Unreal Engine 5 Revealed

[Tangled-Universe]

I guess I'll need a better graphics card. I can't see how a bunch of unoptimised Megascan assets can run on a low or mid-range card without severe bottlenecking.

Well, I don't know...?
You'd probably need a new graphics card anyway, by the end of 2021, but not for the reasons you just mentioned I'd guess.

No way you could load a single or few slabs of such rock geo and textures into your GPU nowadays.
From the video you can tell that they developed some kind of algorithm/optimization which allows for the use of unoptimized megascans assets.
My guess is that since megascan assets are available in several LOD's that they found a way to dynamically blend between LOD's over distance, kind of similar to TG's rasterized renderer.

Second reason is that this engine is supposed to work with Playstation 5 and some of those hardware specs are known/leaked by now (Ryzen and AMD GPU, I believe).

[KyL]

they found a way to dynamically blend between LOD's over distance, kind of similar to TG's rasterized renderer

My guess too. I remember reading about an automatic LOD generation in Unreal 4.something. But this looks like the next level. Probably using a form of GPU tessellation on the whole scene geometry...

Anyway we are still far from there. Games we see using raytracing are all doing the same thing: sharp mirror-like reflection. This is because at the moment it is the only thing that developers can afford to do otherwise it gets too expensive for the current hardware (especially with low-end platforms such as old generation consoles)

This demo probably gives us a glimpse of what game *could* be in probably 3 or 5 years. But I am pretty sure an extensive amount of work went into optimizing only to reach playable level, as always with real time.

Nevertheless this is still an impressive achievement and what I consider a true milestone. This feels like a real "next gen" now. What is the most impressive is the handling of scene geometry and I am really curious to know how they pulled that off!

[WAS]

This demo probably gives us a glimpse of what game *could* be in probably 3 or 5 years. But I am pretty sure an extensive amount of work went into optimizing only to reach playable level, as always with real time.

From what they were saying, not much was done to achieve it, they literally slapped in the assets on UE5, UE5 is handling it for the most part, right down to spacial sound. When the showed the meshes, you could see meshes rebuilding as the screen moved, which looks like normal screen-space LOD. They'll probably utilize shader optimization to do batch processing of geometry, while still shipping with some basic default LODs for those that can't wait for shader optimization processes.

While this is running on the PS5 and the new RDNA 2 AMD GPU, we can go back to UE4 and see some pretty amazing real-time stuff with only batch-processing megascans in houdini to achieve a target framerate above anything else. Since this is now handled by UE5, you can imagine you don't actually have to do a lot to get playable FPS, on a single 1080 TI.

[Tangled-Universe]

Very good information here on the combination of techniques they used to obtain these results.
It seems they indeed used different LOD's for geometry, but there's a lot more information about other aspects  in the video:

[Tangled-Universe]

A lot of the same, but now in text:

https://www.eurogamer.net/articles/digitalfoundry-2020-unreal-engine-5-playstation-5-tech-demo-analysis

[WAS]

The geometry scaling aspect was only really a matter of time. If I can bake out an 8k normal map on my GPU in less than a second now, there is no reason GPUs can't slave this process dynamically. I do wonder if there is some sort of predictive model to engage this feature though for smooth transitions. Looking at other demos, screen-space is kinda "poppy" (z-depth on camera poo maybe).

[amandas]

The geometry scaling aspect was only really a matter of time. If I can bake out an 8k normal map on my GPU in less than a second now, there is no reason GPUs can't slave this process dynamically. I do wonder if there is some sort of predictive model to engage this feature though for smooth transitions. Looking at other demos, screen-space is kinda "poppy" (z-depth on camera poo maybe).

It looks like UE5 is using screen space as a reference space for LODing objects at runtime, accompanied by tillable normal maps for detailed displacement. Much simpler idea than I thought. They are using old concept of micropolygons (perfected by 3Delight, which I remember had extra fast displacement rendering 15 years ago). I am guessing any popping may be due the algorithms used for LODing transitions and paradigms used (occlusion/distance culling or maybe some other in-house culling?).

[WAS]

The geometry scaling aspect was only really a matter of time. If I can bake out an 8k normal map on my GPU in less than a second now, there is no reason GPUs can't slave this process dynamically. I do wonder if there is some sort of predictive model to engage this feature though for smooth transitions. Looking at other demos, screen-space is kinda "poppy" (z-depth on camera poo maybe).

It looks like UE5 is using screen space as a reference space for LODing objects at runtime, accompanied by tillable normal maps for detailed displacement. Much simpler idea than I thought. They are using old concept of micropolygons (perfected by 3Delight, which I remember had extra fast displacement rendering 15 years ago). I am guessing any popping may be due the algorithms used for LODing transitions and paradigms used (occlusion/distance culling or maybe some other in-house culling?).

It's simply not possible to use screen space anything at run-time. Screen space is live, based on the viewport engine and runtimes are already running.