Slashdot Mirror
Crytek Shows 4K 30 FPS Ray Tracing On Non-RTX AMD and NVIDIA GPUs (techspot.com)
dryriver writes: Crytek has published a video showing an ordinary AMD Vega 56 GPU -- which has no raytracing specific circuitry and only costs around $450 -- real-time ray tracing a complex 3D city environment at 4K 30 FPS. Crytek says that the technology demo runs fine on most normal NVIDIA and AMD gaming GPUs. As if this wasn't impressive already, the software real-time ray tracing technology is still in development and not even final. The framerates achieved may thus go up further, raising the question of precisely what the benefits of owning a super-expensive NVIDIA RTX 20xx series GPU are. Nvidia has claimed over and over again that without its amazing new RTX cores and AI denoiser, GPUs will choke on real-time ray tracing tasks in games. Crytek appears to have proven already that with some intelligently written code, bog ordinary GPU cores can handle real-time ray tracing just fine -- no RTX cores, AI denoiser or anything else NVIDIA touts as necessary.
What nvidia means by "ray tracing" with their RTX thing and the AI denoiser is actually path tracing, which uses incoherent rays and actually does simulate light bounces in a physically accurate way. Effects like depth of field, soft shadows, caustics, ambient occlusion, and diffuse interreflection are a natural result of the path tracing algorithm, but have to be specially accounted for in other algorithms like ray tracing. A good reference for this is Physically-Based Rendering, by Matt Pharr. Because the rays in a path tracer are incoherent, it's an inherently noisy algorithm that requires many samples to reduce variance to acceptable levels. That's where the AI denoiser comes in - it's able to take a noisy image made with fewer path-traced samples and reduce variance to an acceptable level in realtime.
The guys over at brigade also have an actual realtime path tracer, and while the work is world-class and draw-droppingly impressive, you can see how noisy it still is.
As someone who works with raytracers you're spouting gibberish.
There are two components to raytracing: ray intersection and shading. No shit that Nvidia uses their SHADING hardware to shade raytraced rays. What they added was ray-intersection hardware which yes does take a good bit of processing power on non-trivial scenes.
Yes ray cohesion is important and no Nvidia didn't address it (like arguably Caustic\IMG did with their now-dead OpenRL raytracer) but they are in fact tracing rays.
As to raytracing needing "bouncy rays". You're confusing Global Illumination with "Ray Tracing". Ray tracing is just firing rays and returning the results. You are using a no-true-scotsman falacy to equate one with the other. If you raytrace only primary visible rays (like a rasterizer) you're still raytracing even if you have 0 bounces. And bounces is exactly where the RT cores do help boost the trace rate on the RTX GPUs.
As to why there are Tensor Cores? Because you already mentioned that you need millions of bouncing rays to deliver global illumination. Imagination Technologies almost delivered true ray counts high enough for GI but it was still too noisy. Nvidia without cohesion for its shaders said "how can we do global illumination without ray counts high enough?" and the answer was to apply a denoising neural net. So RT cores + a large tensor unit means they can denoise their low sample count raytracing into something useful.
As to "Rasterization can do anything raytracing can do but better". That's a load of bullocks. Raytraced shadows are infinitely more memory efficient. You also can't have self-reflections with reflection maps. There is a reason every single production renderer in existence today for high end visual effects is now a raytracer. Ray tracing is more efficient, it's faster and it produces far superior quality to rasterization hacks.