Sorry I dont agree with your premise that RT cant be faked
Okay. Let's talk about this fakery. But before that let's talk about what makes RT so expensive.
RT is just an accurate way of modeling light transport. But it's expensive as heck on vanilla GPU cores because the algorithm is (a) extremely divergent and (b) very memory latency sensitive. It's divergent because if you shoot 1000 rays into a scene it's very likely that they will start doing very different things depending on what objects they hit (or don't hit) and where they go after reflection. It's latency-sensitive because the sheer act of figuring out if you hit something involves walking through a data-structure called a BVH sitting in GPU memory, and there's not much compute work to do in between steps of this walk. So the speed at which you walk through the data structure is determined by how quickly the memory is able to return the next element of the walk, but you can't get to the next step until you've finished the current step. This is called a classic "pointer-chasing" workload in computer science. Oh, and the BVH walk itself is highly divergent as well. Now, GPUs are built for extremely high parallelism on code that *doesn't* diverge (e.g., running the same pixel shader on thousands of pixels), and extremely high bandwidth to supply data to those shaders. But they are terrible at divergent, latency sensitive code.
CPUs are actually *excellent* at divergent, latency sensitive code. But they just don't have enough parallelism. In theory, an array of 1000 CPU cores could be an awesome BVH traversal + object intersection engine but it would be too expensive and would need a personal power station to run. And that's why we need dedicated RT hardware to perform this task.
Okay I now realize all that was a tangent but what the heck I felt like lecturing so you have to suffer.
Back to fakery. Yes, under certain constrained situations you can use a variety of more GPU-friendly approximations to avoid tracing rays through world space. For example, you can model reflections in screen space (a la SSR) instead of world coordinates, which is *much* cheaper and works beautifully
as long as reflected objects are on screen. You can even render the entire scene twice for reflections
as long as you're willing to pay 2x the render cost. You can use baked lightmaps (which are pre-rendered offline on CPUs for hours), which with sufficient resolution give you perfect GI
as long as there are no dynamic objects or light sources. You can use probe based GI for dynamic light sources
as long as you're willing to tolerate light leakage. You can use SSAO to apply mascara over the leakage and it'll look kinda okay
as long as you don't look too closely.
All of these hacks/approximations/fakery "almost work" as long as you meet the constraints. If you can build your art around these constraints then it will look good. But the moment your content starts stressing or outright violating constraints the approximations break down. As our art is becoming more detailed and photorealistic in other dimensions (higher geometry, physically based materials, more dynamism) it's becoming harder and harder for the approximations to keep up. The seams begin to show. Can you still in theory pile on a crap ton of approximations to desperately avoid the accurate simulation of light? Probably.
It's stupid though.
Eventually it becomes cheaper to do the right thing, both in terms of human cost to develop the hacks and compute cost to run them all. And our hardware vendors have decided to pay the engineering and silicon cost of dedicated RT hardware to dramatically accelerate this eventuality.
I mean, it's hilarious to see people complaining about a 30-50% performance dip (especially when you have reconstruction techniques to get that back, and then some). The cost was considered to be
100x only a few years back. If you ever took a Computer Graphics undergrad course you would know that real time ray tracing is a 30-year old dream for the field. Everyone knew the cost. Everyone knew a theoretical RT accelerator could narrow the gap, but researching, developing and productizing it into a profitable commercial product seemed unachievable. It wasn't just about the hardware, you need an entire ecosystem for the thing to survive.
But anyway, it happened. And all those people who did take that Computer Graphics 101 course and went on to have careers in game development houses, hardware vendors and platform vendors are now seeing this unfold. It's unreal.
Don't get me wrong, approximations will stay for a *long* time. But we will be slowly switching gears into *different* kinds of approximations. Instead of talking about screen space effects and cascaded shadow maps we'll be talking about denoising, importance sampling, better image reconstruction. And probably many other techniques that I'm not aware of. But these approximations are all centered around the fundamental operation of tracing
some number of rays through world space, which is now achievable in real time. And unlike screen space hacks, they don't break down catastrophically but degrade gracefully.
These are good times.
(PS: just don't think about Moore's Law.)
(PPS: disclaimer: CG is not my primary field so I may have butchered a bunch of stuff. I beg forgiveness from the real gurus here.)
(PPPS: the cake is a lie.)
