4 ms·

Because a lot of the "cheap hacks" aren't even cheap. Famously, using real-time SSR to simulate ray-traced reflections can end up slower than ray tracing them f

Because a lot of the "cheap hacks" aren't even cheap. Famously, using real-time SSR to simulate ray-traced reflections can end up slower than ray tracing them from the jump, the same goes for high-res shadowmaps. Ambient occlusion is a lighting pass you can skip entirely if you render shadows the right way with RT from the start. If you keep stacking these hacks until you have every feature that a RT scene does, you're probably taking longer to render each frame than a globally illuminated scene would. Accelerating ray tracing in-hardware is a literal no-brainer unless you deliberately want to ostracize game developers and animators on Mac. I understand the reactionary "But I don't care about dynamic shadows!" opinion, but there's practically no opportunity cost here. If you want to use traditional lighting techniques, you can still render it on an RT-enabled GPU. You just also have the option of not wanting to pull out your fingernails when rendering a preview in Blender.

flohofwoe 2y ago
Yes, for some specific problems, raytracing definitely makes a lot of sense, but rasterization and 'geometry compression' via triangles also makes a lot of sense for other situations. I think the best approach will always be a hybrid approach instead of a 100% pure raytracing pipeline. The other question is how much of the raytracing features in 3D APIs need to be implemented in fixed-function hardware units, or whether this is just another area where the pendulum will swing back and forth between hardware and software (running on the GPU of course). But then maybe in 10 years we'll have 'signed distance field units' in GPUs and nobody talks about raytracing anymore ;)
- talldayo 2y ago
  From everything I've seen, hardware acceleration seems to be the only option. Software-accelerated RT existed for years and was mainly the reason why it was disregarded as a technique in the first place. Here's the OG M1 Pro, manufactured on TSMC 5nm, getting blown the fuck out by a 3070 Ti on Samsung 8nm: https://youtu.be/NgDTCNPm0vo https://youtu.be/NgDTCNPm0vo The architecture will probably improve on both Nvidia and Apple's side going forward, but in theory both GPUs should be easier to support for longer since they're not focused on accelerating obsolete junk like MSAA or HBAO. It enables a better convergence of features, it makes porting to-and-from Mac easier, and with any luck Apple might even quit being so sheepish about working with Khronos since they're not actively neglecting the Vulkan featureset anymore.

flohofwoe 2y ago

Yes, for some specific problems, raytracing definitely makes a lot of sense, but rasterization and 'geometry compression' via triangles also makes a lot of sense for other situations. I think the best approach will always be a hybrid approach instead of a 100% pure raytracing pipeline. The other question is how much of the raytracing features in 3D APIs need to be implemented in fixed-function hardware units, or whether this is just another area where the pendulum will swing back and forth between hardware and software (running on the GPU of course). But then maybe in 10 years we'll have 'signed distance field units' in GPUs and nobody talks about raytracing anymore ;)

talldayo 2y ago
From everything I've seen, hardware acceleration seems to be the only option. Software-accelerated RT existed for years and was mainly the reason why it was disregarded as a technique in the first place. Here's the OG M1 Pro, manufactured on TSMC 5nm, getting blown the fuck out by a 3070 Ti on Samsung 8nm: https://youtu.be/NgDTCNPm0vo https://youtu.be/NgDTCNPm0vo The architecture will probably improve on both Nvidia and Apple's side going forward, but in theory both GPUs should be easier to support for longer since they're not focused on accelerating obsolete junk like MSAA or HBAO. It enables a better convergence of features, it makes porting to-and-from Mac easier, and with any luck Apple might even quit being so sheepish about working with Khronos since they're not actively neglecting the Vulkan featureset anymore.