Flat screen vs virtual

[Muff_Huggar]

While fooling around with some different settings over the last few days, I've found something on my setup that has surprised me...The load on my computer appears on the surface to be less in VR than in flat screen (FS).

 

Basically I run an I7 4770, 24G ram with a vega 64 reference card. There is no OC/UV on the cpu and the GPU runs in an AMD "turbo" mode with the fans running a little faster sooner. The gathering tool is HWINFO. Everything is up to date on OS (W10 0918), AMD (18.12.2) and IL-2 (3.008). Graphics in IL-2 is set to Ultra; the settings are the same in both VR and FS (FPS is not targeted in flat screen, though). Flat screen runs about 160FPS; VR on a Lenovo WMR headset is a consistent 45FPS with occasional jumps to 90. After flying for a while in both modes, I notice the following:

GPU clock speeds in VR are significantly less than FS (about 1500FS 1100VR). Temperatures are significantly less in VR. Fans are higher in FS. But what I found the most interesting was the core/thread usage numbers. In FS, one of the four cores was signficantly higher than any of the other three. And it wasn't the same core each time. In VR, the core/thread usage was balanced across all cores/threads at about 35%. I have made no assumptions about what this all means, nor have I jumped to any wild conclusions nor even formulated any off-the-wall conspiracy theories. I would be interested in someone else looking at FS vs VR to determine if they see the same thing. It could, of course, just be an aberration or maybe just an old man's cry for attention, eh?

[Mitthrawnuruodo]

Those observations aren't strange at all. Performance in VR is always worse because because the game essentially has to render everything twice. 

 

The CPU/GPU usage numbers might seem strange, but they also make sense. You see, VR performance in Il-2 (and many other games) is heavily CPU-bound. Although it's the graphics card that generates the final images, the CPU must feed it with draw calls to tell it what to do each frame. For stereo images for VR, this CPU rendering overhead more or less doubled. The GPU clocks and temperatures will be lower simply because the graphics card is held back by the CPU. To make matters worse, this CPU rendering overhead all happens on a single CPU thread. This severely limits performance.

[Muff_Huggar]

I think you misread what I wrote. I only see single thread load on the CPU when I'm flat screening. When I'm in VR, the load is balanced across all the CPU cores/threads and is much less than the single core numbers I see in flat screen. And the single core/thread usage in flat is nowhere near what someone would consider "bound." 

 

From the number I saw, I would say performance in VR is BETTER than in flat screening...load on GPU and CPU is less than what I expected.

 

That's why I asked if someone else could do a couple of tests to see if my results can be recreated.

[Mitthrawnuruodo]

The results are entirely normal. Load on the GPU and CPU is indeed lower because it is being limited by that single thread. The CPU-bound nature of Il-2 has been extensively tested. In this plot, taken from popular benchmark results, we see that VR performance is proportional to single-thread CPU benchmark results.

 

The CPU utilization figures are deceiving because a CPU core need not always run the same thread. 100% utilization is not necessary for the rendering thread to be the bottleneck.

 

VR performance certainly can't be better than monitor performance as it requires additional processing. This is shown by the lower frame rates. Note that the extent of the CPU bottleneck may vary with graphics settings and GPU performance.

 

[chiliwili69]

Basically in your monitor test your bottleneck is GPU (look at the GPU load and you will see it is near 100%, that´s why it goes up to 1500MHz).

In the other hand, in your VR test your bottleneck is the CPU. (look at the GPU load and you will see it is far from 100%).

Don´t trust the CPU core/thread usage % number. They are a bit fake. In VR, IL-2 uses 4 threads, but one of those is the heavy one which is always bottlenecking the core where is running.

This thread is jumping from core to core to distribute heat.

More here:

https://forum.il2sturmovik.com/topic/29322-measuring-rig-performance-common-baseline-for-il-2-v3008/?do=findComment&comment=499246

 

Edited December 17, 2018 by chiliwili69