AMD: Speculation, Rumors, and Discussion (Archive)

[CSI PC]

According to Newegg the processor Guru3D uses in their reviews is 3.0GHz, but their test setup for the RX 480 review runs it at 4.4GHz.

Yeah it is an extreme 8-core Haswell.
And so it should in theory seen the same performance benefit-trend as DSOGaming with their enthusiast OC 6-core Ivy Bridge if there is something unusual and improved in terms of thread-core behaviour with Vulkan.
But Guru3d saw no gains on Nvidia cards.
Cheers

 

[CSI PC]

Which again doesn't contradict what I've been saying. It's a CPU limited situation (720p, even lower than the 1080p that we had in the Computerbase.de benchmarks). Nvidia cards in Doom with the limited benchmark site sample size that we have so far indicate that Nvidia cards only see gains on Vulkan when the game is primarily CPU limited.

AMD GCN cards in contrast are seeing gains regardless of whether you are CPU limited or not.

Regards,
SB

However there are examples that go against what your saying about weak CPUs and Nvidia cards; these are not ideal but instead of looking at FPS, look at the ms time AVG for GPU as that gives a reasonable indicator of a broader time window if one does not have a true capture tool - agree not ideal but still.
Doom OpenGL 4.5 VS Vulkan Nightmare Settings 1080P | GTX 1080 | i7 5960X 4.4GHz

How does it run? Doom (2016) Vulkan vs OpenGL - GTX 980 / i7 6700k

And there are forum members on other sites seeing better performance with their Nvidia, and also seeing benefits at 1440p.
A friend on another forum is doing some testing for me with their OC 5820K and 980ti at 1440p, and they are seeing 12% boost pushing everything into Nightmare - measured using PresentMon; that is one of the lower results of those seeing benefits.
So there is something about the settings or OS that possibly is causing a divergence between those seeing zero gains to those that can see great gains at 1080p and still good gains at 1440P (I would say 12% is OKish rather than good).
DSOGaming was also 1080P, I agree that as resolution goes up more pressure would be put onto the GPU and settings will also impact this.
But to know more regarding resolution and pressure it puts onto the GPU in terms of openGL-Vulkan benefits, we would need to see the benchmark for those systems that actually gain a good boost at 1080p also run at 1440p.

Before I come to any conclusions, I would like to see some really careful and in-depth analysis looking at all the primary variables that may affect performance behaviour, including such as Borderless/Full Screen.

That last point has affected performance for a few with AMD cards, and probably will also for Nvidia.
Maybe we should take the Nvidia specific aspect and discussion into their own thread, although the variables and what discussed so far is still relevant to AMD.
Cheers

 

[lanek]

It could maybe be a good idea to move the posts about Vulkan Doom to a Vulkan performance thread, or mix it with the DX12 existing thread.

Because honestly, even if i know where it is, im not sure in 6 months i will think to come in the " AMD discussion, rumor thread for find it.

 

[Silent_Buddha]

However there are examples that go against what your saying about weak CPUs and Nvidia cards; these are not ideal but instead of looking at FPS, look at the ms time AVG for GPU as that gives a reasonable indicator of a broader time window if one does not have a true capture tool - agree not ideal but still.
Doom OpenGL 4.5 VS Vulkan Nightmare Settings 1080P | GTX 1080 | i7 5960X 4.4GHz

How does it run? Doom (2016) Vulkan vs OpenGL - GTX 980 / i7 6700k

And there are forum members on other sites seeing better performance with their Nvidia, and also seeing benefits at 1440p.
A friend on another forum is doing some testing for me with their OC 5820K and 980ti at 1440p, and they are seeing 12% boost pushing everything into Nightmare - measured using PresentMon; that is one of the lower results of those seeing benefits.
So there is something about the settings or OS that possibly is causing a divergence between those seeing zero gains to those that can see great gains at 1080p and still good gains at 1440P (I would say 12% is OKish rather than good).
DSOGaming was also 1080P, I agree that as resolution goes up more pressure would be put onto the GPU and settings will also impact this.
But to know more regarding resolution and pressure it puts onto the GPU in terms of openGL-Vulkan benefits, we would need to see the benchmark for those systems that actually gain a good boost at 1080p also run at 1440p.

