Updated Ashes of Singularity Benchmark-AMD Clobbers Nvidia

[DocSwag]

Sources:http://www.extremetech.com/gaming/223567-amd-clobbers-nvidia-in-updated-ashes-of-the-singularity-directx-12-benchmark and http://anandtech.com/show/10067/ashes-of-the-singularity-revisited-beta/5

The Ashes of Singularity benchmark has now been update, and now supports Asynchronous Compute.

Quote

With asynchronous compute disabled, AMD’s R9 Fury X leads the GTX 980 Ti by 7-8% across all three detail levels. Enable asynchronous compute, however, and AMD roars ahead, beating its Nvidia counterpart by 24-28%. The GeForce GTX 980 Ti’s performance, in contrast, drops by 5-8% if asynchronous compute is enabled. This accounts for some of the gap between the two manufacturers, but by no means all of it.

Asynchronous compute is basically AMD's secret weapon. AMD has had it with GCN since the 7970 but it was never fully taken advantage of. Now, though, we can see how much of an impact it makes. With normal DX12 the Fury X is beating the 980 ti not by much, but enable asynchronous compute and AMD's lead is astronomical.

this could mean a lot for AMD. DX12 seems to really favor AMD already, but enable Async Compute and AMD just crushes Nvidia.

GCN's hidden weapon has finally come out, and AMD is really getting Nvidia with this one.

Unfortunately, most games probably won't support it for a long time.

 

Now some of you might be confused. So what exactly is asynchronous compute?

As most of you know, most CPUs only have a handful of cores, while most GPUs have thousands of individual cores.

Normally, a GPU only does one operation at a time. However, when more cores(or stream processors for AMD) are added, sometimes not all of the cores are utilized.

Asynchronous compute aims to help get rid of this issue by letting the GPU do many things at once.

Most of you probably already know that the 980 ti has 2816 CUDA Cores while the Fury X has 4096. The Fury X also has more computing power than the 980 ti.

However, not all of the Fury X's cores always get utilized. However, by adding asynchronous compute you can utilize all of these.

Normally some of the Fury X's power is not utilized, while most of the 980 ti's is. Now, however, asynchronous compute better utilizes all of the power.

As well, the Fury X and every other gcn card has a sync compute/shading optimizations on the hardware level. Maxwell does not (I don't think so, I could be wrong). This also helps AMD out a lot.

Driver optimization is probably also part of the equation. Nvidia has to build this feature into its drivers to enable it. Anandtech also explains this quite well.

Quote

Again in a traditional environment all of this is fine, however as GPUs have advanced they have begun to test the limits of a single execution queue. As GPUs add ever more ALUs, even embarrassingly parallel begins to break down, and it is harder to keep a GPU filled the more ALUs there are to fill. Meanwhile new paradigms such as virtual reality have come along, where certain operations such as time warping require executing them with far less latency than the traditional high throughput/high latency execution model of a GPU allows. Thus GPU developers and software developers alike have needed the means to concurrently execute multiple jobs on a GPU’s ALUs, and this is where asynchronous shading comes in.

Whereas the traditional model is serial execution, asynchronous shading is executing multiple jobs over the ALUs at the same time. By implementing multiple queues within a GPU’s thread scheduler, a GPU executing jobs in an asynchronous manner can potentially run upwards of several jobs at once; more queues presents more options for work. Doing so can allow a GPU to be better utilized – by filling the underutilized ALUs with additional, related work – and at the same time work queues can be prioritized so that more important queues get finished sooner, if not as soon as outright possible.

However, don't get too optimistic yet. This is only a beta benchmark, and most likely won't get supported for quite a while.

DX12 probably won't be coming to games for quite a while, and even after they do, chances are they won't all support it.

This is interesting as it has given AMD a huge performance uplift. It will be interesting to see how Nvidia reacts to this.

Also, it isn't like DX 12 is making Nvidia's performance worse and AMD's a little better. It is giving both a performance uplift-AMD is just getting more benefit.

Quote

Also, if some of you suddenly saw a ton of information pop up all of a sudden its cause I went back and edited it so that it is more understandable.

Again, NO THIS DOES NOT MEAN AMD IS NOW BETTER THAN NVIDIA. Asynchronous compute won't be coming for at least a year and most likely around 3-4 years. Some games will probably start to support it in around 2 years but most games probably will take longer to do so.

