guru3dDoom Vulkan Benchmarks

[Bacon1]

NV boost probably coming soon:

Oxide said the exact same thing almost a year ago. Still waiting for Maxwell Async Compute to be enabled by drivers...

Hopefully they do manage to get async compute working, because it is free gains for everyone and that means more developers will spend time implementing them, but I'm not holding my breath waiting.

 

[LTC8K6]

Oxide said the exact same thing almost a year ago. Still waiting for Maxwell Async Compute to be enabled by drivers...

Hopefully they do manage to get async compute working, because it is free gains for everyone and that means more developers will spend time implementing them, but I'm not holding my breath waiting.

Apparently happened a while ago:

http://www.gamespot.com/forums/pc-a...ly-enables-async-with-march-drivers-33047280/

IIRC, we have seem some boost for Maxwell in DX12 titles.

 

[Snarf Snarf]

The Fury X gains at 1080 just blew my mind. 66% gains is edging on a classic node shrink and this is all just from switching API. Where is AMD's marketing division right now? I'd be blasting this all over the internet right now "NVidia no gains in next gen API's while GCN has free performance left under the hood"

 

[Bacon1]

Apparently happened a while ago:

http://www.gamespot.com/forums/pc-a...ly-enables-async-with-march-drivers-33047280/

IIRC, we have seem some boost for Maxwell in DX12 titles.

Nope, its still not enabled:

http://twitter.com/PellyNV/status/702556025816125440

 

[Racan]

The Fury X gains at 1080 just blew my mind. 66% gains is edging on a classic node shrink and this is all just from switching API. Where is AMD's marketing division right now? I'd be blasting this all over the internet right now "NVidia no gains in next gen API's while GCN has free performance left under the hood"

8.6 TFlops vs the GTX 1070 with 6.5 TFlops, finally thanks to Vulkan that theoretical advantage can be unleashed.

 

[DisarmedDespot]

And the master plan unfolds! Seriously, if this is a sign of things to come, AMD will be crushing it in new games. Hopefully Nvidia can catch up with a patch or something.

 

[HurleyBird]

The Fury X gains at 1080 just blew my mind. 66% gains is edging on a classic node shrink and this is all just from switching API.

Fiji has a very difficult time utilising its massive shader power under normal circumstances, but graphics APIs like DX12, Vulcan, and Mantle can correct things if implemented properly.

Depending on your perspective, you could either say the Fiji is a completely imbalanced, or a very forward looking architecture.

Unfortunately for the later, the small 4GB frame buffer puts an effective limit on how future proof the Fury line can be. While the architecture may be incredibly forward looking, its product implementation is not.

That's the real tragedy of Fiji. When it shipped neither its current nor its potential future performance were compelling. If it had a larger frame buffer, I would have recommended it over the 980ti every time. When the first benchmarks of the FuryX came out, I predicted that it would initially age better than the 980ti due to its more forward looking architecture, but also that it would eventually fall behind again because of the smaller frame buffer. The first part of that prediction is now coming true.

 

[Irenicus]

Apparently happened a while ago:

http://www.gamespot.com/forums/pc-a...ly-enables-async-with-march-drivers-33047280/

IIRC, we have seem some boost for Maxwell in DX12 titles.

I don't know if this is true, but I think what might have happened is that the automatic workloads the additional scheduling hardware operates on (with gcn) is a NIGHTMARE to optimize for manually, and since nvidia chose not to add the same sorts of hardware schedulers to cut their power use down (that extra power use is NOT for nothing on gcn as we see here when it actually gets used) they have to task scores of engineers to work on the problem for untold man hours (which they have the resources to do).

Eventually, they will likely boost the performance up higher, but even if they do this will completely be a role reversal compared to previous launch benchmarks. Doom did not launch with both opengl and vulkan, but for future games that do launch with both code paths for vulkan/dx12, amd will have a head start because their hardware is completing the optimization work automatically, and so it will be nvidia that is delayed in terms of optimized performance, if they ever get there at all.

 

[Head1985]

https://www.computerbase.de/2016-07...md-nvidia/#diagramm-doom-mit-vulkan-2560-1440

Gigantic boosts for AMD, nothing for nVidia.

yeah AMD drivers suck they need low level API to use most of their hardware.
+66% on furyx performance...
WTF they are doing with normal drivers?

 

[Stuka87]

yeah AMD drivers suck they need low level API to use most of their hardware.
+66% on furyx performance...
WTF they are doing with normal drivers?

Not wasting resources doing a complete overhaul of their drivers for an API that will be going away.

They knew they needed to overhaul them, so they created Mantle. Which then became Vulcan and DX12.

 

[AnandThenMan]

yeah AMD drivers suck they need low level API to use most of their hardware.
+66% on furyx performance...
WTF they are doing with normal drivers?

