Poor DX 12 performance

highly irrelevant since nvidia's software scheduling outperforms amd's hardware one in nearly all cases,

hence amd's need for async compute to even work the hardware properly.

 

Did you even read the context of my post and the question I was answering? Apparently not. Nobody said anything about AMD's hardware instruction scheduler outperforming NVidia's software scheduler or vice versa, so I don't even know why you brought that up.

 

Nope. Pretty much nothing has changed in AMD's DX11 performance during the last two years. It still goes down like a rock when there's a CPU limitation on the driver thread. There are many recent benchmarks which still show this in all its glory.

Both are using software compilation for instructions and the only thing which AMD does in h/w while NV doesn't is instruction stream reordering. This has nothing to do with any API as it's about h/w's innards and this also has nothing to do with the driver. The simple fact is that NV is able to get a lot more performance from same flops while consuming less power at the same time which means that their solution here is just plain superior.

 

No, I just go by my hands-on performance experience vs reading online performance charts. For example, DA:I DX11 vs Mantle. It used to be that Mantle outperformed DX11, esp in regards to minimal FPS. A few months into Catalyst development, and continued into Adrenilin, I can now use either API and get very similar performance. In fact I prefer DX11 because load times are quicker. (All this on a R9 390) And the card performs on par with the GTX970 in DX11 games, which would be the comparable matching hardware.

 

I remember having very similar DX11 <-> Mantle performance in DA:I back in the days with a 290X + 2500K@4.4GHz (1080p60 display with Vsync), especially after the last few game patches. I bought the game around update #8 (so didn't test earlier versions) and I remember being unable to replicate the early benchmark results (as seen in zero-day reviews). And DX11 just got better with the last few game patches and driver releases as I played along (the last DLCs came out later, so it spanned several months).
I had the feeling they stopped caring about the Mantle mode early on (they only tested the new changes thouroughly with DX11 and/or fixed DX11-specific issues but not Mantle-specific ones).

I always preferred DX11 because I experienced a lot more random graphical glitches with Mantle, like completely missing visuals for some magical effects (and, you know the saying: "anyone can be infinitely fast if they are allowed to be infinitely inaccurate").
If that wasn't enough, I remember AMD drivers missing the tripple-buffer option for Vsync with Mantle under Win10 for a long time (if not forever?).
And then the loading times... despite Mantle having it's own shader cache and AMD drivers not having a DX11 shader cache back then...
-> all these issues for a max of ~5% extra on the average fps (I remember finding some places/scenes where Mantle actually produced slightly lower fps, so you cound't even say that "it's at least as good but with occasional extra" because even that was a little erratic).

 

Not sure what you two are talking about since this was launch day DX vs Mantle performance.

 

AMD's Mantle performance was superior to its DX11 performance for as long as you were CPU bottlenecked.

 

That's because you didn't use the hardware and play the game for over a year, and get to know the performance changes intimately. Relying on benchmarks posted online does not provide the insight. That bench doesn't account for engine patches, driver optimizations, fixes, and such.

 

Considering I was using an i7 950 @ 4GHz on X58, I would say it was a possibility.

 

and now you're trying to change the subject, now i know why most pay you no attention.

 

What can I say? There's no accounting for illiteracy. If you can't read and understand basic English, why are you posting in a tech forum?

 

OK I thought that AMD used a hardware instruction scheduler for everything. I know that NVidia's software instruction scheduler is credited with reducing the power consumption significantly, since the days of Kepler.

 

Sorry if I won't hunt it down (it's possibly not even online anymore) but I remember some German site had somewhat different results (may be they used a weaker CPU though). And the game went through a lot of updates. I think Mantle was improved in the beginning (first few updates) but then regressed at the end (last few updates). I guess Mantle was fresh and thus shipped unpolished (especially in DAI), got improved for a while (mostly because a Battlefield title used it too while AMD was still interested in the Mantle marketing) but then silently got thrown out of focus (even by AMD, let alone EA/Dice) and was never worked on again by Dice (it was fully dropped from later Frostbite games) while the Mantle games (DA and BF) still received some updates (but probably with the Mantle mode completely out of focus by then, so they didn't even care if the Mantle mode regressed by changes made with the DX11 mode in focus).

I did. And I used DX11 for >90% of my gameplay hours.

Truth is, we were later told (long after DAI, sometimes around 2017 if I recall, after seeing the first few DX12 games and their benchmarks even after several "foretold to be magical" updates) how game engines have to be almost fully redesigned at a much deeper level (instead of of just ported to a DX12 presenter and tweaked a bit) to really bring out the possible advantages of the new explicit APIs. And, as I see it, we are still waiting to see that premise proved or debunked (most DX12/Vulkan games today are still either "hybrid" and/or "handicapped" in one way or another to be used as etalons). Actually, if that premise is true, then we will never really learn the full truth, since no game will have an equally good DX11 and DX12/Vulkan mode (one of the modes will be handicapped on principle or simply won't be included at all to compare).
And truth be told, as it often happens with these "promised to be better on the really long run" things, the 290X will be hopelessly weak for those "true DX12" games anyways, no matter the APIs and optimizations. Both AMD and nVidia will probably have some new GPUs out on the market by then (so even Vega and Pascal will be replaced, let alone Hawaii and Maxwell which will be heading into obsolescence when the "latest and grates" games with ~full quality are concerned).

 

Another "dx12 is slow on Nvidia" thread? Just use dx11 mate and call it a day.

 

AMD do have drivers don't they? And last time I checked their GPUs weren't able to execute DXBC or SPIR-V bytecode which means that they need to compile the code in software for it to run on their ISA.

There is no "software scheduler". All instruction scheduling - as in what is being executed where on the GPU - is done in h/w on both NV and AMD. The difference lies in what this h/w is able to do with the instruction stream. GCN is able to reorder it in h/w while Kepler/Maxwell/Pascal/Volta run it exactly as provided by the driver side compiler. NV had this ability in Tesla/Fermi too but they've removed it starting with Kepler since the power losses weren't compensated by the increased h/w utilization - which is demonstrated by the comparative power efficiency advantage of Pascal over Polaris/Vega although it's not the only or even main source of such advantage.

Generally speaking, you don't even need this ability as your compiler obviously knows your h/w and can produce highly optimized code for it in which there will be nothing to optimize further on the h/w itself. And since GPU is a highly parallel machine there's always an ability to cover a stall or bubble of one instruction stream with instructions from other thread making this whole "h/w scheduling" kinda unnecessary.

 

AMD is weak in DX11.