High DX11 CPU overhead, very low performance.

Ye it improves, the only doubt that i have, is how it compares to nvidia results now
In respect for overhead, that why the 1018.1 results of the monarch are essential XD

 

You knew that already. And what's your point anway since according to Monarch his draw calls skyocketed just from swapping the GPU. CPU never even affected it.

So either the CPU means everything or the CPU means nothing. Which one is it? You can't have both.


The more likely answer is Monarch doesn't have a 290X. Considering our near identical scores I'd say it was a 280. Still not trolling?

 

What it means is that you and Monarch get similar AMD draw calls with similar CPU's.

Yet when he switches to an NVIDIA card, his draw calls double. What does that mean to you? :infinity:

 

Read above. I'm through with tghis.



Btw I have W10 and was not using modded drivers. The 15.200 108.1 set are WDDM 2.0 The other set are only WDDM 1.3.

So put 2+2 together.

 

Since it is actually relevant, but locked, here are the benchmarks from the original post of the other thread.

And the excellent post by Vbs.

My software is Windows 8.1.1 x64 with all the updates as of today (08/05/2015), my hardware (thanks Blackfyre ) is as you see it on the left. The game settings are everything to the max except non-post-processing antialiasing, unless I provide a link about the specific title's settings. The resolution is a humble 1080p60 (GPU donations accepted ).

I do not include min/max framerates, as they are completely useless and they clout the charts. I have gone the other way around and tried to find some kind of metric that would show the relative variance and smoothness of the experience, instead of only the framerates (which don't mean much from a point on). The 99th percentile number shows in under how many ms the 99% of the frames is presented, and after Vbs' suggestions I saw the light and I'm removing the Standard Deviation that was too linear to represent anything very meaningful, in favor of Variance. It can be better explained here, by people who know much more than me. Variance and the 99th percentile are by far the most important metrics here, as the smaller they are the smoother the frame delivery is.

Let's start with the games then.

Assassin's Creed Unity

Spoiler

Game Settings:
Environmental: Quality High
Texture Quality: Ultra High
Shadow Quality: Low
Ambient Occlusion: HBAO+
Anti-Aliasing Quality: FXAA
Bloom: On

Test Route.
You can see in the frame chart, this driver is much better at delivering frames as both the curves and the numbers are indicating much smoother deliveries. The dual core peformance seems worse though.

Dying Light

Spoiler

Game Settings, Benchmark Route.

Same story here, with many more gains in the quad core test. This will become a pattern with this driver, as you'll see in the next tests too.

Total War Rome 2

Spoiler

Rome 2 is a badly written game (still lags horribly by just selecting a unit in the campaing map). These numbers are taken using the game's benchmark. The 1018.1 plays and feels worse with a dual core and the Variance number confirms it. It is the opposite with a quad though. Whatever these drivers have, it is not focused on dual+HT scenarios.

StarCraft II: Heart of the Swarm

Spoiler

Althoug StarCraft II is a DirectX 9.0 title, we curiously see a difference in peformance, towards the bottom. Again dual core performance is harmed with the new drivers, and the game more or less performs in a similar manner to Rome 2.

Unigine Heaven

Spoiler

Here things are neck to neck and within all kinds of statistical error. For all intents and purposes, the performance is the same, with the 15.4 driver being worse in a Dual Core scenario for the first time, and being more "spiky" in general.

Unreal 4 Elemental Demo

Spoiler

Another case where everything is within the margin of error.

Metal Gear Solid V: Ground Zeroes

Spoiler

Here again we take the paces of the intro sequence of the game (and it is Hojima-long), again the performance is the same between the driver sets, except the 1018.1 being worse in the dual core scenario.

Final Fantasy XIV: Heavensward Benchmark

Spoiler

This is one of the examples where the 1018.1 driver simply gave worse results, especially in the quad core scenario. I run the test three times, the results were consistent with what you see. Much lower average framerate, a tiny bit better frame delivery.

