

As I understand it, the issue is that users who bought Nvidia cards from 2012 to early September 2014 did not receive as much return on investment as users who bought AMD cards in the same time-frame and held onto them for the coming months and years.

What I see here is the dismissal of the conspiratorial belief that Nvidia actively sabotages Kepler. To dismiss the true issue in the above paragraph we'd need broader comparisons to see if they have improved back to pre-Maxwell performances tiers relative to GCN 1 and 2 cards. Furthermore, if they did somehow reset the average back 2.5 years to what it was, does that erase the past 2+ years?

Anyway, it's great to see performance bumps across the board in drivers. Both companies have now done this, so I very much hope the lazy, bench it once and use it for years, reviewers consider complete rebenches in 2017 for all older games. Spoiler alert: many of them won't.

 

[littleg]

Will we still see the same kinds of improvements through drivers in the DX12 era though? It seems to me that there's less AMD and Nvidia can actually do on the driver side to improve performance compared to DX11.

 

[crisium]

I see examples of DX12 gaining in the OP, looks comparable to the DX11 gains for games that have both.

 

[Phynaz]

But how can this be? Nvidia is nerfing performance to get people to buy Pascal!!

 

[96Firebird]

That is the true meaning of the lack of support of Nvidia on older architectures, not that newer drivers yield worse performance. But of course Internet if full of uneducated people that misunderstand most of the things they read.

Except the bolded is exactly what some people on these forums thought, and probably still think. But I guess your second sentence takes care of them.

 

[2is]

Nice to see some definitive evidence that Kepler wasn't intentionally gimped via driver updates. I still have one of the two 680's I was running. Still has acceptable 1080p performance as long as I stay within it's 2gb limit. Difference is today's games can easily exceed 2gb vram even at 1080p at higher settings even without AA.

 

[WhoBeDaPlaya]

[Who] actually thought that nVIDIA was intentionally sabotaging Kepler?!
Piss poor support, yes - hence my 2x 780 -> 2x 290 @ 290X swap a long while back.

Watch your language in this forum.
-- stahlhart

 

[Keysplayr]

[Who] actually thought that nVIDIA was intentionally sabotaging Kepler?!
Piss poor support, yes - hence my 2x 780 -> 2x 290 @ 290X swap a long while back.

Watch your language in this forum.
-- stahlhart

First day at Anandtech forums?
Seriously though are you kidding me?
One of the biggest AMD faithful marketing pushes these forums have seen.
To paraphrase, "Dont by Maxwell as Nvidia will gimp it just like they did Kepler to boost sales their next gen."
Plastered whenever and wherever possible and even a few places it wasnt possible, all over this forum.

 

[antihelten]

First day at Anandtech forums?
Seriously though are you kidding me?
One of the biggest AMD faithful marketing pushes these forums have seen.
To paraphrase, "Dont by Maxwell as Nvidia will gimp it just like they did Kepler to boost sales their next gen."
Plastered whenever and wherever possible and even a few places it wasnt possible, all over this forum.

But as other posters have already pointed out "gimping" doesn't necessarily refer to reducing performance of older architectures through drivers as much as it simply refers to abandoning any improvements to said architectures through drivers (in favor of focusing on the newest architecture).

But by all means if people were specifically claiming that Nvidia was reducing Kepler performance through drivers (and not just abandoning improvements), then I'm sure you won't have any issues finding some quotes, especially since this was apparently "One of the biggest AMD faithful marketing pushes these forums have seen."

 

[Arachnotronic]

But as other posters have already pointed out "gimping" doesn't necessarily refer to reducing performance of older architectures through drivers as much as it simply refers to abandoning any improvements to said architectures through drivers (in favor of focusing on the newest architecture).

But by all means if people were specifically claiming that Nvidia was reducing Kepler performance through drivers (and not just abandoning improvements), then I'm sure you won't have any issues finding some quotes, especially since this was apparently "One of the biggest AMD faithful marketing pushes these forums have seen."

Did it ever cross people's minds that there's just not that much more performance to wring out of ol' Kepler?

 

[96Firebird]

Yet the article that sparked this thread shows that 780 Ti gained more performance than 980 with the Win10 drivers, even after the 980 was released. Abandoning not found.

I'm actually surprised by this, since I figured they would have been able to refine a newer architecture more so than an older one.