AMD: Speculation, Rumors, and Discussion (Archive)

[Ike Turner]

Although it is another review like Guru3D that saw no improvements for any Nvidia.
Something is not right with all the testing because some are seeing still pretty good improvements for Nvidia cards (albeit not comparable to GCN but still into double figure % increase) and some reviews or tests seeing zero.
Bearing in mind this improvement is not just Async Compute but also low level API benefits.
Fingers crossed a publication actually does a very in-depth analysis, looking at the diverse variables that may influence performance.
Cheers

AFAICS Users seeing improvements with NVIDIA GPUs are not using 6700K or similar CPUs. That is the whole point of some of these benchmarks. Even when not CPU bound the GCN GPUs are registering insane performance boosts while the NV GPUs are performing relatively identical to the OGL path. NV's great OGL (and DX11) simply had way less CPU overhead then AMD's but something else is at play here beyond that. Vulkan (and to a lesser extend DX12) seems to be tailor made for GCN.

 

[SimBy]

l

This corroborates my experience with regards to the huge performance boots on the Fury X coupled with a 6700K. It totally nukes a 1070 at 1080p and 1440P (and is probably as fast or faster than a 1080).

https://www.computerbase.de/2016-07/doom-vulkan-benchmarks-amd-nvidia/

Which is pretty much in line with TFLOP difference between 1070 and FuryX. So I don't expect Nvidia 'fixing' this with any kind of driver update. Heck they have been working and demoing Doom on Vulkan since Pascal launch.

https://twitter.com/PellyNV/status/752515505978613760

 

[CSI PC]

AFAICS Users seeing improvements with NVIDIA GPUs are not using 6700K or similar CPUs. That is the whole point of some of these benchmarks. Even when not CPU bound the GCN GPUs are registering insane performance boosts while the NV GPUs are performing relatively identical to the OGL path. NV's great OGL (and DX11) simply had way less CPU overhead then AMD's but something else is at play here beyond that. Vulkan (and to a lesser extend DX12) seems to be tailor made for GCN.

So your saying that the Doom FPS game is still CPU-thread bound with a 6700K but not the 6 and 8 core CPUs....
Possible something about CPU but I am still a bit dubious (although like you I have mentioned CPU also needs checking in other posts).
Cheers

 

[el etro]

Vulkan unleashing all GCN potential is incredible.

Doom will be the Golden showcase guy for Vulkan. Vulkan is the biggest winner here.


It is also good because unties us to proprietary APIs.


I can't thank AMD enough for this.

 

[Silent_Buddha]

AFAICS Users seeing improvements with NVIDIA GPUs are not using 6700K or similar CPUs. That is the whole point of some of these benchmarks. Even when not CPU bound the GCN GPUs are registering insane performance boosts while the NV GPUs are performing relatively identical to the OGL path. NV's great OGL (and DX11) simply had way less CPU overhead then AMD's but something else is at play here beyond that. Vulkan (and to a lesser extend DX12) seems to be tailor made for GCN.

That shouldn't be a surprise. The iD guys have tweeted about how things like Async Compute and other GCN features allowed them to save massively on rendering time on console. At the time I imagine they couldn't talk specifically about the Vulkan version (unreleased, and still in development), but it's nice to see that it appears they've been able to port virtually all of that across to the PC.

Pascal won't be able to benefit from all of that, but if it has decent Async Compute support, that may allow it to see some gains. Nvidia cards prior to Pascal may be SOL, however.

Although it is another review like Guru3D that saw no improvements for any Nvidia.
Something is not right with all the testing because some are seeing still pretty good improvements for Nvidia cards (albeit not comparable to GCN but still into double figure % increase) and some reviews or tests seeing zero.
Bearing in mind this improvement is not just Async Compute but also low level API benefits.
Fingers crossed a publication actually does a very in-depth analysis, looking at the diverse variables that may influence performance.
Cheers

Either limited by CPU in general (as Ike Turner mentioned) or they are finding very specific locations on a level that aren't representative of general gameplay (IE - they are looking for where performance improves and ignoring most of the game where performance either degrades or doesn't improve).

Hence why I wait for actual sites to benchmark rather than relying on user benchmarks which may be tinted by trying to justify their hardware as the "best."

Regards,
SB

 

[CSI PC]

AFAICS Users seeing improvements with NVIDIA GPUs are not using 6700K or similar CPUs. That is the whole point of some of these benchmarks. Even when not CPU bound the GCN GPUs are registering insane performance boosts while the NV GPUs are performing relatively identical to the OGL path. NV's great OGL (and DX11) simply had way less CPU overhead then AMD's but something else is at play here beyond that. Vulkan (and to a lesser extend DX12) seems to be tailor made for GCN.

Just checked it cannot be due to CPU 8 core vs quad core; Guru3d also used an 8-core CPU - 5960X is their test CPU for all evaluations and reviews, and they had the same results as computerbase.de who used a 6700K.
Cheers

 

[Silent_Buddha]

Just checked it cannot be due to CPU 8 core vs quad core; Guru3d also used an 8-core CPU - 5960X is their test CPU for all evaluations and reviews.
Cheers

That's what he's saying. Nvidia users appear to be only be seeing general improvements in rendering time if they are CPU limited.

