Doom Vulkan with different CPUs

[D.P.]

There have been a lot of posts recently showing confusion over what the new Vulkan and DX12 APis bring to the table. Some people mistakenly seem to think there is some kind of magical instantaneous performance boost and any card that doesn't get faster is somehow not supporting the API. this is somewhat flawed logic, the primary advantages, at least in the near term, is reduced CPU overhead. The multi-threading and reduced draw call overheads reduce any CPU bottlenecks.

if performance doesn't increase, that merely means there isn't a CPU bottle-neck with the specific GPU and CPU combo. This is very apparent when looking at a low-end GPU combined with a high end CPU which sadly is frequently the case in reviews. Things get very interesting when the CPU power is reduced

here we can see with a last CPU the 1060 doesn't see a big gain in performance, which indicates there is no CPU bottleneck and the OpenGL have a low overhead and well optimized. In contrast, the 480 sees a very healthy gain in performance which means the card is suffering some kind of bottle neck in OpenGL. A combination of high driver overhead and limitations of the GPU front-end coming to play.

However, with older slower CPU's even the 1060 starts to see very healthy gains in performance using Vulkan, in fact performance keeps very close to the experience with the fast CPU. The 1060, Vulkan API and the NVidia Vulkan drivers are doing exactly what is expected - reducing the bottleneck of the CPU.

The 480 still see a good benefit form using Vulkan but even there the performance is degraded significantly with slower CPUs. The effect is so big that the 1060 commands a significant performance lead, be it in openGL or Vulkan over the 480.



This is soemthign to keep in ind if you have an older computer and are looking to buy one of the new mainstream GPUs . Performance with a very high end CPU may not be painting the true performance picture.

 

[D.P.]

Isn't thi showing that even with no CPU bottleneck the 480 gains a lot with Vulkan, whereas the 1060 does not?

I'm referring to the 6700K results. I'd assume a 6700K at 4.5 Ghz would not be a bottleneck to the 480...

That depends on the quality of AMD's OpenGL drivers which are notoriously bad.

The performance improves can only come from reducing a bottleneck. the primary one being the CPU and the number of draw calls but the GPU front end such as the command processor can be limiting as well. AMD were hinting that they did a lot of work in improving the command processor but it isn't very evident from reviews so far. The performance delta between DX11 and DX12 should be smaller with Polaris than with Fiji and Hawaii but that doesn't look to be obvious.

 

[D.P.]

Doesn't explain why some nvidia users are getting worse performance when running Vulkan.

Bug in the game or driver, who knows. If performance degrades then it is obviously not normal and it needs fixing. Same when you see DX12 results slower than DX11, that means the developer has failed to optimize their code patch as well as the DX11 driver code-path. There is nothing wrong with the DX12 drivers or hardware when that happens, it quiet natural that a game developer under time pressure can;t do as good a job as a nvidia engineer who knows the architecture details to a T.

 

[D.P.]

This is beginning to give more credence to the claims on this forum about better scope for improvement with the RX480. Assuming more developers utilise Vulkan.

Actually I think it shows somewhat the opposite. AMD can still do a lot with OpenGL and Dx11 drivers to close the performance gap.


The other thing is in the future there may be games where the number of draw call increases substantially. Under those scenarios even Nvidia's DX11 and OpenGL might start becoming bottlenecked by the CPU and thus the DX12/Vulkan APIs will facilitate much better performance improvements. As the moment, Nvidia's driver stack simply means that DX12/Vulkan don't always lead to such gains.

One of the exceptions to this is the DX12 Timespy benchmark that has massively increased draw-calls. PascalGPUs do really well at leveraging the API

 

[D.P.]

So either the cpu's are bottlenecking the 480

OR

The 480 is attempting to offloading some work to the CPU that the 1060 does not and it can't keep up with demand.


I have to say ... that is quite a nasty looking graph.

Is there anything else to this, like stupid levels of tessellation going on or something, that isn't immediately obvious here.

No, itis the PAI behavign exactly as expected.

Vulkan and Dx12 reduced CPU overheads. No CPU, bottleneck then no performance gain.

 

[D.P.]

The rest of the results show very little in it between dx11 for nvidia and AMD quite surprised tbh maybe the extra driver work that went into the new GPU.

