I bought a 4090 based PC last week, this is the future for video games. It also helps 4090 is dual use for AI, its like 20x faster for stable diffusion than a RTX2000s card.
Although I do game, I'm mostly interested in AI work. My current baseline 3080 has been just fine, although to be fair I haven't yet experienced path tracing.
The extra horsepower + DLSS frame generation will give you around double the FPS.
https://www.pugetsystems.com/labs/hpc/nvidia-rtx4090-ml-ai-a...
https://lambdalabs.com/blog/nvidia-rtx-4090-vs-rtx-3090-deep...
None of that has real meaning for AI. But the 4090 is just straight up 40% faster as well.
I think one of the strong suits of game graphics is the results at any cost model of rendering tech, where the focus is on using whatever methods get the best results the fastest (and hopefully within frame budgets) wins out. Since the advent of polygon rendering rasterization has been the best solution for attempts at realism (be that hyper- or photo- realism). However, rasterization and the rendering ‘shortcuts’ used to achieve certain effects were also useful for other styles of rendering. However, if every bit of graphics silicon is going to be designed and optimized for a physically based, photorealistic rendering result it may be difficult to bend that functionality toward some non-photorealistic usage.
I worry everything is going to end up having the same stamped out photorealistic appearance, in a similar way to the comments I’ve seen about the emerging Unreal 5 ‘look’. And that to get away from that look will require giving up a large chunk of GPU rendering horse power.
The path tracing mode mentioned here however, does "full raytracing" like blender does (outside of a few exceptions such as particles). Same with Portal RTX, Quake RTX and a few others.
https://www.nvidia.com/en-au/geforce/news/cyberpunk-2077-ray...
Nvidia's own announcement shows the path tracing overdrive mode is still tracing rays from a G-buffer, which would've been generated with the rasterizer.
For now, we'll always have both, as too many cards/consoles can't path trace at a good enough rate. But the artist pipeline wins that come from fully path tracing lights are too good to pass up, so it's coming.
This article alone shows that basically all of the frame is path traced rather than raster. It's coming, and faster you might think.
When watching direct comparison videos path tracing was slightly better when someone pointed out more accurate lighting details but not in a way that was noticeable if someone didn't point it out.
What I found works for path tracing is objects that don't have strong textures or bump mapping. I've tried this when writing my own path tracer in a programming project (not cyberpunk) for simple monochrome spheres. Then in that case it makes it look very real, but in something as visually detailed as cyberpunk it's barely noticeable.
I also think the lighting in cyberpunk is too urban and neon. It's not something our eyes evolved to interprete with extreme fidelity so our brains cant really see the difference. I think the effect is much more noticeable in environments with fewer light sources and light sources that aren't neon. For this reason I suspect red dead redemption 2 would benefit much more from path tracing.
That's just my opinion. I have a 3090.
And there's also a bit of knowing where to look at, what kind of scenes and weather types benefit the most, and so on. If you have someone point out the differences, on specific scenes, even the most casual observer will notice the huge improvements[1]. However if you boot up a random scene, it may look very similar.
What exactly are you referring to? I searched for it and all I got was stuff about actual fishing rods and some Minecraft and WoW guides...
Try to picture you flying a plane in a game. You're over a lake and your plane is occluding a mountain in the distance.
With SSR, those bits of the mountain your plane is blocking won't be rendered in the reflection.
Same with a fishing rod. They're tall and block a lot of things in the distance, thus breaking SSR.
Pay attention to the reflection in the water here when Sonic casts the rod (time-stamped): https://youtube.com/watch?v=IaPMkpDOFng&t=42s
Since it doesn't make sense to have a hole in a reflection when there's not actually a hole in the object, the reflection just stops when something occludes the object being reflected.
In some scenes, the bloom just removes all detail and makes parts of the scene look like a blurry mess.
Fortnite, for example, does it right. I remember watching a character dance and thinking "This is so god damn smooth, why does it look so smooth?" and then I noticed the motion blur. When done right, it's subtle and you don't even notice it unless you're looking for it.
Games that render frames don't actually replicate this because the object is rendered on discreet instances. Thus your eye is not able to properly capture the "motion blur" from a set of stop motion images.
To fix this, games must use artificial motion blur to replicate what the eyes see. Disabling motion blur is a mistake it is required for fidelity.
The problem with motion blur is if it's done for objects moving too slowly. Then it becomes noticeable because if your eyes are fast enough to capture the object without blur then adding blur becomes noticeable.
1. Blur on slow objects, as you mentioned. Something moving 5 pixels per frame probably shouldn't have any blur applied.
2. Blur exceeding the distance moved per frame. If an object moves 20 pixels between frames, the blur should be less than 20 pixels wide. Ideally, I'd even argue that the blur width should be half the movement between frames. The example above of something moving only 5 pixels per frame probably shouldn't have any blur applied.
3. Blur on objects that are not moving relative to the camera. In a racing game being played with the camera above and behind the car, yes, the scenery should be blurred, but your car and the car right next to you that is driving the same speed as you should not be blurred. Though I suppose this is really just a special/specific case of problem 2 mentioned above.
Of course, there's one problem with all of this: the video game camera is supposed to represent your eyes staring strictly in one direction. When your eyes follow an object, the object does not blur, but everything else you see does. This means that if you're tracking an object on your screen with your eyes, that object should not be blurred. Unfortunately, this is not possible for games to do, because they don't know where your gaze is actually directed.
Finally, I can't close out this comment without mentioning the greatest offender when it comes to doing motion blur wrong: simply blending each frame with the previous one. There was an N64 game that did this, and my god it's horrible.
Watch the digital foundry videos. If you watched the GamersNexus ones they are notorious raytracing haters and picked worst case scenes to demo path tracing for most of their videos (where the faked lighting looked as close to PT as possible).
This is where the performance gain for the 100s of lights Cyberpunk might have in a single frame comes from.
https://research.nvidia.com/publication/2021-06_restir-gi-pa...
There is a "Digital Foundry" video which mentioned it.
Yeah I think so too, but I do think that is a long way away. At least 5, maybe more, years. Because the sheer install base of non-path-tracing cards means a developer is silly to ignore it, and sadly the state of the GPU market means that that install base is going to stay as-is for quite a while, I think.
I hope I'm wrong: path tracing with proper dev tooling around it means we get near-perfect lighting in our games, with less development and artist time needed to achieve it. It'll be an amazing day once it's path-tracing-only.
The rasterizer is used to trace the primary rays to generate a G-buffer like any other deferred renderer that's been in use for decades by this point. What's unique is that all light transport is solved with path tracing.
So, a decade from now
>As soon as the average gaming PC and console has the power of a 4080/4090
Ah, more like 15 years.
Then, shortly after that, the mainstream won't be client compute, but cloud compute and game streaming. Although, I wonder if 20 years to build out full fiber for almost everyone in rich countries is realistic. I might be over optimistic here, actually. Maybe 30 for mainstream game streaming.
Secondary rays (GI, non-planaer reflections) though are quite hacky to accomplish with rasterization and benefits immensly more from the availability of raytracing extensions where your "only" problem is that you can't sample as much as you'd like so you need to converge a solution instead.
As soon as the average gaming PC and console has the power of a 4080/4090, rasterization is dead.