AMD GPUs See Lesser Performance Drop on "Deus Ex: Mankind Divided" DirectX 12

[RejZoR]

And I own a GTX 980 this moment... The fact is, AMD and NVIDIA are about the same in DX11, trading blows with exclusive games. But with DX12, AMD has a clear dominance which NVIDIA fanboys refuse to admit.

 

[HD64G]

And let's not all forget. As some have already said, this is one game. And it's an AMD game. It is without any doubt playing to AMD's strengths, just as Nvidia titles play to Nvidia strengths. What is good is that even on Nvidia hardware, it's playing very well. The lower fps (for some cards) isn't a deal breaker, the game is stutter free.
The only kicker is the memory load. I'm playing close to the 6Gb number and that's my cards limit.


Edit: Found this on Google Now.

http://techreport.com/review/30639/...x-12-performance-in-deus-ex-mankind-divided/3

Seems the FPS numbers in DX12 mask an AMD issue after all.

Let's wait for it to be tested elsewhere. They messed up a previous review with their system's latency issues making AMD GPUs work worse than they should and they apologised for it already. And this test you posted doesn't show AMG GPUs gain on DX12 as almost every review clearly shows after all. Fury X losing more FPS than 1070 when turning to DX12 and 1060 keeping the same distance from 480 for the same situation? Something isn't correct there.

 

[RejZoR]

AMD beats NVIDIA in every single DX12 game (except games where both suck hard compared to DX11 for no logical reason). Must be "AMD biased games". Right. It's not because the rendering engine in Radeons is clearly superior for such tasks since HD7000 series when they introduced GCN, it has to be "bias". C'mon people, can you be less of a fanboys?

One thing is GameWorks which is NVIDIA's proprietary tech which is known to gimp AMD cards and another is standardized API like Vulkan or DX12 that everyone can use with same capacity. How is AMD playing dirty here? And don't even try bringing Mantle up, because without its idea, there would be no Vulkan or DX12 in existence. Which doesn't mean it's based on Mantle, it's based on its core idea. And while NVIDIA at least showed a bit of improvement with GTX 1080, seeing how 1 generation older Fury X with half the memory is still beating it or is equally as fast is telling a lot...

 

[the54thvoid]

Let's wait for it to be tested elsewhere. They messed up a previous review with their system's latency issues making AMD GPUs work worse than they should and they apologised for it already. And this test you posted doesn't show AMG GPUs gain on DX12 as almost every review clearly shows after all. Fury X losing more FPS than 1070 when turning to DX12 and 1060 keeping the same distance from 480 for the same situation? Something isn't correct there.

I think this is an ongoing problem, there are varying results from diff sites. I'm under no illusions that AMD generally see a respectable performance uplift in DX12 and Nvidia sees very little, if any.
What I'm interested in is the frame time graphs which I've seen replicated elsewhere in other sites. I was stung years back in DX9 era crossfire with great apparent fps but a bad visual show. It looked very much like Hawaii fixed that with XDMA. Also moving on with Fiji in DX11, even single card latency looked so smooth.
But recently, DX12 has thrown up quite a few instances where AMD has a much higher latency than Nvidia.
I don't point fingers at AMD, I think devs are being really ignorant of their tasks and perhaps simply expecting the hardware to 'work' with the API without much effort on their end.
Like baking a cake with all the right ingredients but rushing the mixing. Cooking too long etc. Just because the 'components' are present, doesn't mean the end result will be right.

 

[ZeDestructor]

Curious how in Vulkan nVidia also has performance gains, but on D3D12 (why do people keep saying DX12 ????), is the opposite...

That's the difference in developer quality and game engine design. As can be evidenced from MD and DOOM, it looks like Eidos, Eidos Montreal and Nixxes are evidently not as good as id at using the new low-level APIs. For both GPU vendors.

Just to sate the question of DX12 vs Vulkan, here's the old (rather AMD-biased cause it originally started out as a Mantle tech demo)Star Swarm benchmark showing healthy gains on nV hardware to match the healthy gains nV gets on DOOM. Not as big of a gain as the GCN cards, but then again, the accepted theory is that GCN is simply particularly bad at DX11-/OpenGL for those of use who eat live and breathe microarchitectural CPU/GPU designs.

