So, Oxide hasnt optimize the engine for nVidia hardware and now it is nVidia's job to fix this mess?
Yeah, not a great explanation...
Often we get asked about fairness, that is, usually if in regards to treating Nvidia and AMD equally? Are we working closer with one vendor then another? The answer is that we have an open access policy. Our goal is to make our game run as fast as possible on everyones machine, regardless of what hardware our players have.
To this end, we have made our source code available to Microsoft, Nvidia, AMD and Intel for over a year. We have received a huge amount of feedback. For example, when Nvidia noticed that a specific shader was taking a particularly long time on their hardware, they offered an optimized shader that made things faster which we integrated into our code.
We only have two requirements for implementing vendor optimizations: We require that it not be a loss for other hardware implementations, and we require that it doesnt move the engine architecture backward (that is, we are not jeopardizing the future for the present).
No, it is not. It is optimized for AMD hardware.
Otherwise nVidia hardware would be faster with DX12 like they are in King of Wushu or Fable Legends.
No, it is not. It is optimized for AMD hardware.
Otherwise nVidia hardware would be faster with DX12 like they are in King of Wushu or Fable Legends.
It's not about optimizations. It's about Hardware limitations. These limitations affect both nVIDIA and AMD GPUs under Ashes of the Singularity but in different ways.
Explains why Maxwell 2 loses as much performance as Kepler and Maxwell 1 with DX12. :|The difference between Maxwell and Maxwell 2 is that Maxwell's Grid Management Unit can only send either a Graphics task or 32 Compute tasks to the work Distributor. It cannot send both in Parallel.
The Grid Management Unit has 1000s of pending grids waiting to submit to the Work Distributor Unit while the WDU is distributing the grids to the available units...Another difference is that AMDs GCN 1.1 (290 series)/GCN 1.2 have 8 independent Asynchronous Compute Engines each able to schedule and prioritize work independently of one another. With Maxwell 2, it's a single Grid Management Unit. You can see why GCN 1.1 (290 series)/GCN 1.2 can best take advantage of the available compute resources.
Nope. These are draw calls and having nothing to do with asynchronous compute. Star Swarm is using the same aproach and it doesnt show negative scaling.Therefore if there are more than 31 Compute tasks (assuming there is a Graphics task which there ought to be because of the amount of Rasterization going on), it takes two cycles for Maxwell 2 to assign the tasks to the Work Distributor. This looks to be the culprit (explaining why Maxwell 2 tends to match, but not beat, AMDs GCN 1.1/1.2 architecture).
What? It only matches a GTX980TI because DX12 is slower than DX11. The GTX980TI is faster with DX11 than the 290X with DX12.Since Ars Technica showed that a 290x can nearly match a GTX 980 Ti and a GTX 980 Ti is a near match to a Fury-X then we can conclude that the 290x and the Fury-X are a near match under Ashes of the Singularity. This points to a common bottleneck between both Hawaii and Fiji architectures.
DX11 doesnt support multi-engine. This is a new approach in DX12. DX11 can never be faster when they are using multi-engine.The GPU utilization will always be better than it is under DX11. They can do so much more at the same time.Fiji and Hawaii are bottlenecked by their Peak Rasterization rates under Ashes of the Singularity while Maxwell 2 is bottlenecked by its ability to handle Asynchronous Shading.
Care to post the links for those benchmarks ??
and there are posters defending this, shudders.Warner Bros refused AMD optimized code.
Ubisoft REMOVED dx10.1 implementation in AC.
Project Cars Devs think that sharing 20 Steam codes with AMD is doing their part.
GameWork DEVs are fully bought out, very unethical bunch, unlike Oxide.
Here you are, hindsight. What about 3-4 years ago when Maxwell was designed & planned? Would NV know that a Mantle-like API would form the basis of next-gen DX12?
It's a single unit. A single block. A single block is not as Dynamic or Parallel as 8 separate units. This single unit is also limited in the amount of queues it can prioritize when compared to the competition. Maxwell2 is thus not nearly as Parallel, as a matter of fact, as GCN 1.1 (290 series)/GCN 1.2. There is no disputing this. It is a hardware fact.
