AMD Polaris architecture – GCN 4.0

[Ferrari100]

No, thats not what i said at all.
Without async the 390x is not fully utilised, its a bigger chip than a 980 and SHOULD be faster.

If they both get 60fps without async, the 390x goes up to 70 with async, the 980 still gets 60 because its already fully utilised at 60, it hasnt gone down, theres no bottleneck, the 390X has had a bittleneck removed and so being a bigger chip it now gets a bigger score

Cool, so you state that AMD GPUs are more powerful when Async compute is used as it can extract more performance from the cores. So if AMD'S single threaded performance was already level with Nvidia and the 390x was already hitting that 70fps without Async compute, what performance increase would you allow on top of that in a game such as Ashes of the Singularity where currently the 390x finds 20% + performance when Async is being used?

I'm a teacher, I ask questions, it's a big part of my job.
If somebody can teach me something then I want to know so that I am confident in what I say when I pass my knowledge on to others.

Also you said 'No, that's not what i said at all'.
I never actually mentioned anything you said in my post, I did however ask a question that did not warrant that response, imo.

 

[pandem0nium]

Dx12 also allows mixed vendor mgpu.. so maybe we'll be running 1080s with polaris 11s for dedicated async compute, like the dedicated phys x cards people run now.

 

[Ferrari100]

Dx12 also allows mixed vendor mgpu.. so maybe we'll be running 1080s with polaris 11s for dedicated async compute, like the dedicated phys x cards people run now.

Get the best from both sides!
Imagine the Nvidia GPU handling GameWorks performance and the AMD GPU handling ASYNC compute.
It will be in interesting to see how many games manage to implement Mgpu.

 

[humbug]

Andy is doing a pretty good job of explaining this.

Put simply the fury is far more bottlenecked than the 980ti in DX11, Nvidia are just far more efficient at DX11.

The 390x has the same ish shading power as the 980ti, but Nvidia has far better instruction scheduling in DX11 so they are able to use the GPUs to full potential.
AMDs bottleneck is overcome in DX12.

Difficult to explain it better than that, again, on a tablet.

 

[pandem0nium]

Get the best from both sides!
Imagine the Nvidia GPU handling GameWorks performance and the AMD GPU handling ASYNC compute.
It will be in interesting to see how many games manage to implement Mgpu.

I'd love to see it adopted more but I can't see it happening much to be honest. A lot of people have igps doing nothing when some of the lighter work could be offloaded.

 

[queamin]

That hypothesis can proven in DX12 games that do not use ASync compute and I believe it will hold true due to DX12 making it easier to use multiple CPU cores where available.

Don't get confused with what DX12 offers in this respect and Async compute, they are two different things that improve performance in different ways.

But it all helps that is all I was saying Async compute is good and hopefully we see what it can do as we get dx12 games built from the ground up.

 

[Ferrari100]

But it all helps that is all I was saying Async compute is good and hopefully we see what it can do as we get dx12 games built from the ground up.

Absolutely. I am looking forward to that as well. It's a good time to be a gamer right now. If AMD can deliver on great price/performance and offer up even better DX12 performance in the future it is great for the industry and consumers alike.

 

[triss]

Andy is doing a pretty good job of explaining this.

Put simply the fury is far more bottlenecked than the 980ti in DX11, Nvidia are just far more efficient at DX11.

The 390x has the same ish shading power as the 980ti, but Nvidia has far better instruction scheduling in DX11 so they are able to use the GPUs to full potential.
AMDs bottleneck is overcome in DX12.

Difficult to explain it better than that, again, on a tablet.

+1

Simply drivers etc on dx11 let nvidia cards work to 100% for eg yet AMDs work at say 90% and dx12 in some cases lets this be 100% . Nvidia cant gain from removal of a bottleneck if there isnt one for them.

 

[Ferrari100]

Andy is doing a pretty good job of explaining this.

Put simply the fury is far more bottlenecked than the 980ti in DX11, Nvidia are just far more efficient at DX11.

The 390x has the same ish shading power as the 980ti, but Nvidia has far better instruction scheduling in DX11 so they are able to use the GPUs to full potential.
AMDs bottleneck is overcome in DX12.

Difficult to explain it better than that, again, on a tablet.

OK so if Async brings the 390x in line with the 980ti simply due to realising it's full potential, are you suggesting that Async compute is a waste of time for Nvidia as their GPUs are already running at their full potential?