AMD beats NVIDIA in every single DX12 game (except games where both suck hard compared to DX11 for no logical reason). Must be "AMD biased games". Right. It's not because the rendering engine in Radeons is clearly superior for such tasks since HD7000 series when they introduced GCN, it has to be "bias". C'mon people, can you be less of a fanboys?

Just because AMD is particularly crap at DX11 doesn't make nV particularly bad at DX12. In terms of FPS per W for equal generations, AMD and nV are actually quite close for DX12 loads, which these days is a much, much better indicator of chip design efficiency than pretty much anything else.

One thing is GameWorks which is NVIDIA's proprietary tech which is known to gimp AMD cards and another is standardized API like Vulkan or DX12 that everyone can use with same capacity. How is AMD playing dirty here? And don't even try bringing Mantle up, because without its idea, there would be no Vulkan or DX12 in existence. Which doesn't mean it's based on Mantle, it's based on its core idea. And while NVIDIA at least showed a bit of improvement with GTX 1080, seeing how 1 generation older Fury X with half the memory is still beating it or is equally as fast is telling a lot...

nV just has a nice name for their gamedev support. AMD could have done the same, and put in a lot more effort into playing their strengths (TressFX in Tomb Raider 2013 anyone?), but they didn't, and for that, they have to suck up the performance losses when nV essentially offers free extra dev time to game devs.

As for Mantle/Vulkan, most of that credit goes to DICE, and later on Khronos and Valve for Vulkan (which basically came about when AMD gave up on Mantle and just handed the entire thing to Khronos, the OpenGL consortium), not AMD. If you read deep enough and between enough lines, it's obvious that AMD provided only as much support as needed for the guys at DICE to build the API up into something workable.

 

[Shambles1980]

Software driver based a-synch compute is retarded. And they need to just admit they didnt do it right and move on to hardware a-synch..
if amd had bothered to make a better gpu than the 480 then nvidia would be surviving 100% on fanboys.
Right now though they have the time to decided to go hardware a-synch before amd release a decent gpu.

 

[Dethroy]

These discussions always end up going round in circles.

Having said that, I'm neither pleased with Nvidia nor with AMD this round. Another generation I'm skipping. My GTX 670 is still serving me well and the current generation of cards seem only like a stop-gap solution to me.

 

[RejZoR]

In tests AMD is like 3fps behind NVIDIA in DX11 (except in very NVIDIA centered games) and everyone is losing their shit on how AMD is "much" worse than NVIDIA. And then we have DX12 titles where AMD takes a lead by 20fps and no one can be bothered because reasons. C'mon, who are you trying to fool?

 

[ZeDestructor]

Software driver based a-synch compute is retarded. And they need to just admit they didnt do it right and move on to hardware a-synch..
if amd had bothered to make a better gpu than the 480 then nvidia would be surviving 100% on fanboys.
Right now though they have the time to decided to go hardware a-synch before amd release a decent gpu.

Kepler (well, big Kepler/GK110 at any rate), Maxwell and Pascal all have hardware-based schedulers. Please stop spouting bullshit

These discussions always end up going round in circles.

Having said that, I'm neither pleased with Nvidia nor with AMD this round. Another generation I'm skipping. My GTX 670 is still serving me well and the current generation of cards seem only like a stop-gap solution to me.

Heh.. so it does, and filled with misinformation too..

I need an upgrade on my end, even though I want to wait for HBM-equipped nV cards thanks to GTX670 No. 2 is well on it's way out :/.

In tests AMD is like 3fps behind NVIDIA in DX11 (except in very NVIDIA centered games) and everyone is losing their shit on how AMD is "much" worse than NVIDIA. And then we have DX12 titles where AMD takes a lead by 20fps and no one can be bothered because reasons. C'mon, who are you trying to fool?

I never said that anything about who's winning, just that pointing to Gameworks nerfing AMD cards or massive DX12 performance improvements for AMD side means very little besides telling us that AMD cards are less fast under certain workloads.

 

[RejZoR]

"Certain workloads" aka proprietary effects specifically tailored to take maximum advantage of one hardware and giving NO option for the competitor to utilize it or utilize it efficiently. But hey, what do I know...

