High DX11 CPU overhead, very low performance.

Nice results. Now i wonder what crossfire will do

 

Mantle is in maintenance mode for a long time now. DX12 and Vulkan are actively developed. I'm sure that MS has done a lot of things to accommodate DX12, the main one being WDDM 2.0, that seems to benefit older DX versions too.

Drivers are very optimized in general. I bet AMD wouldn't want to be known for intentionally gimping the CPU's of 80% of the people who buy their cards. Even the "Quantum" computer they presented had an Intel CPU. Their CPUs will start performing more and more better with DX12 and Vulkan. You can even see that in some of today's properly optimized titles, like BF4, the Witcher, etc. I believe that they actually made the right choice at the wrong time with their CPU line.

 

Actually with 1st versions of 3DMark, API draw call test for Mantle did considerably better than it does now. There has been some kind of degradation either in way 3DMark works with Mantle or in drivers.
And my guess is driver since BF4 was running considerably better with Mantle at launch than with DX11 and now it is better to use DX11 in most cases.

 

Do we also get more drawcalls/less overhead with DirectX 10?
For example in Flight Simulator X, which still runs terrible on most pc's because of the CPU load.

 

I'm sure the game runs so bad because it uses just 1 or 2 cores or something like that?

 

Have you tried by using older driver DLL? Keep in mind also that Mantle was more or less optimized up until Hawaii, if my feeling is correct. I wouldn't even bother with Fiji, for example.

 

That shouldn't be too big of a factor here. Fiji is still a GCN incarnation (thus more or less "backward compatible" on it's own) and Mantle is supposed to be a true low-level API, so the Catalyst part shouldn't even make any significant difference (in theory, unless some fundamental changes happen or bugs occur) but either "work" or "not work", the overall performance and any unique GPU-specific optimization are more or less up to the graphics engine (which now has a "driver" like part) and that shouldn't even be relevant with the draw call benchmark because it's more or less a CPU benchmark (so unique GPU-specific optimizations would be pretty much fruitless anyway, the GPU isn't fully utilized...).

Although I can believe "anything is possible", especially with AMD software support, "anomalies" might occur frequently enough so they look "normal".:infinity:

 

It is a low level API after all. The driver's job is to interface properly with the driver. I honestly don't believe that AMD has really wasted their time with Tonga/Fiji into optimizing Mantle at all, especially since DX12 is basically here now.

 

Nope, it actually uses up to 8 cores.

 

Windows 10 single GPU 960/1500

15.7 (15.7 main driver)

Spoiler

1055 (15.7 main driver)

Spoiler

 

What's your cpu speed? I get 1.3M single threaded on same cpu @ 4.3GHz (Win7).

 

3.9GHz

 

With linear scaling your 4.3GHz score on ST would be 1 332 335. Sounds about right.

 

There's only one way to find out.

 

1,223,000

 

What no performance improvements at all? And that was with 1055? And you sure that was ST? (ofc you're sure but I'm confused :/ )

 

20k improvement yes, with 1055.
Are you running the test with a 7970 clocked at 1.1GHz?

 

Yeah I am, but I get same results as stock aswell. Clocking only makes difference on mantle where you get gpu bound.

Could you test again with setting cpu affinity for 3dmark to use only 0.2.4.6 cores of cpu, so you don't use hyperthreading (I get a bit better; and stable results this way on dx)

 

The 1055.0 driver is another step in the right direction. Almost 10% up for me. The 3DMark score is a tiny bit better because of the Intel driver (since I upgraded them I get a constant +100 on 3DMark) but the API results are representative. Almost 10% up. Color me Impressed. Just 800k draw calls to get to NVIDIA numbers (I'm not even sarcastic).

API Scores.

Spoiler

15.7





1055.0

Graphics Scores.

Spoiler

15.7





1055.0

 

I have a different opinion.

This step-by-step crawling makes me think that this is a brute-force fight, a sequentially repeated power-push through the barricades on the old bumpy road which will probably converge to a lot smaller number at the end through the accumulation of progressively diminishing returns.

I mean, a continuous progress is more probably an indication of optimization rather than a complete rebuilding from the ground up (which would probably manifest in a single huge starting leap, optionally followed by some slow minor crawling as fine-tuning of the new code).

Although, they might do a complete restructuring by replacing parts one-by-one in a modular system. (I have no idea how modular or monolithic their code is. Even less idea if replacing some modules could make any significant difference without aligning the others on the new line.)

It's all just my subjective guessing though but I am less optimistic than you (I don't think they will get too close to nVidia during this stage of development - which could also mean "ever" since DX12 is here...).