Nvidia's Performance Under Vulkan API Explored

[Carfax83]

Have you read it? They are using PresentMon (raw framebuffer output), so those results are quite incorrect . There are still lower FPS displayed on monitor :/

Well then I guess all benches for Vulkan are incorrect then, since every reviewer uses PresentMon..

 

[Red Hawk]

Well, on a 750TI works worse, just tell me how that is better.

The fact that Kepler and Maxwell 1 cards currently have worse performance with Doom running in Vulkan doesn't mean there is no potential benefit for them with low level APIs. Or are you suggesting that Kepler and Maxwell 1 have better OpenGL support than Maxwell 2 and Pascal?

 

[Carfax83]

"Poor performance in Vulkan"? The benches I saw mostly showed Maxwell and Pascal GPUs having measurable gains with Vulkan, just not to the extent that AMD did.

The benches you saw were not an accurate depiction of the performance then, because before NVidia published the 372.xx drivers, the Vulkan path had lots of stuttering and frame lag to the point where it was unplayable. I mentioned this early in the thread.

In any case, it is nice to see Nvidia getting more performance out of Vulkan with Pascal and Maxwell at least. One thing I like about these benchmarks is that it puts to bed the idea that Nvidia's OpenGL driver was just so good that Vulkan really only offered a marginal improvement, and Vulkan's performance leap on AMD cards was representative of what AMD's performance should have really been if they just had proper OpenGL support. I've seen people try to downplay DX12/Vulkan as just an attempt by AMD to shore up their crappy driver support, and these benchmarks prove that no, low-level APIs offer considerable improvement to Nvidia GPUs, too.

I don't know who told you that idea, but OpenGL itself has terrible CPU optimization even when compared to DX11, which means that in the more CPU dependent areas in the game, performance would suffer. If people weren't expecting much of a performance boost with Vulkan, it was likely because Doom itself isn't very CPU intensive with the exception of certain areas..

One particular exception I've already noted. The beginning of the Argent Tower level runs very poorly on OpenGL, because with Vulkan I got a massive 150% boost in framerates, which is almost obscene!

Nvidia hardware still doesn't support asynchronous compute in the way AMD does, so there's no guarantee they can use it for the same kind of boost in a game like Doom.

When are AMD fans going to get it.. Asynchronous compute has no specific implementation. AMD's implementation is no more right than NVidia's.. If asynchronous compute doesn't give Pascal a boost in Doom, then it's because ID failed to optimize it properly..

However I do doubt that NVidia will get as large a boost in Doom as AMD, but that's because Pascal is less likely to have as many execution bubbles as GCN at any given time.

And shader intrinsics are an AMD initiative made easier by the fact that they can be ported over from development on the GCN hardware in the PS4/Xbone. So who knows if Nvidia will even bother with that.

Shader intrinsic functions aren't an AMD initiative. They are a consequence of low level APIs finally being able to expose and make use of these advanced GPU features that have been there for years. NVidia already updated NVAPI to support shader intrinsic functions, and Vulkan will implement them just like they've done for GCN.

 

[Red Hawk]

The benches you saw were not an accurate depiction of the performance then, because before NVidia published the 372.xx drivers, the Vulkan path had lots of stuttering and frame lag to the point where it was unplayable. I mentioned this early in the thread.

Ah, ok then. Hadn't heard that, and I have no Nvidia cards to test myself.

One particular exception I've already noted. The beginning of the Argent Tower level runs very poorly on OpenGL, because with Vulkan I got a massive 150% boost in framerates, which is almost obscene!

Yep, pretty incredible. Doom is pretty much the gold standard for how low-level API conversions should turn out.

When are AMD fans going to get it.. Asynchronous compute has no specific implementation. AMD's implementation is no more right than NVidia's.. If asynchronous compute doesn't give Pascal a boost in Doom, then it's because ID failed to optimize it properly..

However I do doubt that NVidia will get as large a boost in Doom as AMD, but that's because Pascal is less likely to have as many execution bubbles as GCN at any given time.

There may not be a "right" way to support asynchronous compute, but AMD and NVidia's implementations are different, so there's a possibility one is better than the other. Much like how Fermi's tessellation was better than Cypress/Cayman's by a country mile. And the fact that it's taking so long for Nvidia to get it implemented (along with shader intrinsics) doesn't inspire me with the greatest of confidence that it will make much of difference.