 

[Shambles1980]

How can you take amd being 3fps slower than nvidia in driver based dx 11 vs 20fps faster in dx12 as "amd being slower in certain work loads"
When you could easily have said "nvidia is a lot slower in certain work loads"
thats just looking at the situation form a very one sided point.

and you cant claim nvidia are hardware a-sync when the drivers 100% controll it.
And the whole point of dx12 was to just avoid the drivers.

 

[ZeDestructor]

"Certain workloads" aka proprietary effects specifically tailored to take maximum advantage of one hardware and giving NO option for the competitor to utilize it or utilize it efficiently. But hey, what do I know...

How can you take amd being 3fps slower than nvidia in driver based dx 11 vs 20fps faster in dx12 as "amd being slower in certain work loads"
When you could easily have said "nvidia is a lot slower in certain work loads"
thats just looking at the situation form a very one sided point.

and you cant claim nvidia are hardware a-sync when the drivers 100% controll it.
And the whole point of dx12 was to just avoid the drivers.

Because that is exactly what "certain workloads" means. Overall, they are quite well-matched.

To give another, not nV/AMD comparison, but rather a present vs past comparison, AVX code does not run with a processor that doesn't support AVX, instead you fallback to SSE or even just base x86 (depends on the dev's choices) to do it at half the speed. It's certain workloads and certain chips are twice as fast at it. Yes, you can file this as just general evolution, but at the time, it was a real difference that you had to care about as a developer, much like you would for gameworks effects.

Pascal is fully hardware-based. Maxwell 2 and previous were not (the hardware scheduler there disabled for mixed-used scenarios, enabled for HPC) because the scheduling was static and with static scheduling you can easily end up with worse results that driver-level scheduling. All this is written about at length at the link.

PS: before you say AVX loads don't matter in games, Path of Exile uses AVX for particle effects calculations for an almost doubling of framerates.

 

[Shambles1980]

i wasnt stating that you were incorrect in your statment, i was saying that you chose to state it in the manner that portrayed nvidia in the best light.

 

[RejZoR]

The AVX comparison is silly at best. That's like saying if your graphic card doesn't support DX11, it'll fall back to DX9. It's what that is. It's not Intel's AVX is better than AMD's AVX...

And why on Earth would you use CPU to simulate a graphic effect like particles. It's just beyond idiotic. Lets make particles being calculated on HDD controller, because dumb reasons. LOOOOL We're in era where GPU's are being used more and more for even non graphic features and here they stick particles to a CPU. It just makes zero sense.

 

[ZeDestructor]

i wasnt stating that you were incorrect in your statment, i was saying that you chose to state it in the manner that portrayed nvidia in the best light.

I'm a bit biased towards nV because I run Linux on my machines as well as Windows, and honestly, I'm a little bit tired of people fawning all over AMD as though they're being everyone punching bags. Newsflash: AMD dug themselves into this hole their in while their competition gets to dictate the market and give free extremely well-optimized code with free devtime to everyone, possibly as a direct result of the sheer amount of revenue they get allowing them to do so.

The AVX comparison is silly at best. That's like saying if your graphic card doesn't support DX11, it'll fall back to DX9. It's what that is. It's not Intel's AVX is better than AMD's AVX...

Actually, CPU architecture implementation matters even for literally identical CPU code. It's literally why IPC improvements are a thing. And yes, Intel x86 (and extensions) is better than AMD x86 (and extensions) right now. Literally. Compare Intel's Sandy Bridge to AMD's Bulldozer running literally identical code - the SNB chip is faster at a lower clockspeed ffs!!!

And why on Earth would you use CPU to simulate a graphic effect like particles. It's just beyond idiotic. Lets make particles being calculated on HDD controller, because dumb reasons. LOOOOL We're in era where GPU's are being used more and more for even non graphic features and here they stick particles to a CPU. It just makes zero sense.

You still need to simulate the underlying physics after the event has happened before the GPU gets to render it. Maybe async compute would be better, maybe it would not be (there's plenty of things that CPUs run a lot better than a GPU, and besides, PoE is still a 32-bit D3D9 game with 64-bit D3D11 builds scheduled for 2017). PhysX would almost certainly be better.. but wait! That's unfair and anticompetitive cause nV would run better and at higher fidelity.

 

[Shambles1980]

Only because nvidia stoped people using nvidia cards as a dedicated physx card with amd main card.. Which is the definition of anticompetative.

