Yeah i see where you are coming from.
At the moment I'm inclined to believe this 'as a benchmark' is deliberately tweaked to push massive amount of Calls to give the benchmark purpose.
When you think that in DX12 and 8 thread i7 can push 15 to 20X as many calls then it would in DX11 its hard for me to believe that this level of Call demand is normal for this game, it certainly doesn't look like it should require that much, there are a lot of points of light and instances in it but not 15m worth of calls.
How the actual game behaves will also be interesting.
The benchmark is running a full game simulation, that means AI, Path-finding, Weapon ballistics (each shell is a physical object that has its own physics like in supreme commander). And all of that is before graphics become involved.
You also have to take into consideration that each new unit bumps up the work on the AI and path finding threads. So what they have saved in overhead in DX12 is also going towards enabling more units.
I think that if you could pause the game simulation then most of the CPU benchmarks would be far closer to the 5960X benchmark.
I don't even think the game is pushing DX12 anywhere near the Draw call numbers that the GPU's can handle. Considering how close the DX11 Nvidia results are to their DX12 results. The DX11 api results being around 1.2 - 2 million draw calls depending on the processor. Just because you can push that many draw calls does not mean they will. but they will certainly use the extra budget they now have.
Just found it, but considering the Benchmark details from AMDMatts results. the batches go nowhere near what DX12 is capable of pulling. And looking at his result, he was GPU bound the entire time, although his 5960x could have been managing over 100fps if the GPU could keep up.
The funniest part about the above image is that 11k draw calls in DX11 is already a long stretch for most games. The api test might show that a driver can do more, but it is only rendering simple textured cuboids and a simple global light.
The only thing that makes this a reproducible benchmark is that the unit movements and attacks are scripted in their execution, but they still perform all of their path finding and ballistic calculations.
This is as it stated on the benchmark itself, a "Full system test."
I think another major problem is that this benchmark gives a lot of information. but none of the websites were showing all of this extra information or explaining the benchmark correctly.