4 ms·

> CPU or GPU but obv favours gpu Why obviously?

by virtualritz 2y ago

MyOutfitIsVague 2y ago
Rendering is significantly faster and more energy efficient on a GPU.
- fc417fc802 2y ago
  That only applies to video game level stuff. Most "real" scenes from professional movies won't fit in RAM on a typical desktop, let alone the extremely limited VRAM on a GPU. For example (from 2016, so 10 years out of date): https://www.disneyanimation.com/data-sets/?drawer=/resources/moana-island-scene/ https://www.disneyanimation.com/data-sets/?drawer=/resources...
- virtualritz 2y ago
  That's a prevalent misconception among people outside of offline rendering/VFX. The short version is: no, it is not. For this reason render farms for the CGI you see in blockbusters and series nowadays are usually CPU-only, still. How do I know? I work in that industry.
  - MyOutfitIsVague 2y ago
    That's quite surprising to me, as a regular hobbyist Blender user who sees at least 4x speedups on GPU vs CPU. Is it because of VRAM limitations on huge scenes? I assumed that big render farms used rigs with many GPUs with some proprietary technology to share memory between them to be able to fit the massive scenes. Or are you actually responding directly to what I was saying, and it turns out that on big scenes, even apart from the memory use, CPU rendering is somehow more efficient than GPU rendering? I don't think I'm prepared to accept that claim without a lot more substantiation. edit: I found this thread here: https://news.ycombinator.com/item?id=25616372 https://news.ycombinator.com/item?id=25616372 I guess the conclusion is that it is really about memory use, but at least 4 years ago, there wasn't a total consensus on the topic.
    - virtualritz 2y ago
      There are many aspects to this. The biggest is indeed memory. Scene complexity if VFX has always eaten hardware advances. Average frame time used to be 2h throughout my career and it never got smaller, however fast hardware became available (and this includes GPUs). And because scene complexity was always larger than what fit on GPUs for most stuff this was always the limiting factor in how effective they could be. Another issue people in these debates ignore is R&D. The amount of R&D you have to spend to make cutting edge offline rendering work perfectly on the moving target that is GPUs is insane, compared to CPUs. To give you an idea, look at the list of limitations RenderMan XPU (CPU+GPU) has compared to its CPU counterpart (nowadays called Ris) [1]. If you're A CG supervisor that page is a giant red flag for relying on XPU in your production in any way because any limitation you hit has to be worked around. And the biggest expense in production is human labor. Pixar has sunk probably dozens of man years of R&D into XPU (I think it was announced at Siggraph 2017). I.e. that's seven years times number of developers. In a shootout, a year ago (with a scene simple enough to fit on a GPU, mind you), the noise levels of the RenderMan XPU render were marginally better that those of 3Delight, for the same render time.[2] And that's after seven years of R&D and with the aforementioned limitations (none of which apply to 3Delight or to RenderMan Ris but the latter had unacceptable noise levels if you look at the images). 3Delight only supports CPU and their team of just half a dozen people has spend 100% of their time, sinc3 1999, on that. Instead of chasing GPUs. If you have complex scenes or volumes, 3Delight will beat any GPU renderer (and any CPU renderer anyway). I'm happy to bet on this. On a side note, my guess is that GPU-related R&D was a factor in what killed Clarisse. They even successfully ported OSL to GPU -- laudable! But briefly after the company went under ... Maybe their time would have been better spent to make the renderer more robust and faster on CPU instead. Just a guess ofc. This is the high-end stuff. There are lots of GPU path tracers out there and they're all great as long as your scene complexity doesn't cross a certain threshold. And as long as you do not need OSL support, I think only Evee has this. [1] https://rmanwiki-26.pixar.com/space/REN26/19661981/XPU+Features+and+Limitations https://rmanwiki-26.pixar.com/space/REN26/19661981/XPU+Featu... [2] https://github.com/tristan-north/renderBench https://github.com/tristan-north/renderBench
      - MyOutfitIsVague 2y ago

> CPU or GPU but obv favours gpu Why obviously?

