The real-time GI is amazing. Ray tracing is going to be a game changer, not only in terms of visual quality, but also because it's one of those rare times in computer graphics where the more advanced solution is actually making it easier on graphics programmers by replacing a bag of dirty tricks with a unified, physically based solution.
One interesting note: they say the GI is "instant" but you can actually see that they are using temporal stabilization to achieve this effect, and there's a slight lag between when the light changes and when the GI finds its resting point. I suspect removing this lag will be one of the things which makes graphics feel "next gen" when GPUs can handle enough rays per pixel to handle GI in one or two frames instead of a few dozen.
I tried to figure out whether the new GI used raytracing, and it doesn't seem like it is? If it is using raytracing hardware it isn't nearly as impressive. If it isn't, then it's sheer magic.
Feels to me like: https://www.gdcvault.com/play/1026182/ (which Nvidia has wrapped up in their RTXGI library: https://developer.nvidia.com/rtxgi). The idea is to supplement basic irradiance probes with moment shadow maps, so you can efficiently calculate how much each pixel should be affected by each probe. This greatly reduces the light bleeding problem, which is the biggest downfall of irradiance probes. Nvidia uses raytracing to update some percentage of the probes each frame, and I wouldn't be surprised if UE5 does, too (despite the Euro Gamer article). That said, their geometry solution is likely more amenable to random access than traditional rasterization, so that may make raytracing for probes and probe shadow maps unnecessary. Also, judging by when the player character puts her hand on the door, it looks like there's a bit of screen space GI going on, as well.
It certainly isn't completely raytraced, because current hardware can't support that. If they use the new hardware, they're using it in some kind of mixed mode. Whether that means only for certain materials, or whether they're able to use it to do "fat rays" that approximate only the general direction for indirect light, or whatever else. But direct, hard shadows are quite cheap (comparatively) on today's regular, rasterization-based rendering systems, so that part isn't surprising.
It certainly isn't fully path traced, but the result does look to me like they are using ray-traced GI. You can tell by some of the details, like colored bounce lighting which I am not sure can be achieved with this level of detail using other methods.
The current standard for getting real-time performance for ray tracing is to limit the number of rays and bounces, and stabilize the results temporally.
I.e. to get a "physically accurate" result you would have to send hundreds, or thousands of rays per pixel, and bounce them up dozens or hundreds of times. With this method, instead you send maybe even one or two rays per pixel, which gives you a noisy result. But you store the result, and accumulate it over a number of frames, and apply de-noising, and over time you end up with a high-quality result.
I believe that is why when the light moves, the GI lags for a fraction of a second in this demo.
Another trick is in stead of tracing multiple bounces per ray at once, to compute one indirect bounce every n-th frame using intermediate results from the previous (not the current) frame and then spatio-temporally smooth the results, possibly with on-screen bilateral filtering of the past few frames. I think one can see they do this when they move the light source. It takes a while for the indirect bounces to fade out. Perhaps they use screen space ray tracing for this (meaning no indirect bounces for occluded geometry).
Because softer shadows add to the realism, so they would have used them if they had invented a new efficient ray tracing method! Look at 5:12 in the video. The shadows are way too sharp. It's probably good old shadow mapping for the first bounce and then screen space(?) global illumination.
I'm almost certain they are. HW raytracing is of the big selling points of the new console generations, and if they achieved these kinds of results without it, then I would have to reconsider my skepticism of the occult.
For some tasks, ray tracing already outperforms traditional rasterization.
All "photorealistic" 3D renderers are trying to produce output on par with a true ray tracing renderer, and because the computation required is so high, tons of bodges and hacks are employed to approach that level of quality. The resulting complexity of a traditional rasterizer is astronomical compared to the complexity of a ray traced renderer.
Ray traced graphics has been the goal the entire time computer graphics have been a thing; for the past 40 years, at least. That's always been the goal. I can't remember a time when it wasn't.
It's only now that hardware fabrication technology has put us in a place where we're able to slightly open the door to the rendering techniques we've desired for so long. It's an exciting time.
It is not a scam. It may not perform well in most situations currently, but I can promise you, that will change significantly as time goes on.