What is Nvidia doing with their Vulkan drivers? Fury X should not be beating the 1070 by 20%.

 

[Snarf Snarf]

yeah AMD drivers suck they need low level API to use most of their hardware.
+66% on furyx performance...
WTF they are doing with normal drivers?

I think that's the real question here, is the DX11 overhead for Fury that high or is it a combination of the the terrible front end that doesn't keep the shaders fed and the overhead.

66% is HUGE, that's a generational jump plus a tier or performance. There were some people during the Fury hype around here claiming up to 25% faster than Titan X, if they would have been able to keep Fury X fed it might have actually lived up to the hype

 

[Yakk]

Eventually, they will likely boost the performance up higher, but even if they do this will completely be a role reversal compared to previous launch benchmarks. Doom did not launch with both opengl and vulkan, but for future games that do launch with both code paths for vulkan/dx12, amd will have a head start because their hardware is completing the optimization work automatically, and so it will be nvidia that is delayed in terms of optimized performance, if they ever get there at all.

In reality all the heavy work for the new APIs is being finished right now and we are just beginning see the fruits of this labour for PCs.

The bulk of the API optimizations for AMD are being done at the console game engine level for big studio developers, with specific PC optimizations for different configurations. This stands to say, as DX12 & Vulkan continue to be more and more widespread it should cost less for AMD and be faster to optimize thier PC games, with Async actually optimizing a lot of inefficiencies in the code quickly saturating the GPUs with work. Whereas nvidia will still have to spend more time and money to manually optimize for the new APIs compared to DX11 and under.

So gameworks may be nvidia's answer with more proprietary code, and it will work for PC games sponsored by nvidia, but still for games with console ports the heavy work will already be done for AMD in part on either APIs.

 

[Arachnotronic]

What is Nvidia doing with their Vulkan drivers? Fury X should not be beating the 1070 by 20%.

If the Vulkan codepath is designed to be very compute heavy and is optimized to wring the most out of the GCN architecture, then I'm not surprised -- 6.5TFLOPs on the GTX 1070 versus ~8.6 TFLOPs on the Fury X.

 

[seitur]

Well AMD cards since GCN and Keplers have more brute force in their hardware so it makes sense they see higher growths.

 

[LTC8K6]

Nope, its still not enabled:

http://twitter.com/PellyNV/status/702556025816125440

I read that as enabled in the driver, but not the individual app/game. Without both, it's essentially not enabled, such as with Doom?

 

[Arachnotronic]

I read that as enabled in the driver, but not the individual app/game. Without both, it's essentially not enabled, such as with Doom?

With DX12/Vulkan, the developer really needs to explicitly support the different architectures that they're targeting. Asynchronous Compute isn't just a "switch" that the developer can just turn on and have work for all GPUs; it needs to be done essentially on a per-architecture basis.

It's clear that id worked pretty closely with AMD to build a code path that takes good advantage of GCN. It looks like id is still working with NVIDIA to build a code path for the Pascal GPUs that can take advantage of Pascal's asynchronous compute abilities/architecture.

 

[LTC8K6]

With DX12/Vulkan, the developer really needs to explicitly support the different architectures that they're targeting. Asynchronous Compute isn't just a "switch" that the developer can just turn on and have work for all GPUs; it needs to be done essentially on a per-architecture basis.

It's clear that id worked pretty closely with AMD to build a code path that takes good advantage of GCN. It looks like id is still working with NVIDIA to build a code path for the Pascal GPUs that can take advantage of Pascal's asynchronous compute abilities/architecture.

I'm sure there will be a patch or update for Doom and NV cards.

Don't forget this slide:

http://static2.gamespot.com/uploads/scale_super/92/929129/3027998-actualasync.jpg

There are undoubtedly improvements in the pipeline.

 

[Face2Face]

Wow, Fury X just wrecking everything. Finally all those shaders can be unleashed.

And I can confirm those numbers. I couldn't believe what I was seeing when I switched API's on it last night

That's awesome man. It's good to see the Fury spread it's wings. I have a 7970 GE in my other machine, and I'm tempted to put DOOM on it to try it out.

 

[RussianSensation]

The trend is developing with a very high percentage performance gain for AMD, but as far as I know there are not thousands of games yet for Vulkan and DX12 APIs.

It doesn't matter for Fury X vs. 980Ti gen as that's history but RX 480 is only 7-8% behind 980Ti. Pascal 1070 shows no material gains at all which means 1060 will have nothing to show for itself against the cheaper RX 480. There are some huge titles coming out soon with DX12.

As I said in the other thread, NV doesn't make GPUs to last. This has been true going back to GeForce 5, 7, and Kepler.