 

[Unreal123]

Well the new update has quiet some improvement for Nvidia. GTX 1080 getting some good gains in Vulkan.

Vulkan on 1080p ultra settings

Open GL on 1080p Ultra settings

 

[ocre]

Crickets?

Are there any websites doing a formal comparison after update?

 

[antihelten]

Well then I guess all benches for Vulkan are incorrect then, since every reviewer uses PresentMon..

Moreover it would mean that every benchmark that isn't done with FCAT is incorrect, seeing as presentmon is one of the most accurate tools available today (more so than FRAPS)

When are AMD fans going to get it.. Asynchronous compute has no specific implementation. AMD's implementation is no more right than NVidia's.. If asynchronous compute doesn't give Pascal a boost in Doom, then it's because ID failed to optimize it properly..

Whilst there techically isn't any specific defined way to implement Async, and thus no right or wrong as far as the current implementations are concerned, there is certainly a better or worse implementation, with AMD obviously belonging to the first and Nvidia to the latter.

And I don't see how you can possibly conclude that the lack of performance boost for Pascal from async is ID's fault and not Nvidia's.

 

[Arachnotronic]

good, looks like NVIDIA users -- the majority of the people playing id's game on PC -- finally have a decent implementation of Vulkan. Now they need to get SLI/CrossFireX working.

 

[Bacon1]

finally have a decent implementation of Vulkan

Yep glad Nvidia finally got their drivers in order.

Guess that means you should retract this statement:

The (relatively) poor performance on the Vulkan codepath for NVIDIA has nothing to do with NVIDIA's drivers, IMO, as a lot of what the driver handled before is now handled explicitly by the developer.

http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/guru3ddoom-vulkan-benchmarks.2479898/page-11#post-38389181

 

[Arachnotronic]

Yep glad Nvidia finally got their drivers in order.

Guess that means you should retract this statement:



http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/guru3ddoom-vulkan-benchmarks.2479898/page-11#post-38389181

I'm sorry, Doom using GCN shader intrinsic functions (and not the intrinsic functions exposed by NVAPI for NVIDIA GPUs) and Doom not implementing asynchronous compute on NVIDIA hardware is an NVIDIA driver problem?

I don't buy it.

 

[Bacon1]

good, looks like NVIDIA users -- the majority of the people playing id's game on PC -- finally have a decent implementation of Vulkan

NVIDIA hardware is an NVIDIA driver problem?

So all that has changed was the NVidia driver. Yet you are claiming that Nvidia drivers have nothing to do with it?

I don't buy it.

Couldn't have said it better myself.

 

[Sweepr]

I'm sorry, Doom using GCN shader intrinsic functions (and not the intrinsic functions exposed by NVAPI for NVIDIA GPUs) and Doom not implementing asynchronous compute on NVIDIA hardware is an NVIDIA driver problem?

I don't buy it.

And it looks like we were completely right, they rushed this Vulkan patch when only AMD's implementation was ready (drivers + GCN shader instructions + exclusive async compute) to spoil GTX 1060's launch.

Edit: wrong thread.

 

[Arachnotronic]

Just trying to start a flame war, as usual. And it looks like we were completely right, they rushed this Vulkan patch when only AMD's implementation was ready (drivers + GCN shader instructions + exclusive async compute) to spoil GTX 1060's launch. Unfortunately it didn't last long.

I hope not, flame wars aren't fun. Anyway, I was planning on giving this game another play-through next week when I have finished moving my rig into a new case. Glad things are getting better with it for NVIDIA users.

 

[Sweepr]

I hope not, flame wars aren't fun. Anyway, I was planning on giving this game another play-through next week when I have finished moving my rig into a new case. Glad things are getting better with it for NVIDIA users.

That would be nice. Please test it on your Titan X (P) with the older and new drivers.

 

[Arachnotronic]

That would be nice. Please test it on your Titan X (P) with the older and new drivers.

Will do.

 

[Bacon1]

Just trying to start a flame war, as usual. And it looks like we were completely right, they rushed this Vulkan patch when only AMD's implementation was ready (drivers + GCN shader instructions + exclusive async compute) to spoil GTX 1060's launch. Unfortunately it didn't last long.

What are you talking about?

The only thing that has changed was Nvidia's Vulkan drivers to improve performance. That means the game itself (id's work) was perfectly fine all along. There was no giant conspiracy by id and AMD, just lacking Nvidia drivers which have now been fixed.