 

[humbug]

OK so if Async brings the 390x in line with the 980ti simply due to realising it's full potential, are you suggesting that Async compute is a waste of time for Nvidia as their GPUs are already running at their full potential?

Largely tho not completely, Nvidia still gain tho nothing like AMD, keep in mind Fury X does have more grunt that a 980ti

 

[CAT-THE-FIFTH]

OK so if Async brings the 390x in line with the 980ti simply due to realising it's full potential, are you suggesting that Async compute is a waste of time for Nvidia as their GPUs are already running at their full potential?

Because both companies went with different paths - AMD tried to be forward looking like Fermi was forward looking, but the software support was not quite there - it means under dx11 AMD cards are more likely to not hit their full potential as easy as the Nvidia cards. You can even see this with AMD dx11 CPU scaling too.

This is why DX12 and to some degree Mantle did help out.
Plus both the AMD and Nvidia GPUs are probably prioritising different things in their design. Even if Pascal might not be effective at async as say Polaris or Vega,Nvidia can gain performance through other aspects of the design. Also remember even with the Async boost,most of the DX12 performance with AMD cards is still down to other factors - IIRC,at least for the PC it is not more than 10% to 20% improvement with async although I could be wrong on that.

You also got things like conservative rasterisation,etc which AMD does not currently support too.

 

[Ferrari100]

So, one week to go until we see just how bright Polaris actually is.
I guess it will be on shelves within a couple of weeks after that?

 

[Calin Banc]

So does this remove all AMD'S bottlenecks? If so Nvidia are left with a significant bottle neck somewhere.
For example, if a 390x is generally around 980 performance without Async but experiencing a 20% boost in performance when using ASYNC, where does the 980 bottleneck lie that prevents it from keeping up? I'm interested to know.

I thought 390x and 980ti had about the same GFLOPs it is just harder to get Amd cards to max out their GFLOPs before dx12, but that don't mean 390x is as powerful as a 980ti just means sometimes it can close the gap a bit in dx12 to dx11.

Cool, so you state that AMD GPUs are more powerful when Async compute is used as it can extract more performance from the cores. So if AMD'S single threaded performance was already level with Nvidia and the 390x was already hitting that 70fps without Async compute, what performance increase would you allow on top of that in a game such as Ashes of the Singularity where currently the 390x finds 20% + performance when Async is being used?

I'm a teacher, I ask questions, it's a big part of my job.
If somebody can teach me something then I want to know so that I am confident in what I say when I pass my knowledge on to others.

R390x has about 28% more theoretical TF than 980 and ~ 5% more than the 980ti and is also faster in texture rate.

However 980 has 7% higher pixel rate and 980ti 43% higher than R390x.

In general for 980/ti you can add higher numbers since most are coming factory overclocked.

It depends per game which architecture likes best. I don't believe that nvidia has the chip that fully utilized since it's still doing stuff the old dx11 way. Supporting async compute would still mean some percentages gained. If async is not useful, then why console guys went with it and AMD as well? Why not just make a game that runs faster the old way? Probably because of this (it's from Total War guys blog on steam)?

In GPU terms, we’ve shifted our particle simulation pipeline from the pixel shader to the compute shader, which is a more efficient use of the GPU’s time. In fact we’ve done this with several parts of the rendering pipeline, further utilizing the GPU and letting the CPU focus on everything else it has to do.

Doing more efficiently stuff with compute shaders, while doing other stuff in the same time in no specific order surely means a step in the right direction, isn't it?

 

[ICDP]

Doing more efficiently stuff with compute shaders, while doing other stuff in the same time in no specific order surely means a step in the right direction, isn't it?

Exactly this. I cannot fathom why every Nvidia fan isn't asking the direct question "does Pascal have proper hardware aysnc compute"? Instead they are more likely to claim that async compute doesn't matter since Nvidia have most market share in PC graphics and wont be used anyway. Or that Pascal is faster than AMD in DX12 anyway so it doesn't matter, even despite the fact that Fiji closes the gap on Pascal in DX12 (which should be a big warning sign). They ignore the fact that there has not been a single AAA PC exclusive game for years. Console devs are utilising every trick they can to improve performance and that includes async compute on console hardware which means PC GPUs with proper hardware async compute will benefit. We have already seen multiple devs clarify that async compute does bring performance benefits and that they ARE already utilising them. We have already seen some DX12 games that benefit AMD hardware as a combination of removing driver overhead and async compute. There is a pattern in DX12 already emerging where even older AMD GPUs are catching up with, or even exceeding performance on historically faster Nvidia GPUs.