MyOutfitIsVague 2y ago

Rendering is significantly faster and more energy efficient on a GPU.

fc417fc802 2y ago
That only applies to video game level stuff. Most "real" scenes from professional movies won't fit in RAM on a typical desktop, let alone the extremely limited VRAM on a GPU. For example (from 2016, so 10 years out of date): https://www.disneyanimation.com/data-sets/?drawer=/resources/moana-island-scene/ https://www.disneyanimation.com/data-sets/?drawer=/resources...
virtualritz 2y ago
That's a prevalent misconception among people outside of offline rendering/VFX. The short version is: no, it is not. For this reason render farms for the CGI you see in blockbusters and series nowadays are usually CPU-only, still. How do I know? I work in that industry.
- MyOutfitIsVague 2y ago
  That's quite surprising to me, as a regular hobbyist Blender user who sees at least 4x speedups on GPU vs CPU. Is it because of VRAM limitations on huge scenes? I assumed that big render farms used rigs with many GPUs with some proprietary technology to share memory between them to be able to fit the massive scenes. Or are you actually responding directly to what I was saying, and it turns out that on big scenes, even apart from the memory use, CPU rendering is somehow more efficient than GPU rendering? I don't think I'm prepared to accept that claim without a lot more substantiation. edit: I found this thread here: https://news.ycombinator.com/item?id=25616372 https://news.ycombinator.com/item?id=25616372 I guess the conclusion is that it is really about memory use, but at least 4 years ago, there wasn't a total consensus on the topic.
  - virtualritz 2y ago
    There are many aspects to this. The biggest is indeed memory. Scene complexity if VFX has always eaten hardware advances. Average frame time used to be 2h throughout my career and it never got smaller, however fast hardware became available (and this includes GPUs). And because scene complexity was always larger than what fit on GPUs for most stuff this was always the limiting factor in how effective they could be. Another issue people in these debates ignore is R&D. The amount of R&D you have to spend to make cutting edge offline rendering work perfectly on the moving target that is GPUs is insane, compared to CPUs. To give you an idea, look at the list of limitations RenderMan XPU (CPU+GPU) has compared to its CPU counterpart (nowadays called Ris) [1]. If you're A CG supervisor that page is a giant red flag for relying on XPU in your production in any way because any limitation you hit has to be worked around. And the biggest expense in production is human labor. Pixar has sunk probably dozens of man years of R&D into XPU (I think it was announced at Siggraph 2017). I.e. that's seven years times number of developers. In a shootout, a year ago (with a scene simple enough to fit on a GPU, mind you), the noise levels of the RenderMan XPU render were marginally better that those of 3Delight, for the same render time.[2] And that's after seven years of R&D and with the aforementioned limitations (none of which apply to 3Delight or to RenderMan Ris but the latter had unacceptable noise levels if you look at the images). 3Delight only supports CPU and their team of just half a dozen people has spend 100% of their time, sinc3 1999, on that. Instead of chasing GPUs. If you have complex scenes or volumes, 3Delight will beat any GPU renderer (and any CPU renderer anyway). I'm happy to bet on this. On a side note, my guess is that GPU-related R&D was a factor in what killed Clarisse. They even successfully ported OSL to GPU -- laudable! But briefly after the company went under ... Maybe their time would have been better spent to make the renderer more robust and faster on CPU instead. Just a guess ofc. This is the high-end stuff. There are lots of GPU path tracers out there and they're all great as long as your scene complexity doesn't cross a certain threshold. And as long as you do not need OSL support, I think only Evee has this. [1] https://rmanwiki-26.pixar.com/space/REN26/19661981/XPU+Features+and+Limitations https://rmanwiki-26.pixar.com/space/REN26/19661981/XPU+Featu... [2] https://github.com/tristan-north/renderBench https://github.com/tristan-north/renderBench
    - MyOutfitIsVague 2y ago