 

[RejZoR]

You're comparing architecture advantages with extensions. I don't think you actually understand what you're talking about. Architecture is standardized to a point where it can be called x86 or not. Extensions are exact and strict stuff, you can't bend things around. You can just tailor architecture to be more efficient with given extensions, but you can't make them different and still have them the same.

As for physics, it has already been established that GPU's are superior in that regard. It's why fluid dynamics are a reality on GPU's and just not usable on CPU's.

 

[Fluffmeister]

http://techreport.com/review/30639/...x-12-performance-in-deus-ex-mankind-divided/3

Seems the FPS numbers in DX12 mask an AMD issue after all.

Like the 7950 is faster, but the GTX 660 TI is still smoother? Doh!

 

[Italia1]

Hi to all, i'm new here, i want ask: where are dx 12 "admirable" gain promised some year ago ?

 

[ZeDestructor]

Only because nvidia stoped people using nvidia cards as a dedicated physx card with amd main card.. Which is the definition of anticompetative.

I personally don't mind. If one end has the better product, I'll pay for it (within limits), and for now, PhysX is well within said limits at this point in time.

You're comparing architecture advantages with extensions. I don't think you actually understand what you're talking about. Architecture is standardized to a point where it can be called x86 or not. Extensions are exact and strict stuff, you can't bend things around. You can just tailor architecture to be more efficient with given extensions, but you can't make them different and still have them the same.

The ISA is standard, the architecture is not. In terms of just x86, and just Intel, we currently have 4, completely independent, fully x86 (with extensions) ISA-compliant architectures with staggeringly large differences in performance: Silvermont/Goldmont Atom chipsets, Broadwell (under Broadwell-EX/-E/Broadwell-D), Skylake (under Skylake-H/-S/-U/-Y) and Knights Landing (Xeon Phi, loosely Goldmont-based). All of them will run up to at least SSE4 code, and all of them will run the exact same code with vastly different performance results.

Probably the most relevant comparison to illustrate is the Broadwell-E vs KNL comparison. BDW-E is your usual general purpose chip, at at the full ~650mm² has only 24 cores, while KNL has 72 cores in the same ~650mm² die size at the same 14nm node. In performance terms, the BDW chip destroys the KNL chip for the typical CPU loads (databases, general workstation, web servers, games, etc), but the moment you point optimize, highly-thread AVX code at them, the KNL chip will simply be magnitudes faster (GFLOPS vs TFLOPS) than the poor Broadwell chip. In conclusion, they are very different core designs, targeted at very different applications.

If you want an ARM example, just look at the huge range of ARMv6, v7 and v8 implementations out there. Any code targeted for ARMvL will work on any implementation that implements the ARMvL ISA, so ARMv6 code will run on any ARMv6 chip, etc upto ARMv8.

What that means, is that code targeting the base ARMv8 ISA will run on everything from a relatively massive Qualcomm Kryo or Apple Twister implementation all the way down to the tiny, super-low-power Cortex-A35. The big chips will be 3+ times faster than the little A35.

As for physics, it has already been established that GPU's are superior in that regard. It's why fluid dynamics are a reality on GPU's and just not usable on CPU's.

For games, you have the interesting proposition of also needing to run the rest of rendering on the GPU (polygons, texturing, filtering, occlusion), so it may be more effective to do the work on the CPU. Eitherways, this is not relevant to the discussion that the AMD/nV perf difference lying pretty much entirely in the different architectures. I only brought up that particular example because I knew that the first thing that would come from using AVX as an example would be that no games used it. Path uses AVX, and uses it to great effect. How wise/optimal that decision is, I don't really give a shit since my framerates have quadrupled since the last major patch.

 

[Octopuss]

That's like saying the new asbestos is giving you less cancer than the old one.

ROFLMAO!!!!!!!

 

[R-T-B]

And I own a GTX 980 this moment... The fact is, AMD and NVIDIA are about the same in DX11, trading blows with exclusive games.

Not when you consider what the hardware SHOULD be doing given it's specs and bus width/bandwidth, which is about what it does what it does on DX12.

AMD beats NVIDIA in every single DX12 game (except games where both suck hard compared to DX11 for no logical reason). Must be "AMD biased games". Right. It's not because the rendering engine in Radeons is clearly superior for such tasks since HD7000 series when they introduced GCN, it has to be "bias". C'mon people, can you be less of a fanboys?