Conclusion
Well I didn’t manage to run into any compatibility issues with my Radeon RX 480 on these old systems but I did end up finding some interesting results. For the most part the 480 didn’t fare nearly as poorly against the GTX 1060 as I thought it might. There were of course exceptions though such as Doom and Ashes of the Singularity using DX11.
It was interesting, although admittedly not all that surprising, to see just how much slower these new mid-range graphics cards were on these old quad-core systems. It’s also been a long time since I’ve seen the Core i5-750 face the Phenom II X4 955 and honestly I had forgotten how much better the Intel processor was for gaming. The performance difference is likely being amplified by the use of modern games that take full advantage of quad-core processors, with the exception of ARMA 3.
In short it appears you’re going to see similar performance margins between the RX 480 and GTX 1060 on older hardware. On that note you’re also going to see considerably greater performance when using a modern processor or at least a relatively modern processor overclocked. I didn’t have time to overclock the Core i5 and Phenom II X4 processors for this video but perhaps I can do a more indepth video in the near future featuring more GPU and GPU configurations if you guys are keen to see it.

http://www.hardwareunboxed.com/gtx-1060-vs-rx-480-in-6-year-old-amd-and-intel-computers/

yeah, in general when you run a modern GPU on old PCUs then the PCu becomes the performance bottle neck. that is particularly exciting but explains why reviewers will sue a fats PCU to highlight differences between raw GPU performance.


What I found interesting was the strong evidence that Pascal GPUs do indeed see a good benefit form running Vulkan in scenarios where you expect Vulkan to help over OpenGl.

The other common scenario is very fast GPUs, and again we see the 1080 see a healthy performance boost form using vulkan when the slower 1060 doesn't on a fast CPU.

 

[D.P.]

The interesting thing here, is that graph taken in isolation makes the RX 480 look total trash for those on older hardware. But then if you goto the link, and look at the other graphs it is NOTHING like the graph shown for doom. The 480 is mostly behind the 1060 for sure, but certainly nothing like the graph posted at the top.

The purpose of this thread wasn't compare the 480 to 1060 across all games. With old CPUs you expect to see a smaller difference between GPUs of different performance.

The interesting part is looking Vulkan and/ DX12 perform in scenarios where you expect there to be a benefit.



When looking at the relative performance with an old slow CPU the average differences are smaller, the 1060 tend to be slightly faster, but there are no anomalous leaps like in Doom under Vulkan.

People recommending the RX480 due to the Doom Vulkan performance should be careful to acknowledged the CPU dependence. In contrast with an old PU it is hard to argue one or the other form a purely performance perspective. And People claiming Nvidia don't see a benefit form Vulkan/DX12 shoudl check their facts.

 

[D.P.]

i find it hard to believe that my 6600k i5 was bottle necking my mid range rx 480.

got an awesome boost with vulkan over ogl.

AMD's OpenGL driver are notoriously bad.

 

[D.P.]

We need more tests, one benchmark and a chart for that matter isn't enough.

You act surprised at something which should be absolutely expected form the entire purpose of Vulkan/DX12.

 

[D.P.]

So please explain to me why the 6700k would be the most CPU limited in that original chart? Surely the 480 would be LEAST limited by that CPU compared to the others? But then it makes a massive jump in performance. Your explanation is the exact opposite of what we are seeing in that chart.

The RX480 may well have other bottlenecks in the GPU front-end that are resolved with Vulkan on a faster CPU, or AMD's VUlkan driver have issues on slower CPUs. There are a lot of variables at pay and changing the CPU can have additional effects.

Non of that is particularly relevant to the main point of discussion though.

 

[D.P.]

I was talking specifically about the original chart you posted.

(to be utterly clear, I am NOT trying to 'catch you out' or anything like that, I just want to know what the hell is going on with those results).

As an aside, I sure as hell wouldn't but the 480 for one game.

TBH, I really don't know exaclty what happens witht he RX480, looks like there are compounding factors there.

My main discussion point was how Vulkan and Dx12 help Pascal GPUs in the scenarios where they are supposed to help the most.

I do find it interesting that the RX480 gains so much with a fast CPU but the whole architecture is a bit of a mystery. It massively under-performs and uses far too much power. AMD supposedly made lots of improvements to the command processor but it is hard to see that in benchmarks. The RX480 really performs the exact opposite of what i would have expected. I would have hoped Poalris would show a much smaller difference between dX111 and DX12/Vulkan, not a bigger difference. There may be some large DX11 driver gains to come.

 

[D.P.]

There are actual 2 games out there using Vulkan, The Talos Principle is seemingly ignored by fans of certain hardware brands.

