N
Nhranaghacon
Copypasta:
Soo, ray tracing they said, let's talk about "
realism" shall we (and I don't mean Pixar stuff rendered by a freaking render farm, but by NVs tech demo),
Corners:
if you need to be reminded how they look like, welp, a real photo:
you can read what's going on in this wonderful blog post:
Now, let's move to "
full RT" shall we? Let's be generous,
Quake.
it takes 2060 about 20 seconds to generate a decent quality frame.
So how do they render it in a fraction of a second? Meet
Green RT Fakery:
1) Temporal denoiser + blur
This is based on previous frame data, so with the textures turned off and the only image you're seeing is what's raytraced. Top image was taken within a few frames of me moving the camera, bottom image was the desired final result that took 3-5 seconds to 'fade' in as the temporal denoiser had more previous frames to work from. Since you are usually moving when you're actually playing a game, the typical image quality of the entire experience is this 'dark smear', laggy, splotchy mess that visibly runs at a fraction of your framerate. It's genuinely amazing how close to a useful image it's generating in under half a second, but we're still a couple of orders of magnitude too slow to replace baked shadowmaps for full GI.
2. Resolution hacks and intelligent sampling zones to draw you eye to shiny things at the cost of detail accuracy (think of it as a crude VRS for DXR)
Here's an image from the same room, zoomed a lot, and the part of the image I took it from for reference:
A - rendered at 1/4 resolution
B - tranparency, this is a reflection on water, old-school 1995 DirectX 3.0 dither hack rather than real transparency calculations
C - the actual resolution of traced rays - each bright dot in region C is a ray that has been traced in just 4-bit chroma and all the dark space is essentially guesswork/temporal patterns tiled and rotated based on the frequency of those ray hits. If you go and look at a poorly-lit corner of the room you can clearly see the repeated tiling of these 'best guess' dot patterns and they have nothing to do with the noisier, more random bright specs that are the individual ray samples.
So, combine those two things together. Firstly we have very low ray density that is used as a basis for region definitions that can then be approximated per frame using a library of tile-based approximations that
aren't real raytracing, just more fakery that's stamped out as a best guess based on the very low ray coverage for that geometry region. If I was going to pick a rough ballpark figure, I'd probably say that 3% of the frame data in that last image is raytraced samples and 97% of it is faked interpolation between regions and potato-stamped to fill in the gaps with an approximation. This works fine as long as you just want an approximation, because the human brain does great work in filling in the gaps, especially when it's all in motion. Anyway, once it's tile-stamped a best-guess frame together out of those few ray samples, each of those barely-raytraced frames are blurred together in a buffer over the course of several hundred frames. There will be visual artifacts like in my first point anywhere you have new data on screen, because temporal filtering of on-screen data only means that anything that has appeared from offscreen is a very low-resolution, mostly fake mess for the first few dozen frames.
By
Crispy
www.pcgamer.com
www.thefpsreview.com
.
-avg-fps-chart.png)