 

[Arachnotronic]

What are you talking about?

The only thing that has changed was Nvidia's Vulkan drivers to improve performance. That means the game itself (id's work) was perfectly fine all along. There was no giant conspiracy by id and AMD, just lacking Nvidia drivers which have now been fixed.

The updated drivers aren't going to add support for shader intrinsics exposed via NVAPI, and it's unlikely that async compute support will get magically turned on just from an update from NVIDIA. NVIDIA may have improved their Vulkan driver but the AMD code path still appears to have gotten more optimization.

 

[Bacon1]

The updated drivers aren't going to add support for shader intrinsics exposed via NVAPI, and it's unlikely that async compute support will get magically turned on just from an update from NVIDIA. NVIDIA may have improved their Vulkan driver but the AMD code path still appears to have gotten more optimization.

And so performance could be improved more which no one is disputing, but that doesn't negate everyone saying that it was not the fault of drivers. Here is your quote again:

The (relatively) poor performance on the Vulkan codepath for NVIDIA has nothing to do with NVIDIA's drivers, IMO, as a lot of what the driver handled before is now handled explicitly by the developer.

That is obviously false as the driver update has increased performance a lot.

 

[PowerK]

Well the new update has quiet some improvement for Nvidia. GTX 1080 getting some good gains in Vulkan.

Vulkan on 1080p ultra settings

Open GL on 1080p Ultra settings

By update, do you mean updated GeForce driver (v372.70 WHQL) or game update ?
If it the latter, I don't see Steam updating my Doom.

 

[Azix]

So the async compute games have been using is now considered "AMD's" implementation? Really doubt it works that way. Nvidia's way of doing things is just a hack. Not sure why people even call it async. All they are doing afaik is better balancing compute and graphics tasks across shaders. I would guess how that goes should really be up to their drivers. The only thing a developer can do is make it a really basic implementation that pascal can handle, like 3dmark time spy.

The vulkan results just show improvements can be made with driver optimizations, as before. Seems mostly CPU related

 

[dogen1]

By update, do you mean updated GeForce driver (v372.70 WHQL) or game update ?
If it the latter, I don't see Steam updating my Doom.

New driver.

I would guess how that goes should really be up to their drivers. The only thing a developer can do is make it a really basic implementation that pascal can handle, like 3dmark time spy.

What is "a basic implementation" supposed to mean exactly?

 

[Red Hawk]

I'm sorry, Doom using GCN shader intrinsic functions (and not the intrinsic functions exposed by NVAPI for NVIDIA GPUs) and Doom not implementing asynchronous compute on NVIDIA hardware is an NVIDIA driver problem?

I don't buy it.

That's not what you said.

"The (relatively) poor performance on the Vulkan codepath for NVIDIA has nothing to do with NVIDIA's driver"

This has been directly demonstrated as false, as driver side changes resulted in improved performance for Nvidia. If the poor performance had "nothing to do with Nvidia's driver", as you claimed, these improvements should not have happened. Looks like there were things holding back Nvidia's performance other than shader intrinsics and asynchronous compute.

 

[raghu78]

So Nvidia got their Vulkan drivers in good shape for Doom and we are seeing a good perf improvement. People who were arguing that the poor Doom Vulkan performance was due to id were just plain wrong.

 

[ThatBuzzkiller]

The updated drivers aren't going to add support for shader intrinsics exposed via NVAPI, and it's unlikely that async compute support will get magically turned on just from an update from NVIDIA. NVIDIA may have improved their Vulkan driver but the AMD code path still appears to have gotten more optimization.

Just for comedic effect I wanted to point out that there is no such thing as NVAPI for Vulkan. NVAPI are a set of D3D driver extensions and I'm pretty sure inserting D3D or HLSL interface extensions in Vulkan/SPIR-V does not work ...

Your mistaken if you think there's no way for Nvidia to support shader intrinsics when they could add them in their drivers, even the AMD specific extensions!

 

[Azix]

New driver.

What is "a basic implementation" supposed to mean exactly?

predictability might be the main thing. But

http://www.overclock.net/t/1606224/...-less-compute-parallelism-than-doom-aots-also
http://www.overclock.net/t/1606224/...llelism-than-doom-aots-also/180#post_25365469

What stands out to me is that the compute tasks wait for graphics tasks in the chart futuremark provided.

I doubt think it will transfer to actual games.