EDIT: Check this article by Anandtech out http://www.anandtech.com/show/10067/ashes-of-the-singularity-revisited-beta goes pretty in depth to it. Nvidia seems to not have it optimized in their drivers, so that could be something to take note of.

Also if you guys are still confused check out this video made by AMD that explains a sync compute/shading quite well 

Thanks to @Jahramika for showing me the video! It really gives you a good visual to imagine it.

[Citadelen]

DX12 reinforces my decision to upgrade to Zen over Polaris, the 290X has massive gains when the driver overhead is removed.

[CommandMan7]

DANG IT. I just posted this a minute late - please let me finish my post, moderators. I am going to make a very high quality write up.

[Kobathor]

I cannot wait for a-sync compute games to come out, even if I get a new GPU I'd like to test them on my current 7970.

[don_svetlio]

So do we now recommend the Fury X over the 980 Ti? sort of confused

[Humbug]

3 minutes ago, Citadelen said:

DX12 reinforces my decision to upgrade to Zen over Polaris, the 290X has massive gains when the driver overhead is removed.

what?

[ivan134]

4 minutes ago, Citadelen said:

DX12 reinforces my decision to upgrade to Zen over Polaris, the 290X has massive gains when the driver overhead is removed.

Bruh, Zen is CPU and Polaris is GPU. Unless you meant not needing to upgrade your GPU and just upgrade your CPU. It depends. Current GCN cards are not dx 12_1 compliant.

[Citadelen]

5 minutes ago, ivan134 said:

Bruh, Zen is CPU and Polaris is GPU. Unless you meant not needing to upgrade your GPU and just upgrade your CPU. It depends. Current GCN cards are not dx 12_1 compliant.

I know, that's exactly what I meant.

[Unhelpful]

10 minutes ago, don_svetlio said:

So do we now recommend the Fury X over the 980 Ti? sort of confused

Not yet, still far too early for that. I think this will be more of a factor (or a factor at all) with the next generations of GPUs.

[don_svetlio]

Just now, JoeyDM said:

Not yet, still far too early for that. I think this will be more of a factor (or a factor at all) with the next generations of GPUs.

Fury X went from being 10% behind to 25% ahead - that's a 35% increase

[ivan134]

Oi....I almost feel bad for them, but fuck planned obsolescence.

[Humbug]

11 minutes ago, don_svetlio said:

So do we now recommend the Fury X over the 980 Ti? sort of confused

no. on average the 980ti is a faster GPU.

 

AMD wins most of their performance battles across different price ranges, but not this one.

[Unhelpful]

Just now, don_svetlio said:

Fury X went from being 10% behind to 25% ahead - that's a 35% increase

Yes... In a single benchmark. I'm waiting for actual/more games to recommend Fury X over 980 ti.

[Bensemus]

1 minute ago, ivan134 said:

snip

How does that show planned obsolescence? All those cards a current gen.

[Belgarathian]

This information has been around for a long time, and in initial testing done by anandtech these results stood. Basically GCN architecture has multiple compute capabilities, whereas Nvidia run a very efficient single pipeline so using the GPU to run calculations & rendering workloads at the same time causes bottlenecks in the pipeline for Nvidia cards as they have to wait for a task to complete, and then start the next task.

 

DX12 is wondrous, but you have to harness it and I doubt we'll see that happen for a number of years to come. 

 

Disclosure... This is kind of correct. 

[BonSie]

Guru3d also did the test and they found something off with the FCAT analysis.

 

AKA V-Sync is forced on at the end of the pipeline, which might skewer the results depending on where the measure point is taken in-engine. It's a driver issue.

Quote

Here's the thing, if your GPU is fast enough the numbers the benchmark produces exceed 60 FPS / 60 Hz. The game measures at the rendering port (and not the display end), then is post processed, passed to DX12 and the output rendered. So if at that stage your score can be like 92 FPS. However after that segment in the pipeline a lot of other stuff happens. In our case a 60Hz / FPS framerate . Where we measure with FCAT is definitive though, it's what your eyes will see and observe. 

 

Update: hours before the release of this article we got word back from AMD. They have confirmed our findings. Radeon Software 16.1 / 16.2 does not support a DX12 feature called DirectFlip, which is mandatory and the solver to this specific situation. AMD intends to resolve this issue in a future driver update.

 

 

[ivan134]

1 minute ago, Bensemus said:

How does that show planned obsolescence? All those cards a current gen.

