Crystal with / without dynamic foveated rendering compared?

[Fish]

For those of us considering a Pimax crystal light, it would be helpful to know what's the difference in performance with and without DFR, if pimax crystal owners could weight in. A percentage difference would be helpful rather than an FPS, since there's other things to consider in hardware setup. 

Thanks in advance. 

 

 

[Dangerzone]

9 hours ago, Fish said:

A percentage difference would be helpful rather than an FPS, since there's other things to consider in hardware setup. 

Or both.  But definitely a percentage difference for sure! (For those who are lucky enough to actually own both )

[edmuss]

In theory you can sort of replicate the expected performance by changing your current headset resolution to suit the different pixel counts to see how it impacts on performance.

I'm using the varjo aero using varjo_foveated, rather than the crystal using quadviews/pimaxXR but the same still stands as the functionality of the software is very similar.

Running the aero at 39ppd foveated gives an equivalent stereo resolution of around 4620x3960 per eye (around 36.6mp total pixel count) and at this I get 85 - >90fps, if I were to uninstall varjo_foveated and override the resolution in openxr toolkit I guess I would be getting around 40fps.  The total pixel count with my varjo_foveated settings is about 5.8mp, equivalent of 1700x1700 per eye stereo.

So essentially, run your headset at stereo resolution of 4620 wide per eye and then run it at 1700 wide per eye and you should be seeing the performance delta between the two systems.  Bear in mind that there is additional CPU load from having to render the scene 4 times with dynamic foveated rendering.

Now the fixed foveated rendering that is being touted by pimax can never gain the performance uplift of dynamic foveated rendering without absolutely slaughtering the visual quality, from what I've seen, it's only a 5-15fps performance increase; nothing to be sniffed at but far from the 80-100% increase I see in the aero.

Do note that the varjo implementation of dynamic foveated rendering is more efficient than quadviews and does garner slightly improved performance, likely because it is working with native code implementation within the runtime and DCS.  Also that the higher the resolution, the larger the potential gains will be with eyetracking.

[Mr_sukebe]

I’d apply some caution to the above method.  Whilst it’s a way to replicate the theoretical visible resolution, the Quad views process does have an additional overhead that eats up some of the freed resources.

Don’t get me wrong, DFR does work, use it with my Quest Pro, but it’s not quite as simply as saying “it’s only processing 1/3rd of the number of pixels, therefore…”.

[edmuss]

Absolutely, hence my caveat of the additional CPU load, essentially if your CPU frametimes are anywhere near close to your GPU frametimes then dynamic foveated rendering isn't going to help.

[Fish]

Thanks for the input guys. I can see there's complexity in this measurement. And it looks like what happens with the Aero or Quest pro has little bearing on how the Pimax performs.

Would it not be plausible to take a render resolution, which ensures you are generally not CPU bound, and carry out bench mark with the crystal. Then apply a degree of DFR, for which the user sees no noticeable deterioration in quality and carry out the benchmark again (I understand this setting will be purely subjective). This in theory would give and estimate of the expected (not CPU bound) gain using Quadviews DFR (guessing this is the way pimax did their promo video for it). It would probably need to be done with a high end system including a 4090. And I absolutely understand it's not an exact science, but would give some measure of the 'potential' benefits of spending the almost double on the eye-racking version 

[edmuss]

The crystal is essentially a copy of the aero specifications wise, the crystal and aero both run foveated rendering in DCS with software written by mbucchia, the mechanism is the same and the resolutions can be compared directly.  The only difference would be that the varjo implementation is native to DCS and the pimax/quadviews is emulated.

I see no deterioration of visual quality between stereo render and dynamic foveated render, but I do see an 80-100% uplift in performance.  For whatever reason (I think it's to do with the DX11 engine) DCS seems to bottleneck really hard when the the pixel count gets high, this is why the G2 can be such a pig to run at 90Hz.  Sure you can brute force it with a massive GPU but there are much more elegant solutions such as eye tracking; it allows you to slash the rendered pixels to valve index levels whilst retaining insane clarity

edit: just watched the video linked, that's the sort of performance I get with my 3080ti except with simple aircraft like the SU25 I'm locked to >90fps.  With a 4090 I would be able to supersample even higher.

Edited June 3, 2024 by edmuss

[mbucchia]

10 hours ago, Fish said:

Would it not be plausible to take a render resolution, which ensures you are generally not CPU bound, and carry out bench mark with the crystal

The thing you need to understand is that quad views is going to **increase** your CPU utilization, regardless of FFR or DFR. That should be your primary concern. FFR vs DFR is a secondary concern and is more about how sensitive you are to peripheral view degradation. There is also the Pimax OpenVR FR, which will cost you no CPU, but will provide much smaller gains on the GPU. With this one there is 0 difference in performance between FFR and DFR.

For short there is no "somebody please benchmark it for me", because too many variables.

You can test out the impact of both quad views or Pimax OpenVR FR today by installing the Quad-Views-Foveated tool or the PimaxMagic4All tool respectively. You can configure the resolution to get close to what you'd get with the headset you are eyeing. This will give you a much better idea of what it will look like.