The article contains the benchmark from intel towards the end, showing how the light built time differs from processor to processor.
Sure, this doesn’t equate directly to “Intel does something special for unreal”, but at the same time unless you can find me some equally significant chart from an a AMD site, I’d go on assuming there are absolutely 0 optimization or concerns from their point of view.
Whereas, at the very least Intel puts some thought into it - be it at the base/core level or not.
Not that this fact helps the direct problem anyway.
If not by saying again: go bother AMD to optimize stuff for ue4.
You really should have run it for at least a couple of minutes, lightmass can initially drop utilization when skylight radiosity starts but it should quickly recover to max utilization… Why does this happen? No idea.
This is from one of my test scenes, it dips briefly and rebounds to 100% indefinitely:
Regardless I can only wildly speculate here. Maybe your cooling solution isn’t sufficient for your CPU and its thermal throttling. You mentioned you modified the grocery scene to make it more modular, perhaps the new lightmap sizes combined are significantly larger than the original huge mesh which used a 2k lightmap resulting in longer build times. You never mentioned what results you’re getting either, if the difference is fairly small maybe perhaps you’re not using the XMP profile for your memory. Maybe it really is the chipset drivers.
Or perhaps none of this matters because it is building on preview settings, maybe higher core counts just don’t scale as well in preview and the 3900x comes out ahead just because it has a higher base clock speed.
There are so many variables here that it is anyones guess at this point.
So an interesting update. I updated my bios and chipset but that really did not do anything. However, we moved our project to 4.26 to test the bake time and on the mercantile, on preview settings, the light bake had almost no change. I think that can be attributed to the single thread per mesh aspect you and others mentioned and the fact there is a very large single mesh in that scene.
On our scene, the light build was dramatically different and the cpu utilization went from 6% during the radiosity portion to 100% utilization throughout the entire build process! The total light build time went from 21 minutes to 6 minutes. Our scene is much larger than the Mercantile scene but the meshes, although not modular, are well broken up. We also do not have nearly the amount of static lights the Mercantile has which seems a bit much IMHO. some of those lights could have been handled emissively but whatever.
I am not sure what all of this means but at least we have better build times in our project. The one thing I also wanted to point out is that in another thread, it was mentioned that ultimately, Unreal Engine is a single thread application. As such, I just don;t think the 3950x is probably the best processor for it at this moment. 3950x is relatively new and maybe the newer version of 4.26 had better code or whatever to help the 3950x but who knows. Where the processor does really shine is in multithread applications which is why when I use it for Vray and other offline rendering, it really cooks.
None of this will probably matter since I heard that 4.26 has GPU light mapping which should make things faster. Also, with the advent of Unreal Engine 5, who knows where things will go.