To make Maxwell efficient, they threw out the backseat (hardware schedulers) and went "hey guys, look how much faster and lighter our car is." Lack of hardware schedulers is what is causing this performance disparity. Pascal is not skimping on compute capability and is most likely going to have hardware schedulers, which means people with very capable Maxwell cards will need to upgrade again.

This is the same shit they did with Kepler. Maxwell is better at tessellation so they spam it in GameWorks games which tanks performance and offers no visual improvements over a certain level (imagine running AA at 16x) and refuse to include the ability to change the levels in their drivers. This artificially decreases performance and makes people want to upgrade from their already expensive and very capable graphics cards. If you look at older games where tessellation is applied reasonably, the 970 is only slightly faster than a 780. In most GameWorks games that came out within the last 2 years, a 970 is faster than a 780 ti which used to cost about $700 at launch.

[Humbug]

21 minutes ago, ivan134 said:

Oi....I almost feel bad for them, but fuck planned obsolescence.

Planned obsolescence has nothing to do with this though. They just placed their bets differently.

 

Basically you can think of this as a method of ensuring that the GPU is very well fed with work and the utilization stays at 100%. So it's allowing the GCN hardware to be utilized to it's max.

 

As such asynchronous compute should be viewed as a positive for AMD, rather than a negative for current-gen Nvidia.

[ivan134]

1 minute ago, Humbug said:

Planned obsolescence has nothing to do with this though.

 

Basically you can think of this as a method of ensuring that the GPU is very well fed with work and the utilization stays at 100%. So it's allowing the GCN hardware to be utilized to it's max.

 

As such asynchronous compute should be viewed as a positive for AMD, rather than a negative for Nvidia.

Maxwell cards were advertised as dx 12 ready cards, and Nvidia knew way ahead of time what that would imply. Here's my more detailed response:

9 minutes ago, ivan134 said:

To make Maxwell efficient, they threw out the backseat (hardware schedulers) and went "hey guys, look how much faster and lighter our car is." Lack of hardware schedulers is what is causing this performance disparity. Pascal is not skimping on compute capability and is most likely going to have hardware schedulers, which means people with very capable Maxwell cards will need to upgrade again.

This is the same shit they did with Kepler. Maxwell is better at tessellation so they spam it in GameWorks games which tanks performance and offers no visual improvements over a certain level (imagine running AA at 16x) and refuse to include the ability to change the levels in their drivers. This artificially decreases performance and makes people want to upgrade from their already expensive and very capable graphics cards. If you look at older games where tessellation is applied reasonably, the 970 is only slightly faster than a 780. In most GameWorks games that came out within the last 2 years, a 970 is faster than a 780 ti which used to cost about $700 at launch.

 

[Bouzoo]

 

I like how it says High, Extreme and Crazy detail.

And interesting how NV gets lower FPS with Async enabled.

[Bensemus]

17 minutes ago, ivan134 said:

snip

We still don't have any released games and we don't know how prevalent Async will be. Plus Nvidia loves software solutions so they may be able to close the gap again. 

[ivan134]

Just now, Bensemus said:

We still don't have any released games and we don't know how prevalent Async will be. Plus Nvidia loves software solutions so they may be able to close the gap again. 

It was the same thing in the Fable Legends dx 12 benchmark. Hitman will have Async compute. That's 3 games already with Deus Ex: Mankind Divided likely to follow. Star Citizen will be using it when it gets a dx12 patch.

No, they cannot close this gap with software anymore. When Ashes 1st released, AMD smoked them. 3 months later, Nvidia closed the gap with software. We're back here again and with proper support for Async Compute and this is what happens. Pascal is not going to rely on software solutions because that would be suicide for Nvidia.

[Kinda Bottlenecked]

6 minutes ago, Bensemus said:

We still don't have any released games and we don't know how prevalent Async will be. Plus Nvidia loves software solutions so they may be able to close the gap again. 

 

Software solution may not be very effective in this case and can introduce higher driver overhead. We've seen how larger driver overhead affected amd cards. It's anything but good. 

 

Fucking broken qoute fuck.

[NateGSR117]

Now if only every other game supported DX12 with async 

[Masada02]

Well, looks like I'll be buying some next-gen cards. Someone is gonna get some KPE 980ti's on the cheap, lol.

 

I think what is really going to be remembered from this generation of cards (from both companies) is how well AMD GCN cards held their value, and how much Nvidia cards depreciated. No one is going to want Maxwell cards a year from now meaning ultra-low used card prices.