Before I come to any conclusions, I would like to see some really careful and in-depth analysis looking at all the primary variables that may affect performance behaviour, including such as Borderless/Full Screen.

That last point has affected performance for a few with AMD cards, and probably will also for Nvidia.
Maybe we should take the Nvidia specific aspect and discussion into their own thread, although the variables and what discussed so far is still relevant to AMD.
Cheers

All of those are CPU limited scenarios to some extent. None of that contradicts Computerbase.de or Guru3d. Which all show that when you are not CPU limited, you have little to no gain and in some cases performance regression compared to OGL on Nvidia hardware.

And as I've stated numerous times, we have a very limited benchmark sample size. 2 sites so far. And one did very limited benchmarking. This gives an indication but is not definitive. I'd like to see more sites give benchmarks.

Regards,
SB

 

[Ike Turner]

Heads up...people seeing improved perfs on NVIDIA GPUs are for the most part running through the Foundry map.. The Vega Central Processing map actually has the opposite effect..it's slower on Vulkan....-10% in this case

 

[CSI PC]

All of those are CPU limited scenarios to some extent. None of that contradicts Computerbase.de or Guru3d. Which all show that when you are not CPU limited, you have little to no gain and in some cases performance regression compared to OGL on Nvidia hardware.

And as I've stated numerous times, we have a very limited benchmark sample size. 2 sites so far. And one did very limited benchmarking. This gives an indication but is not definitive. I'd like to see more sites give benchmarks.

Regards,
SB

How can you tell where both Guru3d and Computerbase.de captured their performance as it is not in their articles?
I must admit I cannot see how a FPS game can be CPU bound-limited with some of its maps when using say an OC 6-core Ivy Bridge enthusiast or newer I7 CPUs (such as in those videos Extreme Haswell and also 6700K).

I know you stated numerous times but so have I that assumptions cannot be made including by current review benchmarks.
Anyway to re-iterate the person testing for me on another site is using an internal zone, and avoiding any chance of CPU limitations while still gaining 12% so far at 1440p with nightmare setting and nightmare shadows, which is rather noteable compared to the 0% from both Guru3d and Computerbase.de
And I cannot see how you can call it indicative with those 2 reviews when there are plausible indicators suggesting others are seeing performance gains, albeit more anecdotally and using the average ms time with the internal benchmark such as DSOGaming and also Golem.de (who received a mention from Computerbase.de).
That said we do agree that nothing can be taken for granted and nowhere near enough information to date.

But separately, should we expect FPS/open environment (such as Mirror Edge Catalyst) genre on DX12/Vulkan in future to be CPU limited with good I7 type models on certain map-zones, and should this also be a consideration to those games DX11-to-DX12 already benchmarked?
This has not been raised so far to date by any publications even for existing DX11-to-DX12 games.
Cheers

 

[CSI PC]

Heads up...people seeing improved perfs on NVIDIA GPUs are for the most part running through the Foundry map.. The Vega Central Processing map actually has the opposite effect..it's slower on Vulkan....-10% in this case

That was the 2nd video I linked. [Mistake different video from same person]
However the 1st video using a 1080FE was not there, it was The UAC.
Where are you getting the information that most people are seeing the benefit in Foundry map rather than any other (possibly true but it is not the only one)?

Cheers

 

[Ike Turner]

That was the 2nd video I linked.
However the 1st video using a 1080FE was not there, it was The UAC.
Where are you getting the information that most people are seeing the benefit in Foundry map rather than any other?

Cheers

Nah this is another video by the same guy but this time running the Vega Central Processing Map..
Been browsing Reddit and several other places that's how I found this and comments about some maps being more demanding than others.