Just like how early CPU hardware was slow and expensive, early ray tracing hardware is slow and expensive. Over time, CPU hardware has gotten cheaper and faster, and the same will happen with ray tracing hardware.
Real-time ray tracing is not impressive in the current generation of games because they’re optimized to maximize the quality of traditional lighting solutions. Most games use mostly static assets and lighting, because baked lighting can get a nearly photorealistic result as long as nothing moves.
Highly dynamic environments are where the technology shines, because it can achieve results which traditional solutions simply can’t. I think ray-traced Minecraft is the best current example of the potential of this technology.
It seems to me that this game doesn't actually take in account all the seemingly existing light sources when computing the shadows: see for instance at around 7:45, there is an obvious light source in the bottom right region (actually a bunch of synchronized pulsating lights on some sort of cylinders) which clearly should affect the surfaces close to it, but it doesn't in either mode! In the next comparison, there's a fire burning close to a wall, and likewise the fire doesn't seem to emit any light on the wall. Actually, the commentators do mention that there isn't that much dynamic lights handled in the game, limiting the effect of RTX.
I don't know how that RTX tech is working, but from what I've seen of the port of Quake2, it looked much more convincing - did you see it before making that judgement?
Current games that use raytracing often only use it for 1 light. For example, Metro Exodus only uses it for the sun. This means parts that take place underground look virtually identical with or without RT turned on. It's possible that they do the same thing in this game
Oh, I see. Wouldn't that make sense that each surface would be associated with its closest light source? It's still a rough approximation, but I guess it's better than one global light source?
IMO, ray tracing is a scam as currently presented by Nvidia, et. al. GPUs are not the correct way to approach this problem. RTX is an approximation at best and a complete joke at worst. That said, I do believe ray tracing is the future and can become the standard way in which we engage with all 3D graphics concerns.
Contrast the GPU with the newest generation of x86 CPUs. They have more cores than you can typically use, provide ridiculous amounts of pipelined throughput, and have specialized vector instructions which can dramatically accelerate common 3D graphics tasks (i.e. matrix multiplication).
Most practical ray tracing algorithms are trivially parallelized (e.g. just throw each scanline at a thread) and also exhibit properties which can leverage the deep stack depths and OoO execution enabled by modern CPUs. Ray tracing is inherently a recursive activity with a profound number of potential branching opportunities. x86 has absolutely no problem dealing with this kind of scenario. This is exactly what it was built for. GPUs on the other hand require yet more specialized hardware that has its limitations baked-in at the transistor level. x86 does not have these same kind of limitations.
Consider the hypothetical benefits of having your entire graphics pipeline implemented within 1 cache-coherent memory domain and on top of a single instruction set. Imagine no Direct3D/OpenGL/Vulkan/drivers/etc are involved to ruin your day. What if you could fit all of your required scene, texture and model data into L3 cache? How many times can you fill L3 from RAM per second on a 3950x with reasonable memory configuration? How many times per frame at 60fps? I feel we need to take some time to look at the emerging opportunities with the new hardware that is coming to market.
I worry that the developer community is so out of touch with the hardware aspect (e.g. 22ms hello world) that we are basically saying "lol no just use GPU magic graphics rectangle" and calling it a day. I feel like specialization of ASICs is fundamentally taking us in the wrong direction now that we are able to put so many general purpose cores onto a single package/die in a very cost-effective manner. Maybe this is why AMD is taking so long to bring a proper Nvidia killer to market. At some point there has to be a certain # of cores where someone raises their hand and asks "why do we still need a separate GPU?".
I don't know the implementation details of this, but I use both a 3950x and a 2080 Ti for rendering in Blender's Cycles, which is a non-realtime path-traced engine. I've experimented with several different configurations, and the GPU using its raytracing cores is much faster than the CPU for identical results. Given that next-gen cards are expected to increase the number of these cores by 2x or 4x, I don't see how software raytracing could compete.
One interesting note: they say the GI is "instant" but you can actually see that they are using temporal stabilization to achieve this effect, and there's a slight lag between when the light changes and when the GI finds its resting point. I suspect removing this lag will be one of the things which makes graphics feel "next gen" when GPUs can handle enough rays per pixel to handle GI in one or two frames instead of a few dozen.