Asynchronous Compute

This would partly explain why at highest resolutions we actually see a drop in performance on AMD cards in dx12 games like QB. Makes sense as these cards weren't made for such demanding use of AC.

 

What do you mean by a drop? They probably benefit most from async shaders at high resolutions (provided they have the compute to spare) because pixel shader workload is significantly higher.

On the other hand at low resolutions and higher framrlerates there are more likely to be geometry bottlenecks stalling the pipeline, and async shaders can make use of that stall time

 

Have a look here ;

http://www.overclock3d.net/reviews/gpu_displays/quantum_break_pc_performance_review/8

http://www.overclock3d.net/reviews/gpu_displays/quantum_break_pc_performance_review/9

http://www.overclock3d.net/reviews/gpu_displays/quantum_break_pc_performance_review/10

 

I see what you mean, I think that's actually a memory limitation coming into play for the Fury X. I've seen 5.4GB at max settings 1440p

Nonetheless results are all over the place, as far as I know this game is vsync locked, so those averages are 60s and 30s and nothing in between.

 

Yeah, possibly memory limitations coming into play aswell.

That temporal reconstruction thing though...we don't need that feature on PC.

http://w ccftech.com/quantum-break-720p-base-resolution-confirmed-temporal-reconstruction/ (remove space between w and c).

 

Edge blurring is atrocious, looking at hair or dense foliage makes me nauseous

 

AC load is in no way connected to a rendering resolution.

The reason why the AC gains may become lower in higher resolutions is because the caches are getting a much higher traffic in them and since AMD's AC is hitting the caches pretty hard (you have to run several workloads on different datasets on the same h/w) the evictions and caches overflows can happen more often when you have more pixels to render. But that's just a theory. It may be that concurrent AC is just completely unpredictable on anything but the card it was specifically tuned for.

 

There are no indications that QB is even using async compute at the moment.

 

Depends how much workload is on the compute queues really, but yeah entirely possible

Easy enough to test if async is being used, gpuview profile

 

Does it even work with UWP apps?

 

No clue, but it should.

 

From what I saw by high resolutions is mostly Gpixel fillrate (ROPs) that matters the most, not compute or async..

Crysis1 was perfect example, this one loves higher Gpixel fillrate. Seen it every time I upgraded my gpu, 8800gts had 13?, 250gts, 25, 780 50, 980ti 110..



//Btw which FFXIV was tested there by power consumption in another thread? The latest - heaven sward?

 

Yeah but the FuryX doesn't seem rop limited, it tends to perform better at high resolutions generally, relative to the 980ti

Latest one yeah heavensward, Laithan convinced me to try it for stability testing and I have to agree, better than furmark and certainly better than 3dmark

 

I believe we see that behavior because the 980Ti is both memory bandwidth starved and has less efficient ROP's than FuryX.

 

I saw Fury is ROP limited too, that HBM saves it a bit, but overall its still crippled by 64ROPs that's why sometimes 390X comes so close to it.

Ah ok, well for me personally I noticed it can be a little cpu test, gpu also but not much as idk 3dmark2011 or RE6 or Hardreset @ 4xMSAA (powerlimit whoe). Had this with 780GTX OC too, although I knew its limit 1215mhz and anything higher failed in FFXIV too..

The best "cpu" gaming test for me was RE5 in dx9 @ 720p & Lp2 dx9 @ 720p, both with custom jobthread >> in my case 8 or if you want to test try 12.

 

Fury X isn't ROP limited:

Source

 

Well it still cant come close to 96 rops with much higher mem bandwidth.. so it is limited - anandtech pov up or down. I saw this in another review..

Imagine if it had 96, those numbers would be in 50 - 60 gpixel for sure.



Anyway this is totally offtopic now.

 

fury x is geometry limited !

 

In high resolutions, across general landscape? I find that unlikely. Fiji is definitely underperforming for its specs and ROPs and/or MCs seems like the natural place to look for the bottleneck.

 

Good news for Pascal.