DX12 Performance Discussion And Analysis Thread

Or perhaps you should. Did you miss the part where they said the game crashes all the time?

An Alpha release of a game suffers from crashes?
cNT8XxV.jpg





Did you see the terrible scaling?

Here's your definition of terrible scaling:
7g6iOsG.png
 
I think we can do without meme images. ;-)

Well the Alpha only supports AFR for now, so a weak iGPU may not make that much of a difference if you have a medium-range desktop card or better.

I meant in the general sense (AFR sucks anyways; it's actually disappointing they spent anytime at all on it, but that's a different discussion). What if multi-adapter support becomes the norm? Even if the cross-vendor stuff doesn't pan out I still think it'll change the equation on which cpu/apu gamers should buy. Suddenly that 300w apu might make sense even if you plan on using a discrete gpu.

You could even conceive a case that buying the second fastest cpu (zen apu) and the second fastest dgpu (amd dgpu) is faster than the fastest cpu (intel) and fastest dgpu (nvidia gpu). :D
 
Unreal Engine Explicit Multiadapter:
https://channel9.msdn.com/Blogs/DirectX-Developer-Blog/DirectX-12-Multiadapter-Unreal-Engine-4

This benchmark uses Intel iGPU to do the the post processing (NVIDIA discrete GPU renders everything else). No AFR, so no micro-stuttering or other problems. Pretty much everybody has Intel iGPU in their CPUs right now. Getting that 11% perf boost is a nice addition.
And I assume much less stable CPU overclocking and turbo.
Old card + new is much more interesting combination.
 
You might find that CPU overclocking becomes far less of a 'requirement' with lower overhead API's anyway. Real world/game proof of that still needs to be proven as well as what games will do with the higher draw call ceiling in time, but it does look like CPU clock speeds will be less important than they have been.
 
Scaling appears to be bad, well bad is not the right term, but not as good on NVidia as AMD cards. So maybe that's why people think it has bad scaling.

GM200 salvage + GM200 = 46% better than GM200 salvage.
Fiji salvage + Fiji = 66% better than Fiji salvage.

It is interesting that Fiji primary + GM200 secondary nets better results than GM200 primary + Fiji secondary.

I wonder if NVidia will be able to address that in drivers, or if it's an architectural deficiency compared to GCN preventing at least similar scaling.

Regards,
SB
 
Unreal Engine Explicit Multiadapter:
https://channel9.msdn.com/Blogs/DirectX-Developer-Blog/DirectX-12-Multiadapter-Unreal-Engine-4

This benchmark uses Intel iGPU to do the the post processing (NVIDIA discrete GPU renders everything else). No AFR, so no micro-stuttering or other problems. Pretty much everybody has Intel iGPU in their CPUs right now. Getting that 11% perf boost is a nice addition.

Shure... if it´s a 4400 or above. But many people have 4000, not supporting DX 12!

My i7 4770k has a 4600, so... I´m good! But my laptop iGPU is a HD 4000! :(
 
The feature is not even Alpha yet so I wouldn't read much into it.
The problem is that the feature is never likely to leave Alpha stage. Too complicated, requires too much from game developers and GPU driver writers, for too little gain.

It'll probably work for doing post processing on iGPUs, but AMD+NVIDIA? This is likely to stay an Alpha feature.
 
I don't think AFR is worth investigating at all.

It would be exciting if they could find a way to leverage the IGP that goes unused in most gamers' machines. But that is even more difficult. Plus many folks haven't upgraded CPU in years for obvious reasons. My 3770K has an older IGP that doesn't support DX12 AFAIK.

This all seems a little stillborn to me.
 
Using the IGP for post-processing seems nice enough.
What I will find most interesting in DX12 is actually the linked mode. It seems to make SFR much more doable and whatever goes into the GPUs' memory doesn't have to be cloned.
In memory-intensitive scenarios where a single GPU would run out of memory and be forced to swap time-sensitive stuff using the PCIe bus (e.g. very high resolutions on a 4GB Fury), we may be getting dual card setups where the result surpasses the sum of two single cards.
How cool is that?
 
AFR on really asymmetric configurations just sounds awful. I have to wonder what the frame times are like.
Regardless, it's mentioned in the article that AFR is only their first implementation (I guess just as a basic test of EMA to start?).
Hopefully they'll implement something better in the future.

I could see some potential for this sort of thing for rolling PC upgrades - You buy a new graphics card every couple of years, and you use your old card + your new card at any given time. Of course, probably wouldn't work out so well in practice due to differences in feature set/architecture, but one can dream.
 
Well AFR on (not that much) asymmetric configurations may not be all that bad if you're using adaptive sync, for example.
 
Well AFR on (not that much) asymmetric configurations may not be all that bad if you're using adaptive sync, for example.
Adaptive sync is not magic... it can't handle very high frequency refresh changes like you'd have in an asymmetric AFR scenario and it won't solve any of the back-pressure/timing problems in these configs. Adaptive sync stuff helps in cases where you are running a relatively consistent throughput but less than the monitor refresh. It doesn't fundamentally alter the fact that you can't predict when a frame "will hit the display" and still have to use rolling averages based on GPU back-pressure to approximate that.
 
Yes, AFR is the least interesting of potential split GPU rendering methods. I'm guessing they tested that first as it would be easy to see how it performed versus traditional AFR methods as used by SLI/Crossfire.

I'd be interested to see them do something similar to what Firaxis did with split GPU rendering with Mantle. Except this time with GPUs from different vendors.

Regards,
SB
 
If they don't see some visible transitions with SFR, it would interesting to hear the explanation. Whether there is a visual discrepancy between single-card and AFR might be worth investigating, as there shouldn't be pixel-perfect consistency between hardware implementations, much less hardware vendors.
AFR would at least keep things consistent within a frame, and AFR has some clear transition points between frames for data that is being reused.

SFR may not have that same luxury, and the limited discussion of frame time variance makes me wonder about how frame pacing is being handled.
Alternating frames does give more time to hide GPU-variable delays with frame syncs, whereas SFR removes the timing margin that various pacing methods add.

The multi-adapter method where specific stages of the rendering process are divvied up seems more amenable, since the transition between generating a resource and reading it is a natural demarcation, and slightly different but self-consistent input from one stage is fine if it is processed consistently in the next.
 
Last edited:
You might actually have a point about the terrible scaling. Comparing a single gpu Guru3D review published today using original AOS benchmark at 2560x1440.

index.php


http://www.guru3d.com/articles_page...ce_gtx_980_ti_platinum_edition_review,24.html

its things like these that make so many consumers suspicious of certain sites. I was wondering why they used a fury nano in those tests and went checking the other pages. On those pages they used a Fury X. So they had one. Maybe I missed something?

I don't think AFR is worth investigating at all.

It would be exciting if they could find a way to leverage the IGP that goes unused in most gamers' machines. But that is even more difficult. Plus many folks haven't upgraded CPU in years for obvious reasons. My 3770K has an older IGP that doesn't support DX12 AFAIK.

This all seems a little stillborn to me.

That's true for older intel CPUs. But APUs from AMD would be one area of benefit. Given the APIs lower overhead, the CPU with a relatively solid iGPU and relatively cheap dGPU could be really nice on the low end. Could be interesting to see how minimum requirements change.
 
its things like these that make so many consumers suspicious of certain sites. I was wondering why they used a fury nano in those tests and went checking the other pages. On those pages they used a Fury X. So they had one. Maybe I missed something?

Ultimately the cards are in the same price bracket so I don't see a problem.
 
Back
Top