Regards,
SB

 

[CSI PC]

That's what he's saying. Nvidia users appear to be only be seeing general improvements in rendering time if they are CPU limited.

Regards,
SB

So how are users seeing improvements with their Nvidia cards?
At 1st it was because computerbase.de was using a quad core for the null benefits, but now there is the same results for a CPU that is not CPU-thread bound and being 8-core.
So Nvidia users cannot be using quad core Haswell/Skylake, nor using 6 or 8-core Haswells.
Thanks

 

[Silent_Buddha]

So how are users seeing improvements with their Nvidia cards?
At 1st it was because computerbase.de was using a quad core, but now there is the same results for a CPU that is not CPU-thread bound and being 8-core.
So Nvidia users cannot be using quad core Haswell/Skylake, nor using 6 or 8-core Haswells.
Thanks

They are seeing better results because they are using less powerful CPUs and/or running at lower resolutions. Hence, why users see some gains when they are CPU bound. Once they are GPU bound, those gains disappear or they actually lose performance versus OGL.

So lower resolutions as in the Computerbase.de graphs show some improvements. They are CPU limited at those resolutions.

Regards,
SB

 

[CSI PC]

They are seeing better results because they are using less powerful CPUs. Hence, why users see some gains when they are CPU bound. Once they are GPU bound, those gains disappear or they actually lose performance versus OGL.

So lower resolutions as in the Computerbase.de graphs show some improvements. They are CPU limited at those resolutions.

Regards,
SB

What do you deem as a weak CPU from Intel (including 6-cores)?
I think you are ommitting a lot of the footprint using 980tis by assuming they are on weak CPUs.
Thanks

 

[Silent_Buddha]

What do you deem as a weak CPU from Intel (including 6-cores)?
I think you are ommitting a lot of the footprint using 980tis by assuming they are on weak CPUs.
Thanks

Weak as in weaker than the ones typically used to benchmark games by sites that benchmark hardware. And weak as in lower resolutions.

If all you do is game at 1080p you'll likely see some gains on Nvidia hardware. If you game at 1440p or higher, you're less likely to see gains unless you're using top of the line Nvidia hardware or you lower graphical IQ settings. And even then your gains are already going to be reduced. And if you have less than top tier Nvidia hardware, you might start to see worse performance at 1440p on Vulkan versus OGL. And at 4k, there's a good chance you may see worse performance in general regardless of your Nvidia card.

That's in contrast to AMD hardware which appears to see universal performance gains on any generation of GCN hardware on any CPU used.

However, it's still early. Only 2 sites have done benchmarking that I've seen so far, and only 1 has done more than cursory benchmarking. Additionally, Nvidia may release new drivers that might make things better, we'll have to see. And there's still some hope at least on the Pascal front if Pascal actually does well with async compute as iD are working with Nvidia to try to get that enabled on Pascal.

Regards,
SB

 

[lanek]

Vulkan unleashing all GCN potential is incredible.

Doom will be the Golden showcase guy for Vulkan. Vulkan is the biggest winner here.


It is also good because unties us to proprietary APIs.


I can't thank AMD enough for this.

The real problem for Vulkan is still basically the same as for OpenGL...Doom have been developped on OpenGL first, ported to Vulkan then. Its a logical process in this case, similar of port a game from DX11 to DX12.

I can see Vulkan completely remove OpenGL of the equation for games developpement ( keep it maybe just for retro compatibility )..

But how many DX ( 12 ) games will be ported to Vulkan or developped directly on Vulkan instead of DX ? ( outside indies one, and then again with Unreal Engine and Unity, you get DX in the line ).

When you see that studios stilll seems having problem to assume multi-port developpement ( money wise ) between consoles and PC....

Ofc there's the question of SteamOS; but there again, difficut to know where it will go

 

[CSI PC]

Weak as in weaker than the ones typically used to benchmark games by sites that benchmark hardware. And weak as in lower resolutions.

If all you do is game at 1080p you'll likely see some gains on Nvidia hardware. If you game at 1440p or higher, you're less likely to see gains unless you're using top of the line Nvidia hardware or you lower graphical IQ settings. And even then your gains are already going to be reduced. And if you have less than top tier Nvidia hardware, you might start to see worse performance at 1440p on Vulkan versus OGL. And at 4k, there's a good chance you may see worse performance in general regardless of your Nvidia card.

That's in contrast to AMD hardware which appears to see universal performance gains on any generation of GCN hardware on any CPU used.

However, it's still early. Only 2 sites have done benchmarking that I've seen so far, and only 1 has done more than cursory benchmarking. Additionally, Nvidia may release new drivers that might make things better, we'll have to see. And there's still some hope at least on the Pascal front if Pascal actually does well with async compute as iD are working with Nvidia to try to get that enabled on Pascal.

Regards,
SB

Well by this chart even 6-core enthusiast Ivy Bridge are still as strong as enthusiast Haswell....
These results would not be influenced by higher resolutions as that puts the pressure on the GPU rather than CPU.

