Discussion in 'Videocards - AMD Radeon Drivers Section' started by spectatorx, Jul 15, 2016.
Just run them again...
Async ON vs OFF
Seems much better now.
I think this is usefull to post here aswell:
I did some tests on custom benchmark settings with stock Fury X. Turns out tesselation ain't having any real difference.
Graphics test 2:
Max tesselation factor / Tesselation factor scale /fps
1(0) 28,31 fps
8(1) 27,89 fps
16(1) 27,65 fps
32(1) 27,54 fps (default)
64(1) 27,15 fps
Interesting, because in Firestrike Tess factor made a huge difference.
Bigger indeed. Just tested FS Extreme:
Graphics test 1
Tesselation level / Max tesselation Factor
10/8 40,99 fps
10/24 38,48 fps (default for FS extreme)
So yeah, there is bigger difference there.
It does support it. It supports both. Concurrent execution is how DX12 does things. Async is an optional technique over that. I've heard and seen so much bs form AMD fan boys the last couple of days has made my hair fall off. The same people who defend the use of the tessellation limiter in games, have a problem when NVIDIA's INTERNAL pipeline doesn't work in a specific way they imagine, despite NVIDIA hardware both showing the SAME output AND paying the performance penalty for their lack of support.
Tesselation = "Optional" DX11 era "feature", is enabled as per default in Firestrike benchmarks
-> can be disabled -> will INVALIDATE score
(has negative performance impact on all vendor's cards, more on AMD, less on NVIDIA)
Context Switching = "Optional" DX12 era "feature", is enabled as per default in Time Spy benchmark
-> can be disabled -> will NOT INVALIDATE score
(has positive performance impact on AMD GCN cards and NVIDIA Pascal cards, has negative performance impact on Maxwell cards)
Parallel Asynch Compute = "Optional" DX12 era "feature" is NOT integrated in Time Spy benchmark (purposely? "forgotten"? ...)
(would have hugely positive performance impact on AMD, can't say if negative or no impact for NVIDIA)
"Shame upon him who thinks evil upon it."
Tessellation is mandatory in dx11 isn't it?
it's mandatory in firestrike benchmark to be valid, as i wrote.
but it isn't mandatory in "DX11" as you can disable tessellation in either the most game options or driver side (radeon settings / ccc)
I had a run with 1200/1500 and scored 4679 gpu score on my 290x but
then i had a second run with 1200/1550 and did 4400. I think the score you got
is pretty low and i think its because of the memory clock, try dialing back to
1500 and see if it bumps up to 4600+
Ughh.. you say 2x290x but on the pic i see just a 290.. got me confused there.
Tessellation is a feature first pioneered by AMD. It will invalidate the score if you disable it, because there are VISUAL DIFFERENCES.
Context switching is a pipeline feature. It will not invalidate the score if you disable it, because there are NO VISUAL DIFFERENCES.
Async Compute is a pipeline feature. It will not invalidate the score if you disable it, because there are NO VISUAL DIFFERENCES.
WTF is that about?
Since when do we care on how the pipeline works as long as the visual results are the same? Why don't you care about Catalyst A.I.? If NVIDIA's way of doing things is slower, they pay for it (and they do, in Time Spy the GTX 980 is slower than the RX 480, the Fury X is faster than the 980Ti), and if AMD's way of doing things (bad tessellation performance) is worse, they pay for it.
Let me illustrate:
We don't care how we get the result, we care if the result is what the developers envision. All the rest are performance penalties that differ between architectures. There should be an uproar if there are differences in the final image quality, and not about engineering decisions either by AMD or NVIDIA.
Don't you just hate it when the map gets rendered with a parallel technique.
yes bro i have 2x290 but the bench only showing one card maybe the cf is broken
and i bench with one card.......... i disable cf in ccc
i will try with lower mem 1500
Nono don't, I thought you had 290X because you said 2x290x in your original
post, but then i looked at the picture and it said just 290.
aaa ok i have 1x290 +1x290x
and i use 1x290x
There's already a thread about this guys.
And I obviously don't agree with PrMinisterGR. This has nothing to do with fanboyism. It's a legitimate issue. It's not related to visual output/IQ, true, but it does go against what TimeSpy developers intended. And what they intended is to throw compute and 3D queues at the GPU. Nvidia prevents this on a driver level and prevents a performance DROP.
Quote from TimeSpy dev:
"The reason Maxwell doesn't take a hit is because NVIDIA has explictly disabled async compute in Maxwell drivers. So no matter how much we pile things to the queues, they cannot be set to run asynchronously because the driver says "no, I can't do that". Basically NV driver tells Time Spy to go "async off" for the run on that card. If NVIDIA enables Asynch Compute in the drivers, Time Spy will start using it. Performance gain or loss depends on the hardware & drivers."
If you do the same and disable 'AC' via TimeSpy benchmark, score is invalid.
So very simple stuff really.
Final result of the image is all that matters. Naturally maxwell or kepler cards ain't using async in driver level as it doesn't work properly. Nothing wrong in that. You still run the same code.
This is like saying nvidia isn't allowed to use multitheaded driver since developer didn't mean scenes rendered that way. That's what drivers are for. To execute code as effiiciently as possible with current hardware.
So, NVIDIA prevents a performance drop, with NO VISUAL DIFFERENCES SINCE THE SAME THINGS ARE CALCULATED and this is bad HOW? It's not even per case, this is not specifically for 3DMark, that's what the Maxwell driver does for everything. They DO throw 3D and Compute queues to the GPU, and that's how the GPU HANDLES them. NVIDIA doesn't prevent 3DMark from asking whatever it wants from the GPU, but the way they fulfill the request doesn't matter.
And they do pay the performance penalty for that, right? So how is the benchmark biased? Async makes the Fury X faster than the 980Ti. How is that bias FOR NVIDIA? They LOSE potential performance.
They don't lose anything. Not gaining or taking a hit are two different things.
And I already explained in very simple terms why this is an issue so I'm not repeating myself again. Do a run in TimeSpy, disable 'async compute' and post your invalid score here.