 

[CSI PC]

Nah this is another video by the same guy but this time running the Vega Central Processing Map..
Been browsing Reddit and several other places that's how I found this and comments about some maps being more demanding than others.

Thanks and yeah I think his intro caught me out.
Yeah so negative performance like you say on his 980 on that map lol, if he did not mess up his settings or cause a glitch would be interesting to see if it can be replicated on other systems and AMD cards - (if it is not a settings mistake-glitch then probably only Nvidia issue)
Cheers

 

[mczak]

I really dont see Apple starting to promote Metal for games developpement on MacOSX.
In reality, i dont think that Apple is interested to promote any games environnement on their product outside the iOS ecosystem ( tablet and smartphones, TV box. )

The problem with GL on macs is that it is _really_ lagging behind now. Feature wise it's basically frozen since mavericks (osx 10.9) - still stuck at GL 4.1 (with very few extensions on top of 4.1). AFAIK 10.12 won't change anything there. So if you want to use something "fancy" like compute shaders, you have to use metal...

 

[Ryan Smith]

The problem with GL on macs is that it is _really_ lagging behind now. Feature wise it's basically frozen since mavericks (osx 10.9) - still stuck at GL 4.1 (with very few extensions on top of 4.1). AFAIK 10.12 won't change anything there. So if you want to use something "fancy" like compute shaders, you have to use metal...

Apple is all-in on Metal. OS X, iOS, tvOS. If you're going to program graphics for an Apple device, they want you using Metal.

 

[sir doris]

Thanks and yeah I think his intro caught me out.
Yeah so negative performance like you say on his 980 on that map lol, if he did not mess up his settings or cause a glitch would be interesting to see if it can be replicated on other systems and AMD cards - (if it is not a settings mistake-glitch then probably only Nvidia issue)
Cheers

I think Vega Central is where I am now (the one where you have to shoot out the 4 cooling systems with all the deamons running wild?), and where I was when I transitioned to Vulkan. It's is smoother on Vulkan, IIRC under OpenGL I was seeing about 40-50 FPS with Vulkan that was 70-80 sustained on my 7970. The worst place I've see so far was the Curcibal after you've released the two doors, that was hovering around 30FPS under OpenGL, I've just retried it under Vulkan and there was no slowdown compared to the rest fo the level/s. 60-80ish FPS when it all kicks off.

To be honest I'm absolutely astounded Doom looks better than RAGE and runs much better than RAGE (even after all the pacthes for AMD), hats off to ID.

EDIT: Just checked the video, Vega Central is the level I was thinking about.

 

[Jawed]

Why not wait a while for the drivers and Vulkan code paths to settle down?

It appears id is serious about Vulkan for this game, so it seems likely there'll be more tweaks.

This may well be the best code available yet for the IHVs to target. In theory Vulkan, like D3D12, doesn't give the IHVs much scope for tuning their drivers. So I expect they'll dive deep, which will take a while. It is, after all, a whole game, not just a tech demo.

It would be interesting if it turns out that the game places more emphasis on compute with Vulkan. The gap between Fury X and R9 390 increases substantially with Vulkan over OpenGL, with that delta increasing at 1440p versus 1080p. If that's the case then compute could be a ceiling on NVidia's performance. Though compute and texturing are very hard to disentangle with GCN.

 

[Anarchist4000]

It would be interesting if it turns out that the game places more emphasis on compute with Vulkan. The gap between Fury X and R9 390 increases substantially with Vulkan over OpenGL, with that delta increasing at 1440p versus 1080p. If that's the case then compute could be a ceiling on NVidia's performance. Though compute and texturing are very hard to disentangle with GCN.

Not sure what effects ID is using compute on, but TSSAA would seem like an obvious one. I'm thinking the bottleneck is more likely bandwidth then compute in this case. That would explain why the Fury is so far out ahead in those benchmarks. Difficult to know what, if any, compression techniques are actually working there. Techniques for compression in compute shaders will probably be a significant concern in the coming years. Are there any instances where color compression works in a compute shader?