Monster Hunter Online Benchmark

Spoiler

Another one where it is more of the same. Practically identical performance, except that in the dual core scenario the 1018.1 is a bit worse again.

Project C.A.R.S.

Spoiler

For the first test everything was at Ultra except Track Detail which was Low, and Detailed Grass which was off.



Practically identical performance from both drivers, with a tiny bit more of frame stability for 1018.1 in the quad core scenario.

For the second test Track Detail and Detailed Grass were at Ultra. There is something wrong with this game, nothing being displayed could ever possibly justify the performance hit.



Still, this is one of the cases where the 1018.1 driver carries its own weight, and for the first time in a dual core scenario too. The 15.4 is in practice unplayable.

Civilization V

Spoiler

With its support for DX11 multithreading, Civilization V is the personification of "last, but not least". The game provides its own benchmarks and metrics, with the LateGameView simulating a game of 300 rounds played already, and the Units test filling the screen with, well... units.



The results are the same. Apart from a tiny increase in the Units Full Render score, things are more or less the same.

Unfortunately I didn't have more games/hardware available for testing, but the results are quite fascinating. In almost all the tests we have a drop in the dual core performance of the new drivers, which is simultaneously combined with smoother frame delivery in quad core scenarios. The StarCraft II results are surprising, as DirectX 9.0 performance was supposed to be the same and they might be an indication that this is a change in a lower part of the driver (possibly the scheduler itself).
This doesn't solve the problem of low performance with lower end CPU's, it actually makes it worse. It is good news for people with substantial hardware though, as things seem to move in a direction where the GPUs seem to have more of leeway given by the higher-end CPUs.
If this trend continues in subsequent releases, the stutter gap with NVIDIA will probably close, but the lower cpu/thread cap will probably remain and worsen.
It took me some time to post this, as it has been a long time I had to run benchmarks and check things so thoroughly.
I await your comments, extra benches and hardware for me to test. :infinity:

 

I gonna explain to you very careful how the things work.
His drawcalls skyrocket cause he switch from an AMD to an nvidia which have massivaly better draw calls, since their driver have less overhead, like you sayed.

Altough pairing an higher CPU/overclock will gonna give him better draw calls results.

I can remove the overclock of my CPU and show to you, if you want too.
But CPU affects too, unless you have such high end CPU that the driver/GPU will not allow your CPU to archive better results

 

Yet on equal hardware 15.4 is 40% slower draw call wise to the 1018.1, in Windows 8.1 with WDDM 1.3.

My math is fine, thank you.

 

I use 15.4 drivers and I get ~1mil dx11, ~17mil Mantle.

 

PrMinisterGR just ignore Pill;P

Also got a question, which driver are which now. I get confused.
Modded here, VBS there.

I want to try my 6970 with the driver that supposedly might reduce DirectX 11 overhead on Windows 8.1 (or perhaps it only was on W10?).

Sorry if it's obvious, get confused with all the names;S

 

This is the driver we are talking about.

Can you do me a huge favor and get a 3dmark score screencap with your current driver, and then with this one?

 

Sure thing, is it the Basic 3Dmark i should download and do the API Overhead Test?

Any certain settings?

 

Leave it on default and make sure you get an overhead test. I believe it is included on the standard bench run.

 

this is my first post but I thought it might help all. AMD need to know that I tried windows 10 with driver AMD-(15.200.1018.1)-WHQL-WDDM 2.0-DX12 driver for Win 10 TP-April 1, 2015 and windows 7 with the latest beta driver.

now my system specs

radeon 280x 8 gig of ddr2 ram and a phenom x4 b35 a 3 core wth the fourth core unlocked running at 3.48. resolution I used was at 4800x1200 3monitors.

