Unreal Engine 5, [UE5 Developer Availability 2022-04-05]

[Andrew Lauritzen]

Doesn't this come with the downside of larger assets and filesize bloat?

Obviously one should still try and remove detail that is literally impossible to see in a shipping game, but Nanite's compression is quite good, and being able to rely on instancing more heavily cuts down on sizes a lot too (vs. creating variants of pre-combined kitbashed stuff). Versus UE4 stuff I think you generally can achieve both a bump in quality and mild reduction in asset sizes with Nanite. There's some example stats for some of the demos in the Nanite docs I believe. Also if memory serves the textures (BC compressed) have still tended to dominate the asset sizes in the demos to date.

Would Nanite work under DX11?

Yes Nanite works in DX11 via vendor extensions. There's an increasing number of reasons not to use DX11 though; in various places performance is becoming quite poor in DX11 due to the lack of wave ops, good transient memory allocator and more, not to mention lack of hardware raytracing support.

 

[JoeJ]

Somebody tried to shed some light on Lumen:
https://blog.en.uwa4d.com/2022/01/25/ue5-lumen-implementation-analysis/
Almost got it, maybe.
But still wondering about SS probes and their palcement. I assume it's irregular and similar to Frosbites Surfels.

 

[zed]

Somebody tried to shed some light on Lumen:
https://blog.en.uwa4d.com/2022/01/25/ue5-lumen-implementation-analysis/
Almost got it, maybe.
But still wondering about SS probes and their palcement. I assume it's irregular and similar to Frosbites Surfels.

Very nice site, never seen it before, cheers I'll be checking that out when I have time

 

[DavidGraham]

Somebody tried to shed some light on Lumen:
https://blog.en.uwa4d.com/2022/01/25/ue5-lumen-implementation-analysis/

The blog seems entirely focused on the software based tracing part of Lumen, which makes it somewhat outdated, I would have loved a detailed dive into the hardware portion of Lumen, as this is the latest and most advanced iteration (as seen in The Matrix demo).

 

[Dampf]

The blog seems entirely focused on the software based tracing part of Lumen, which makes it somewhat outdated, I would have loved a detailed dive into the hardware portion of Lumen, as this is the latest and most advanced iteration (as seen in The Matrix demo).

I don't think that version of Lumen is available to the public with the currently released UE5 EA builds.

 

[JoeJ]

The blog seems entirely focused on the software based tracing part of Lumen, which makes it somewhat outdated, I would have loved a detailed dive into the hardware portion of Lumen, as this is the latest and most advanced iteration (as seen in The Matrix demo).

I wonder if HW RT allows them to reduce the complexity of the system, or if it adds even more.
It's unlikely a complete replacement? The argument of 'RT can't handle full Nanite detail' is not relevant, as even low lod proxies would be more accurate than SDF volumes.
But RT is basically point sampling, so it could not easily replace voxel cone tracing one might think.
Though, iirc they trace 8 x 6 cones for a voxel probe, which is a lot and the cones end up quite narrow. Getting rid of SDF and voxels would feel very attractive to me.

It's interesting that Lumen seems to carry over a lot of the VCT approach which was promised for UE4 but did not made it finally.

 

[DavidGraham]

I wonder if HW RT allows them to reduce the complexity of the system, or if it adds even more.
It's unlikely a complete replacement? The argument of 'RT can't handle full Nanite detail' is not relevant, as even low lod proxies would be more accurate than SDF volumes.

I don't know if you have been following the latest updates or not, but Lumen is fully hardware accelerated now, this allowed it to render precise GI with more quality (indirect and diffuse), allowed widespread RT reflections, enabled area light shadows and allowed much higher performance too! Enough that the Matrix demo runs well on Xbox Series S.

 

[Andrew Lauritzen]

I don't think that version of Lumen is available to the public with the currently released UE5 EA builds.

Pretty much the latest and greatest (only a few days old usually) is available to the public in the UE release stream on github if you build it yourself. Of course YMMV with a bleeding edge stream like that as there will definitely be bugs and such, but if anyone wants to run the latest or just look at the code, it's all publicly available (one of the great things about working here IMO).

 

[JoeJ]

I don't know if you have been following the latest updates or not, but Lumen is fully hardware accelerated now, this allowed it to render precise GI with more quality (indirect and diffuse), allowed widespread RT reflections, enabled area light shadows and allowed much higher performance too! Enough that the Matrix demo runs well on Xbox Series S.