The 1060 averages 119.6FPS at 1080 with Vulkan, the RX480 averages 68.9 FPS, making the 1060 a whopping 74% faster than the 480.


Are the results weird? yes, by no means if this going to be representative of every Vulkan game but then neither is Doom. It is very clear that different GPU architectures respond to different API in very different ways and it is impossible to draw any good conclusions from a very small sample size.

 

[D.P.]

Talos is mostly ignored by reviewers as well because it uses a Vulkan Wrapper. Just like Linux reviewers mostly ignore games that use DX11 wrappers, instead of running natively in OpenGL/Vulkan. It's not comparable to native versions.

All game engines use a wrapper around the graphics API, especially any engine that lets you test 2 or more APIs together like DX11 and DX12.

 

[D.P.]

Rubbish, any port of a game and engine that uses a Wrapper is significantly worse than a native port.

Many multi-api engines have native support for both and do not use wrappers which greatly impact performance in some cases; along with bringing in compatibility issues.

The DOTA OSX wrapper from DX to OpenGL was significantly worse performance wise than the later multi-api native version released by Valve. It's the case for many.

I don't think you have a clue what you are talking about. Exactly how many game engines have you worked on or developed?

Have a look at the unreal engine, all the graphics APIs are abstracted away in wrappers. No sane programmer on the planet wants to constantly access low level API commands, this is even more the case with DX12 and Vulkan. This is the entire purpose of these new APIS, the developer has to create a more encompassing wrapper that takes more control over the command queue and state

 

[D.P.]

My understanding these other games are ignored because they not true Vulkan API games.

They are just as true as any other Vulkan game.

The developers may not have spent as much efforts optimizing for particular hardware, but that is an important aspect of the new low level APIS. Instead of AMD and NVidia optimizing everything, the develop has far more responsibility.

 

[D.P.]

It's got nothing to do with anyone winning.. It's not a native Vulkan API game. So therefore it can't be compared to how well the api can perform.

Doom isn't a Native Vulkan game either by your definition - it has both an OpenGL and Vulkan support abstracted away under wrappers.

 

[D.P.]

Least with Doom it was built alongside the game and used to its full potential, other so called Vulkan games are not using it how it should be, reason performance increase is lacking.

ID love OpenGL and Vulkan will have been jumped on by them long time ago Doom is the first truly Vulkan API game.

No, with Doom they first coded in openGL and then they added Vulkan as a patch.

How do you know Doom is using Vulkan to it full potential? I would hope there are much more optimization to come in the future otherwise it is very disappointing.

Why are other games not using Vulkan as it should be? Can you provide any links to how Vulkan should or shouldn't be used?

 

[D.P.]

This whole 'truly' thing strikes me very much as a 'true Scotsman' sort of fallacy. If you're using the Vulkan API, you're using the Vulkan API. This talk about 'wrappers' and whatever nonsense is kind of ridiculous.

Vulkan and DX12 are like any other API in that they can be used to great effect or to poor effect or anywhere in between. At any point on the scale, it is still a 'DX12/Vulkan' title. Just like a DX11 title is still a DX11 title even if uses none of the multi-threaded capabilities it offers. Really, even moreso when it comes to a low level API because there's an entire world of optimization that you can bother with or not bother with depending on your level of experience/talent/resources/time/motivation/etc. Which is again why it's very dangerous to go around making blanket statements about what it means for a game if it's using one of these low level API's.

I'd say that even by your own True Scotsman definition, Doom will probably not be considered a 'true' Vulkan game in 5 years time when DX12/Vulkan have advanced far enough and engines have inherent adaptability with the new API's and there are far better gains to be seen. Or rather, what I expect to happen is not for games to simply be 'DX11 + better performance', but I think the extra capabilities will be used for actual game design advancements. The huge increase in draw calls possible make for game design decisions that simply aren't possible with DX11/OpenGL. But it also requires that DX11/OpenGL be ditched completely. So I think it's quite foolhardy to go around saying that Doom is a 'true' Vulkan game when it's still very much rooted in OpenGL development as a fundamental base.

This really. The increase in draw calls will allow some interesting advances in the future, but these aren't possible if a game still supports DX 11 or OpenGL. Even then it requires new generation of game engines.

 

[D.P.]

Read the Q&A above even the developer say they need to code more to get more out the API and gain performance. It's not using Vulkan the same way ID is on DOOM.

ID also need to do more to get more performance out of Vulkan, that is the nature of software development.

None of this makes the Talos Principle results any less relevant than Doom's. They both are useful sources of information as valid as each other.