DX11 will become less and less relevant over the coming years as DX12 and Vulkan become the norm. If Pascal does not have proper hardware async compute then that is a definite disadvantage because we know for a fact AMD have it and DX12 games are using it more and more.

I would be waiting on purchasing any Pascal based hardware until it was 100% confirmed either via Nvidia themselves, or from more DX12 testing that async hardware is included. Because like it or not it is clear async compute IS being utilised in DX12 and if Pascal does not have hardware async compute it WILL be a serious lack of future proofing. Instead of Nvidia fans saying "yeah, give us this free performance increase as well please", we are hearing "async compute is pointless anyway". It's not a good look guys.

It reminds me of this.
https://climatesanity.files.wordpress.com/2011/01/cartoon-from-trenberth-ams-paper.jpg

 

[flopper]

Exactly this. I have been reading from a number of the defenders of all thing Nvidia that async compute doesn't matter since Nvidia have most market share in PC graphics. Or that Pascal is faster than AMD in DX12 so it doesn't matter, even despite the fact that Fiji closes the gap on Pascal in DX12 (which should be a big waring sign). They ignore the fact that there has not been a single AAA exclusive console game for years. Console devs are utilising every trick they can to improve performance and that includes async compute on console hardware which means PC GPUs with proper hardware async compute will benefit. We have already seen multiple devs clarify that async compute does bring performance benefits and that they ARE already utilising them. We have already seen some DX12 games that benefit AMD hardware as a combination of removing driver overhead and async compute. There is a pattern in DX12 already emerging where even older AMD GPUs are catching up with, or even exceeding performance on historically faster Nvidia GPUs.

DX11 will become less and less utilised over the coming years and if Pascal does not have proper hardware async compute then that is a definite disadvantage because we know for a fact AMD have it and DX12 games using it more and more.

I would be waiting on purchasing any Pascal based hardware until it was 100% confirmed either via Nvidia themselves, or from more DX12 testing that async hardware is included. Because like ot or not it is clear async compute IS being utilised in DX12 and if Pascal does not have hardware async compute it WILL be a serious lack of future proofing.

Pascal as well as Maxwell cant do two things at the same time which for example AMD can do with their ACE so there is no latency hit for AMD when such async is used. There is with both Pascal/Maxwell no way around that for nvidia. There is simple a lack of hardware support.

Poalris as well as Vega however will have async support with hardware.
LiquidVR with Mantle allowing a low latency along the way.

Depending what AMD did for Polaris with previous DX versions bottlenecks we can expect Polaris/Vega be superb for DX12 and Vulkan and BF1 will run splendid on AMD hardware and be the leader in the field. Nivida might have some benefit still for DX11 but thats going away and your not buying cards for past games solely as you also want to be playing newer games as they come along.

Like BF1

or like the upcoming warhammer
http://gameranx.com/updates/id/54685/article/total-war-warhammer-directx-12-benchmarks/
Dx12 is a big thing.

"The DirectX 11 build averaged about 45 frames per second on the R9 290 at 1080p. The framerate dropped to 38 to 42 when the camera was taken close to the ground to inspect a group of soldiers and reached a maximum of 51 when pulled out. Meanwhile, the DirectX 12 benchmark averaged 74.8 frames per second. That’s a stunning 66.22% increase. The minimum for the DX12 benchmark was 68.5 frames with a maximum of 81, 30 frames higher than what we were getting in the DX11 custom battles."

 

[jupiter_man]

When are we gonna get more news from AMD?

 

[N19h7m4r3]

When are we gonna get more news from AMD?

Computex June 1st, and then E3 again when they hosting the PC Gaming show again.

 

[Hiijinx]

Hurry up AMD, all this talk about 1080's is actually pulling me back to the dark side out of rebellion!

 

[D.P.]

Exactly this. I cannot fathom why every Nvidia fan isn't asking the direct question "does Pascal have proper hardware aysnc"

Because Pascal does have proper async co lyre, it just isn't the same implementation that AMD has and there is absolutely no requirements to. That makes everything else you say completely irrelevant.

 

[D.P.]

Pascal as well as Maxwell cant do two things at the same time

Completely wrong again.