At 1440p, latest benches show:

$300 280X ~ $500-650 780!
$400 290 > $700 780Ti!
$430 390X > $550 980
https://www.techpowerup.com/reviews/EVGA/GTX_1070_SC/24.html

If we had more DX12/Vulkan games, Fury X would have beaten 980Ti.

Unfortunately RX 480 only has 32 ROPs. Still, if there are more Vulkan/DX12 AAA games, this actually could give AMD a fighting chance with Vega.

At this point it probably matters a lot more how a modern GPU performs in TW:Warhammer, DE:Mankind Divided, BF1, etc. than it does in Crysis 3, Dragon Age Inquisition, Metro LL, BF4, etc.

Very impressive how well optimized Doom is compared to many other modern games that look and run worse.

 

[valici]

I think the issue is simple: AMD has a parallel architecture and drivers and NVidia has a serial architecture and drivers. In DX11 AMD has to emulate a serial architecture and in DX12 Nvidia has to emulate a parallel architecture.
Both see a decrease in performance on the API they were not designed for.

Also that explains NVidia's higher frequency. Like having more pipeline stages in a CPU.

 

[LTC8K6]

Why assume this is the end of the performance envelope for pascal, though?

People act as if Pascal cannot get any better.

Doesn't make sense.

Of course Pascal will get better.

AMD will get better, too.

 

[Bacon1]

I read that as enabled in the driver, but not the individual app/game. Without both, it's essentially not enabled, such as with Doom?

No its quite clear, the game has support but their drivers do not:

Async Compute is NOT enabled on the driver-side

With DX12/Vulkan, the developer really needs to explicitly support the different architectures that they're targeting. Asynchronous Compute isn't just a "switch" that the developer can just turn on and have work for all GPUs; it needs to be done essentially on a per-architecture basis.

It's clear that id worked pretty closely with AMD to build a code path that takes good advantage of GCN. It looks like id is still working with NVIDIA to build a code path for the Pascal GPUs that can take advantage of Pascal's asynchronous compute abilities/architecture.

It doesn't need to be done per architecture. I seem to recall Nvidia having id promote Vulkan with the 1080 release how are they working closer with AMD than Nvidia?

We have yet to see any Nvidia card take advantage of Async Compute in any game. There are perf increases from removing driver overhead, but that's all we've seen.

 

[RussianSensation]

Why assume this is the end of the performance envelope for pascal, though?

People act as if Pascal cannot get any better.

Doesn't make sense.

Of course Pascal will get better.

AMD will get better, too.

No one says Pascal cannot get better. It's a matter of magnitude/scaling. RX 480 shows 40-48% increase in performance, and Fury X even more at 52-66%. The more serial architecture - GCN - has Command Processors and ACE engines designed for high throughout. We have seen enough DX12 benches to see that DX11 never allowed these hardware benefits to come into play based on how the code is issued. The extra hardware in GCN was underutilized because the architecture wasn't made for serial APIs.

If you go way back to 2011 when Eric Demers unveiled GCN, he explained all of this. It's just too bad it took 5 years to see that he was 100% correct. DX11 is simply outdated and hopefully by the time PS5/XB2 come out DX12/Vulkan can put DX11/OpenGL out to pasture.

That's not even discussing the immense CPU overhead benefits for strategy games that DX12 already shows.

If NV finally designs a parallel + hardware Async Compute architecture for 2018-2019, the entire gaming industry will benefit as all major modern GPUs and consoles will be on the same page.

Under no circumstances should a Fury X be beating a 1070 by 26% in a modern game. We also already see that NV's architectures have utilization issues by comparing 1070 vs. 1080. The extra shader and TMU performance on the 1080 is just wasted. That means while Pascal is better utilized under DX11 than GCN is, Pascal itself is starting to run into its own bottlenecks.

DX11 API is already creating a CPU bottleneck for 1080 SLI based on Guru3D and TPU benches. This isn't about AMD vs. NV but the fact that DX11 is simply outdated and last gen.

 

[krumme]

No its quite clear, the game has support but their drivers do not:





It doesn't need to be done per architecture. I seem to recall Nvidia having id promote Vulkan with the 1080 release how are they working closer with AMD than Nvidia?

We have yet to see any Nvidia card take advantage of Async Compute in any game. There are perf increases from removing driver overhead, but that's all we've seen.

A synch is absolutely a per architecture feature. Its crazy complex stuff and my guess is the lack of granularity in pascal and lack of consoles support, little market share plus basic unfit arch will be the reason we don't see it in pascal with solid benefit besides from a few marketing cases. Its a Cuda feature. It would be good if some devs like zlatan chimed in here btw.

The only reason async is here is because the devs need to extract so much power of consoles as possible with compute power not available in the cpu. It doesn't come cheap. And there is no incentive to make it for pascal where the benefits would be slimmer.

And remember asynch is just a small part of dx12 mantle Vulcan. There can be many reasons eg the ace why gcn benefits so much.