DSO Gaming used an Intel i7 4930K that they boosted to the level of a 4960X.
So if that is a weak CPU, we need to exclude all of I7 Haswell and some of the Haswell enthusiast models as well.
Not being pedantic, but the behaviour is more complex than just saying it is a weak CPU for reasons those say they see a noteable boost with a 980ti, compared anyway to Guru3d and Computerbase.de
There may be some kind of behaviour with the thread-core performance, but that should be seen even with Haswell-Extreme, but for some reason it is not replicated at Guru3d who do use such a processor.
Yeah appreciate these are DX11, bit tricky to find a useable chart with DX12 games, but the 5960x is very close to the 6700K with DX12 such as AoTS.
Thanks

Edit:
I am not sure even the OS can be ruled out as influencing performance, although I appreciate it is not meant to be tied to WDDM.

 

[Deleted member 13524]

But how many DX ( 12 ) games will be ported to Vulkan or developped directly on Vulkan instead of DX ?

In a parallel universe where Valve still makes games, they're probably using Vulkan.

 

[Dave Baumann]

There may be some kind of behaviour with the thread-core performance, but that should be seen even with Haswell-Extreme, but for some reason it is not replicated at Guru3d who do use such a processor.

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.

 

[Silent_Buddha]

But how many DX ( 12 ) games will be ported to Vulkan or developped directly on Vulkan instead of DX ? ( outside indies one, and then again with Unreal Engine and Unity, you get DX in the line ).

I don't think games will get ported to Vulcan from Dx12 in general.

However, if Apple ends up supporting Vulkan (somewhat doubtful as they've spent considerable time developing Metal), then we may see some developers (ones that also release games on Mac, like Blizzard) switch to developing on Vulkan rather than Dx12.

If that doesn't happen, not much changes I think. Developers that also do Mac ports will likely stay with Dx12/OGL as it'll make PC/console releases easier to manage. Doom was a special case as iD has always been fond of OGL due to also wanting to release their games on Linux. Vulkan is just a natural extension of that.

In the long term, I could see some PC exclusive developers targeting Vulkan rather than Dx12 as that also opens up easier Linux ports for sale on Steam.

A lot will also depend on rendering engine middleware. If UE/Unity/CryEngine/etc. embrace Vulkan in addition to Dx12, we could see more developers experiment with it.

Regards,
SB

 

[Silent_Buddha]

Well by this chart even 6-core enthusiast Ivy Bridge are still as strong as enthusiast Haswell....
These results would not be influenced by higher resolutions as that puts the pressure on the GPU rather than CPU.

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

 

[lanek]

I don't think games will get ported to Vulcan from Dx12 in general.

However, if Apple ends up supporting Vulkan (somewhat doubtful as they've spent considerable time developing Metal), then we may see some developers (ones that also release games on Mac, like Blizzard) switch to developing on Vulkan rather than Dx12.

If that doesn't happen, not much changes I think. Developers that also do Mac ports will likely stay with Dx12/OGL as it'll make PC/console releases easier to manage. Doom was a special case as iD has always been fond of OGL due to also wanting to release their games on Linux. Vulkan is just a natural extension of that.

In the long term, I could see some PC exclusive developers targeting Vulkan rather than Dx12 as that also opens up easier Linux ports for sale on Steam.

A lot will also depend on rendering engine middleware. If UE/Unity/CryEngine/etc. embrace Vulkan in addition to Dx12, we could see more developers experiment with it.

Regards,
SB

Completely agree with you, just for Metal, this API is destined to the iOS mostly, i will say that compatibility on MacOSX is just there because, you can use Metal for other graphic and image processing than for games. 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. )

 

[Grall]

DDU in safe mode

Naw. I did use AMD's cleanup utility to remove the previous driver though, but it hasn't been updated since last year, so maybe it doesn't do a proper job anymore...

Downloaded the DDU (again) now. I'll see what happens after giving that one a go. Thanks for the reminder!

 

[Ike Turner]

One thing to note (and this may be more relevant to the DX12 benchmarking thread) is the use of Shader Intrinsic Functions in the Vulkan Path. There's been a lot of chatter in the past 24Hrs about how Vulkan seems to be more "capable" than DX12 given the huge perf boost seen on GCN hardware compared to what we have seen in DX12 titles... BUT (and I may be totally wrong here so a developer like Sebbi can probably chime in to correct me) I wouldn't be surprised if DOOM is the first (and only) game/software to currently make use of Shader Intrinsic Functions and this may be why (in addition to Async Compute) we are seeing such a boost here and not in DX12 games yet. AMD released Shader extension access for Direct3D and Vulkan on May 24, 2016 (which is after all of the DX12 games where released besides the latest Tomb Raider patch) and then a few days later took down the Direct3D functions because of un-named issues requiring new drivers:

Note: Shader extension access for Direct3D has been temporarily removed while we wait for an updated driver that fixes certain issues.

It looks to me that DX12 apps/game are currently only leveraging Async Compute and the benefits of lower CPU overhead while Doom with Vulkan has the benefit of Shader Intrinsic functions on top of that. but then again I may be totally wrong there..