No, did not follow anything gfx related recently, so thanks for the update.
Would be interesting to investigate further, but i better get back under my cozy rock of ignorance to get done some time.
When i get back to sunlight, i hope the supply situation has improved as well...

Series S level Steambox 2.0 would be nice, Gabe!

 

[Dampf]

https://compusemble.com/insights/ho...th-fast-storage-to-reduce-memory-requirements

Not directly related to UE5, but a really interesting demonstration nonetheless.

I do wonder, can UE5 compete with this technique? The things we have seen so far in regards to Virtual Texturing+World Partition are incredible, but 2 GB used VRAM with 300 GB's worth of texture data sounds even more impressive.

I suppose there is even more possible with modern hardware, a combination of UE5's techniques and Sampler Feedback Streaming + DirectStorage could set the bar even higher and leverage hardware to its full extent. Amazing times ahead!

 

[PSman1700]

https://compusemble.com/insights/ho...th-fast-storage-to-reduce-memory-requirements

Not directly related to UE5, but a really interesting demonstration nonetheless.

I do wonder, can UE5 compete with this technique? The things we have seen so far in regards to Virtual Texturing+World Partition are incredible, but 2 GB used VRAM with 300 GB's worth of texture data sounds even more impressive.

I suppose there is even more possible with modern hardware, a combination of UE5's techniques and Sampler Feedback Streaming + DirectStorage could set the bar even higher and leverage hardware to its full extent. Amazing times ahead!

3080Ti teamed to a R5 3600 Impressive to say the least anyway, ultra fast nvme drive/speeds in there, with some amazing technologies to back that up.

 

[pjbliverpool]

Even more impressive that this was seemingly without DirectStorage and with a very modest bandwidth requirement.

 

[cwjs]

I do wonder, can UE5 compete with this technique?

Would they need to? Ue5's texturing is relatively standard virtual textures, and fully compatible with approaches like this, right? (haven't looked into the texturing side at all)

 

[Dampf]

Would they need to? Ue5's texturing is relatively standard virtual textures, and fully compatible with approaches like this, right? (haven't looked into the texturing side at all)

Yes I suppose that is the case. But I'm not entirely sure. After 2 years they still haven't made a statement whether or not they plan to integrate DX12U features into Unreal Engine.

Because of that (and the lack of DX12U features in UE4), I'm not hopeful DX12U will ever come to UE5 and thus I regard virtual texturing as their competing technology.

 

[function]

I'm optimistic that if there's an advantage to using SFS, Epic will integrate it into their virtual texturing system.

Granular, accurate and fast streaming of textures with minimal footprint seems to be their goal and if this offers a benefit I expect it'll be integrated when the whole cost / time / customer HW support equation comes into balance.

 

[Remij]

Nice to finally see a demo we can run ourselves.

I hope when DirectStorage releases there's some demos and benchmarks that utilize it and show people very clearly what it's capable of. I'm sure the 3DMark team will come out with one at some point. I'd love to see a demo from them which absolutely thrashes storage bandwidth and throws around some insanely complex visuals.

 

[Nesh]

https://80.lv/articles/a-fantastic-...oGT_rJHnkwAoh0bJg4VZI29WsFTpPyy6QtF4_mSidnN5c

https://twitter.com/x/status/1483806969777508352

Thats mighty impressive

 

[Dampf]

Nice to finally see a demo we can run ourselves.

I hope when DirectStorage releases there's some demos and benchmarks that utilize it and show people very clearly what it's capable of. I'm sure the 3DMark team will come out with one at some point. I'd love to see a demo from them which absolutely thrashes storage bandwidth and throws around some insanely complex visuals.

I am very sure ULBenchmarks and Unigine are working on a DX12U benchmark as we speak.

It has been a while since their last ones. Netheless to say, these benchmarks utilizing all DX12U features as well as HW-RT to a greater extent and modern SSDs are going to look unbelieveable. Should be the biggest leap in graphics in a long while. I am very much looking forward to it!

 

[milk]

https://80.lv/articles/a-fantastic-...oGT_rJHnkwAoh0bJg4VZI29WsFTpPyy6QtF4_mSidnN5c

https://twitter.com/x/status/1483806969777508352

Thats mighty impressive

Is it really? Haven't we already seen small scale fluid sim demos like this running real time in GPUs preceding the PS4/One generation?

EDIT: case in point:

 

[Karamazov]

There will be some fluid simulation in HZFW in the PS5 version only. We'll see soon how well it's done.