I'm really not being a fanboy. I'm just looking at it as logically as I can. But heck, maybe even I am wrong. I'm hardly a GPU engineer.

 

[Vayra86]

And I own a GTX 980 this moment... The fact is, AMD and NVIDIA are about the same in DX11, trading blows with exclusive games. But with DX12, AMD has a clear dominance which NVIDIA fanboys refuse to admit.

I wouldn't go so far as to say dominance, it's just that AMD is finally getting their money's worth out of their 'metal'.

- AMD uses a wider bus
- AMD uses more shaders
- AMD runs at lower clocks
- Polaris provides about similar (or slightly higher) perf/clock to Pascal
- Polaris still has a lower perf/watt than Pascal
- GCN has not radically changed since HD7xxx.

AMD just runs a wider GPU across the board, as they have done for a long time. GCN is geared to be an extremely balanced arch that has some overcapacity on the VRAM end. It is built to let the core do all the work it can do, whereas Nvidia's arch is always focused at 'efficiency gains through tight GPU balance' - Nvidia obtains that balance by cutting heavily into bus width and removing everything from the GPU core that isn't required for gaming. They've tried several things, of which DP was the first thing they dropped with Kepler, then delta compression enabled them to further reduce bus width. This is also why Nvidia's cards don't stretch their legs at higher resolutions, but rather lose performance. Only the GDDR5X-supported 1080 avoids that fate.

On DX11, AMD GPU's were just fine and they excelled only at higher resolutions. Why? Not just because of VRAM, but because of the fact that higher res = lower CPU load. In DX12, GCN gets to stretch its legs even earlier and also at lower resolutions, in part also because of the better CPU usage of that API. Vulkan is similar. That CPU usage was the last hurdle for GCN to really come to fruition. Say what you want, but AMD has really made a smart move here, even though we can doubt how conscious that move has really been. They have effectively gained architectural advantage by letting the market do most of the work.

The irony is that the market for gaming has moved towards GCN, and GCN has seen very minimal architectural changes, while the market is moving away from Nvidia's cost/efficiency improvement-focused GPU architecture. At the same time, Nvidia can almost eclipse that change through a much higher perf/watt, but that only hides so much of the underlying issue, an issue of Nvidia GPU's having to clock really high to gain solid performance, because they lack not only a wide bus right now, but also raw shader counts.

I think it is inevitable, and safe to predict, that Nvidia has now reached a new cap with regards to clock speeds on the core. The only way forward is for them to once again start building bigger and wider GPUs. AMD, on the flip side, has more wiggle room and a lot of things left to improve - clocks, efficiency, and judging the RX480, they also have space left on the die.

 

[jabbadap]

The AVX comparison is silly at best. That's like saying if your graphic card doesn't support DX11, it'll fall back to DX9. It's what that is. It's not Intel's AVX is better than AMD's AVX...

And why on Earth would you use CPU to simulate a graphic effect like particles. It's just beyond idiotic. Lets make particles being calculated on HDD controller, because dumb reasons. LOOOOL We're in era where GPU's are being used more and more for even non graphic features and here they stick particles to a CPU. It just makes zero sense.

Well it depends, if you use some old 128bit avx code(sse relic) then amds avx can be little faster. But all newer code uses mostly 256bit avx2, where intel implementation is far superior than amds. Amds implementation split 256bit code to two 128bit, while intel has 256bit wide simd.

Well cpu only havok FX has particle simulation, I don't really know does it uses avx2 or is it even practical to use SIMD for particle effects. Even PhysX software(cpu) uses avx instructions on some effects(cloth).

 

[RejZoR]

For games, you have the interesting proposition of also needing to run the rest of rendering on the GPU (polygons, texturing, filtering, occlusion), so it may be more effective to do the work on the CPU. Eitherways, this is not relevant to the discussion that the AMD/nV perf difference lying pretty much entirely in the different architectures. I only brought up that particular example because I knew that the first thing that would come from using AVX as an example would be that no games used it. Path uses AVX, and uses it to great effect. How wise/optimal that decision is, I don't really give a shit since my framerates have quadrupled since the last major patch.

Particles are actually just a bunch of tiny polygons with texture attached to them. Why wouldn't you run them on GPU? Especially since we have specialized features like Geometric Instancing to handle just that, hundreds of identical elements.