It is only a big deal if you have games which make use of the amount of Parallelism that Ashes of the Singularity use. nVIDIAs response, to the Ashes benchmark, points to the fact that nVIDIA doesn't believe that Ashes of the Singularity is an overall good example of future DX12 titles. This is likely the same logic on which they based their decisions when building Maxwell2.
Fury-X only has a big bandwidth advantage on paper. In practice, however, nVIDIAs compression algorithms even up the score. See here: http://techreport.com/review/28513/amd-radeon-r9-fury-x-graphics-card-reviewed/4
Since Ars Technica showed that a 290x can nearly match a GTX 980 Ti, and after looking at that memory bandwidth graph, we can tell that memory bandwidth is not what grants the Fury-X its lead.
Since both Fiji and Hawaii retain the same amount of Hardware Rasterizers (and the same Peak Rasterization rate expressed in Gtris/s) we can conclude that both are bottlenecked by their Peak Rasterization rate (ability to draw triangles/polygons)
No, it is not. It is optimized for AMD hardware.
Otherwise nVidia hardware would be faster with DX12 like they are in King of Wushu or Fable Legends.
So, nVidia has less hardware limitations with DX11 than with DX12? :hmm:
Explains why Maxwell 2 loses as much performance as Kepler and Maxwell 1 with DX12. :|
The Grid Management Unit has 1000s of pending grids waiting to submit to the Work Distributor Unit while the WDU is distributing the grids to the available units...
I guess you didnt really understand Hyper-Q.
Nope. These are draw calls and having nothing to do with asynchronous compute. Star Swarm is using the same aproach and it doesnt show negative scaling.
What? It only matches a GTX980TI because DX12 is slower than DX11. The GTX980TI is faster with DX11 than the 290X with DX12.
DX11 doesnt support multi-engine. This is a new approach in DX12. DX11 can never be faster when they are using multi-engine.The GPU utilization will always be better than it is under DX11. They can do so much more at the same time.
In DX12 it is possible to use as many engines as it is supported by the hardware. A hardware with 32 supported compute queues will not be slower than DX11 with only one queue. A GTX770 isnt slower than a GTX960 in the benchmark so the performance impact of DX12 over DX11 is not result by using multi-engine.
This is purely semantics. It's like arguing which is more parallel, a quad core CPU or a dual core CPU with hyperthreading..
The GMU is a single unit, but since it doesn't execute anything, it doesn't really matter. The execution is done by the SMM units, and those are all expressly parallel.
If NVidia were as crippled with asynchronous compute as you claim, I believe that the performance deficit would be much larger. But that's not what we see when we look at the benchmarks.
AMD has a lead yes, but it's not huge or anything. In the PCper review, the 390x has at most a 10.6% lead over the GTX 980:
The compression algorithm doesn't take MSAA into account. MSAA in and of itself has a massive impact on bandwidth..
If you looked at the benchmark you posted, you'll see that the medium preset was run with MSAA and the high without. In the medium preset, the Fury X is 12% faster than the 980 Ti in DX12. On the high preset with MSAA turned off, the Fury X has a 7% lead.
If you factor in that the 980 Ti is somehow faster in DX11 than in DX12, then the Fury X's lead is even smaller.
I'm skeptical of the Ars Technica benchmarks, as they are the only ones showing such to my knowledge.
It's extremely unlikely that Fiji and Hawaii are bottlenecked by rasterization. I don't think I've ever seen any modern GPU being bottlenecked by it's rasterizers in an actual game, only in synthetic benchmarks..
But if AMD rasterizer will be bad than why AMD is leading the Overhead test of the 3dmark?
3dmark overhead test is also a pure rasterizer test with a huge polygoneoutput !!!
- AnandtechTo do this, Futuremark has written a relatively simple test that draws out a very simple scene with an ever-increasing number of objects in order to measure how many draw calls a system can handle before it becomes saturated. As expected for a synthetic test, the underlying rendering task is very simple – render an immense amount of building-like objections at both the top and bottom of the screen – and the bottleneck is in processing the draw calls. Generally speaking, under this test you should either be limited by the number of draw calls you can generate (CPU limited) or limited by the number of draw calls you can consume (GPU’s command processor limited), and not the GPU’s actual rendering capabilities. The end result is that the API Overhead Feature Test can push an even larger number of draw calls than Star Swarm could.
No, it is not. It is optimized for AMD hardware.