my settings were
all fx part on max other than rain and in the main graphics settings it was

texture res=high
anisotropic=x8
vsync=off
all anti aliesing off other than fxaa =high
refections high
env=medium
car detail=high
track detail=low
shadow=med
motion blur=med
grass =off
particles=med

on don't forget im running 3 monitors with a old cpu lol well

with windows 10
in a clear race 12 cars min fraps=67 max=74

in bad weather thunderstorm wit 12 cars min fraps=47 max=54

now with windows 7 beta drivers

in a clear race 12 cars min fraps=47ax=74

in bad weather thunderstorm wit 12 cars min fraps=32ax=52

that's a min fraps increase of nearly 20 fraps I did with 1080p on windows 10 and had over 100 fraps

lol

hope that helped

with that 20 fraps increase if you look at the benchmarks for project cars that puts the radeons level with nvidea the 280x will perform the same as a 770 when windows 10 is released

 

15.4 Beta


AMD 15.200.1018.1 (Driver installed with 15.4 Catalyst installed)


I didn't quite get how i would use DDU as i would lose Catalyst if i just installed the driver. So i just installed it over the 15.4 one, but hopefully it worked as well.

EDIT:

Oh wait i am blind..
will redo the VBS one.

 

CCC is installed with the modded driver.

 

Yeah noticed that afterwards;P

Well i installed it (got a BSOD first time though, second time it installed so not sure what happened).



Worth noting is that VSR doesn't exist for me (I may be able to get it working with the hack, got it working with the 15.4 driver).

 

Thanks for that, that's where the 280x should be (if not a bit higher). Now for Windows 8.1

Make sure you use DDU in safe mode between installs, and that everything gets uninstalled (including CCC). The modded driver should give you a working CCC with the VSR option.

 

I did DDU in safe mode.
Start in Normal Windows.
Tried to install Driver got BSOD(some catalyst files got installed though i think).
Tried again next boot, succeeded and Catalyst worked fine. No VSR though.

Remember that i got 6970, that might be the problem.

(Using the VSR tool just causes Explorer to do some weird restart thing after boot as far as i can tell, and does not add any resolutions).

 

In light of the benchmark thread being closed, I think we should put things into perspective to what conclusions can be drawn from PrMinisterGR's tests.

There were two independent variables being tested: (a) Driver versions (15.4 and 1018.1) and (b) Number of CPU cores (dual and quad-core, corresponding to a logical 2+2 or 4+4).

Let's also assume that the performance of each game reflects a combination of: (a) Driver overhead and (b) game engine efficiency.

These allow us to identify four possible comparisons:

1. On a quad-core CPU, 15.4 vs 1018.1: Compares driver overhead between driver revisions;
2. On a dual-core CPU, 15.4 vs 1018.1: Compares driver overhead between driver revisions;
3. Using drivers 15.4, quad- vs dual-core: Compares driver overhead CPU scaling plus game engine CPU scaling, between both CPU types;
4. Using drivers 1018.1, quad- vs dual-core: Compares driver overhead CPU scaling plus game engine CPU scaling, between both CPU types;

Out of all these scenarios, (1) and (2) will be the most interesting, as the quantity being measured is reduced to just the driver overhead, while game engine CPU efficiency is mostly constant.

Comparisons (3) and (4) are also useful to see the differences in performance resulting from CPU scaling between the number cores, but won't be as useful as (1) and (2) to pinpoint driver-related overhead issues (as they reflect two quantities changing simultaneously), which is the main focus of this thread.

PrMinisterGR, using "bcdedit /set numproc 4" would make windows boot up with just a dual-core (4 logical cpus). I wonder if there's any meaningful difference in doing that vs forcing an affinity mask in process lasso.

You can use "bcdedit /deletevalue numproc" later, to delete it.

 

Would that keep the first four logical or the first four actual processors? I don't want it to keep CPU 0,2,4,6. My guess the difference would be that there would be no CPU to attend to Windows tasks in the background. I don't have now the time to test. I'll try to expand the testing a bit too to some more titles.

We can really only compare between driver versions and guess about efficiencies, since I don't have an NVIDIA GPU here to compare directly.

What made more of an impression to me was that DX9.0 tests looked like they were affected, which means that the changes might run deeper in the driver. Unfortunately I have no heavy OpenGL title to test with now, but I might have Wolfenstein The Old Blood later in the week.