 

[Ike Turner]

So my guess was confirmed by AMD. DOOM is the first and only game so far to make use of Shader Intrinsic functions (probably one of the reasons for the big difference in performance boost between DX12 games and the Vulkan path in DOOM and why GCN perf > NV GPU right now).

AMD answer to my question on Twitter:
https://twitter.com/SasaMarinkovic/status/753055630756159488

 

[Jawed]

Not sure what effects ID is using compute on, but TSSAA would seem like an obvious one. I'm thinking the bottleneck is more likely bandwidth then compute in this case. That would explain why the Fury is so far out ahead in those benchmarks. Difficult to know what, if any, compression techniques are actually working there. Techniques for compression in compute shaders will probably be a significant concern in the coming years. Are there any instances where color compression works in a compute shader?

The results for Fury X at 2160p (I didn't know those results were available until just now) contradict my point somewhat.

RX480, with delta colour compression should have more effective bandwidth R9 390, if DCC is "perfect". I'd be very surprised if compute shaders can consume a DCC-formatted render target. The hardware for decompression will be near the ROPs and in theory not on the path from memory to CU.

TSSAA should feature multi-frame accumulation, but I don't know how many frames. But being multi-frame, asynchronous compute is a perfect way to overlap "start of new frame" with "post-process and show old frame". So at the very least the CUs are less idle per frame.

 

[MDolenc]

So my guess was confirmed by AMD. DOOM is the first and only game so far to make use of Shader Intrinsic functions (probably one of the reasons for the big difference in performance boost between DX12 games and the Vulcan path in DOOM and why GCN perf > NV GPU right now).

AMD answer to my question on Twitter:
https://twitter.com/SasaMarinkovic/status/753055630756159488

So we now officially live in an era when games will receive significant changes even months after release (multiple DX12 updates on Rise of the Tomb Raider, Vulkan update on Doom) and completely distinct rendering paths for each vendor.
Shader intrinsics are available on DX12 as well. And on NV hardware (though I don't think NV has equivalents for all of them).
Should make life of reviewers much easier. :smile:

 

[CSI PC]

Not sure what effects ID is using compute on, but TSSAA would seem like an obvious one. I'm thinking the bottleneck is more likely bandwidth then compute in this case. That would explain why the Fury is so far out ahead in those benchmarks. Difficult to know what, if any, compression techniques are actually working there. Techniques for compression in compute shaders will probably be a significant concern in the coming years. Are there any instances where color compression works in a compute shader?

Also I assume the shadow setting details, as Async Compute is also enabled with AA off.
Would be interesting to compare on AMD between AA off and TSSAA, and the various shadow details.
Cheers

 

[OlegSH]

That would explain why the Fury is so far out ahead in those benchmarks

For whatever reason GTX1070 simply doesn't scale proportionally compared to GTX970 in these computerbase benchmarks. I mean that GTX1070 is 62-66% faster on average in GPU limited resolutions - https://tpucdn.com/reviews/NVIDIA/GeForce_GTX_1070/images/perfrel_3840_2160.png
So GTX1070 looks to be limited by CPU in these tests, 3840x2160 results only confirm this, the issues could be in driver or in the game code itself

DOOM is the first and only game so far to make use of Shader Intrinsic functions

Many of these functions have been available in DX11 for years via nvapi, shuffles for example, I wonder whether developers have been using them

 

[mczak]

RX480, with delta colour compression should have more effective bandwidth R9 390, if DCC is "perfect". I'd be very surprised if compute shaders can consume a DCC-formatted render target. The hardware for decompression will be near the ROPs and in theory not on the path from memory to CU.

As far as I can tell from the open-source driver, this isn't quite true. You can read from textures with dcc enabled (unlike msaa compression or depth compression where this doesn't work), both using samplers and shader images. You cannot, however, write to dcc enabled shader images.