Otherwise nVidia hardware would be faster with DX12 like they are in King of Wushu or Fable Legends.
It does matter. You have a Multi-Core (Parallel) CPU feeding 8 Asynchronous Compute Engines independently vs a Multi-Core (Parallel) CPU feeding a single unit. That's on one end.
On the other you have 8 Asynchronous Compute Engines prioritizing work and sending the work off, independently, to the available compute resources (Out of Order) vs. a single unit feeding a work distributor (extra level of latency) which then assigns the work to the available compute resources (In Order).
That's a pretty big difference.
It doesn't need to be a huge lead and I don't expect AMD to have a huge lead. I'm not rooting for AMD. I'm objective. AMD lacks the Rasterization rate needed to draw all the triangles which make up the swarm of units on the screen.
More units are drawn in the heavy preset vs the medium preset (peak rasterization rate coming into play). You would have to bench the medium preset with and without MSAA to draw any conclusions regarding memory banwidth but I'm not sure what you're trying to say here regardless. You won't draw any logical conclusions from comparing apples to oranges.
Nobody else benched a 290x. PCPer benched a 390x but did not bench a GTX 980 Ti. Therefore we cannot derive any conclusions from that.
Have you even seen a game like Ashes of the Singularity? Nope... no game has that many drawn simultaneous units on the screen.
Let's look at this differently. The entire purpose of the GMU from what I've read, is to attempt to utilize the GPU resources as much as possible in an efficient manner.
So for your assertion to be true, you would have to believe that NVidia's engineers are so incompetent that they would stick with a single GMU through two architectures despite your notion that a single GMU obviously isn't enough to keep the GPU busy..
But when you look at the real world evidence, it becomes clear that NVidia's Maxwell architecture is much more efficient than AMD's GCN, as it's able to do more with less basically.
In fact, it's AMD's GCN architecture which has difficulty utilizing the shaders effectively, not Maxwell.. So in light of this, I still don't think your argument has any merit.
What do you think about Digidi's comment that AMD leads in the 3D Mark overhead test?
PCper and Computerbase.de used a 390x and a 390 respectively, both of which are either faster or equal to the 290x.
In the PCPer review, the 390x was competing against a GTX 980 and whilst the 980 was a bit slower, it was far from being trounced.
In the computerbase.de review, the GTX 980 Ti faced a Fury X, which is much faster than a 290x, yet the performance was very similar between the two..
I don't think Ashes of the Singularity is unique when it comes to the amount of objects drawn.. What it is unique in doing, is how they are drawn. The objects or units are all likely unique in some manner and require separate draw calls like Star Swarm..
The difference between Maxwell and Maxwell 2 is that Maxwell's Grid Management Unit can only send either a Graphics task or 32 Compute tasks to the work Distributor. It cannot send both in Parallel.
Therefore you're correct in pointing out that with Maxwell 2 the communication between the Grid Management Unit and Work Distributor is now Parallel.
The problem is that this doesn't change the fact that Maxwell 2 still only contains a single Grid Management Unit. This still remains as a bottleneck.
nVIDIAs Parallelism, under Maxwell 2, is thus limited to 1 Graphics and 31 Compute tasks. AMDs Parallelism, under GCN 1.1 (290 series) and GCN 1.2 is limited to 1 Graphics and 64 Compute tasks.
Another difference is that AMDs GCN 1.1 (290 series)/GCN 1.2 have 8 independent Asynchronous Compute Engines each able to schedule and prioritize work independently of one another. With Maxwell 2, it's a single Grid Management Unit. You can see why GCN 1.1 (290 series)/GCN 1.2 can best take advantage of the available compute resources.
Let's look at this differently. The entire purpose of the GMU from what I've read, is to attempt to utilize the GPU resources as much as possible in an efficient manner.
So for your assertion to be true, you would have to believe that NVidia's engineers are so incompetent that they would stick with a single GMU through two architectures despite your notion that a single GMU obviously isn't enough to keep the GPU busy..
But when you look at the real world evidence, it becomes clear that NVidia's Maxwell architecture is much more efficient than AMD's GCN, as it's able to do more with less basically.
In fact, it's AMD's GCN architecture which has difficulty utilizing the shaders effectively, not Maxwell.. So in light of this, I still don't think your argument has any merit.