Epic's Unreal Optimization Disaster | Why Nanite Tanks Performance!

Summary

This video challenges the popular belief that Unreal Engine's Nanite technology improves performance, using real-world tests to show that it can actually slow down performance, especially on smaller meshes and in complex scenes with high overdraw. The speaker critiques Nanite’s Auto LOD algorithm and highlights the inefficiency of virtual shadow maps and Lumen in certain scenarios. Proposing an AI-driven solution for optimizing topology and reducing overdraw, the video advocates for more efficient mesh handling techniques, and calls for a shift from reliance on Nanite to better performance-focused workflows.

Takeaways

• 😀 Enabling Nanite does not always result in better performance, especially on meshes with fewer triangles (e.g., 1,000).
• 😀 A scene with 6 million triangles ran 50% faster without Nanite due to better GPU quad utilization.
• 😀 Overdraw is a key issue with Nanite, where unused GPU power is wasted on triangles too small for efficient rendering.
• 😀 LODs (Level of Detail) reduce performance issues by merging smaller triangles as objects move further from the camera.
• 😀 Temporal aliasing, like pixel crawl and shimmer, is directly linked to overdraw and can impact visual quality.
• 😀 Nanite's overdraw increases the surface area shading cost, which results in slower performance and higher GPU resource consumption.
• 😀 Traditional mesh optimizations, such as merging meshes and instancing, are not supported efficiently with Nanite in Unreal Engine 5.
• 😀 Virtual Shadow Maps (VSSM) are slower on traditional meshes but work well with Nanite meshes, contributing to performance inconsistencies.
• 😀 There is a misleading trend in Unreal Engine forums and YouTube videos, where enabling Nanite on skeletal meshes is presented as a performance booster without proper LOD management.
• 😀 A new AI-driven solution that optimizes photogrammetry scans and LOD management could potentially provide better performance than Nanite, with fewer visual issues.
• 😀 Modern upscaling technologies like DLSS can mask the performance impact of overdraw, but they also introduce new visual problems and are becoming less effective as the resolution improves.

Q & A

What is the main focus of the video?

The video focuses on analyzing the performance differences between Nanite and traditional LOD systems in Unreal Engine, highlighting how Nanite may not always be the performance boost developers expect, and addressing common myths surrounding its use.

What is the myth about Nanite performance that the video debunks?

The myth is that Nanite always improves performance. The video debunks this by showing that Nanite can actually perform slower than traditional LODs, especially on smaller meshes with fewer triangles, despite claims to the contrary.

Why did the 6 million poly mesh scene run faster without Nanite?

The 6 million poly mesh scene ran faster without Nanite because of a phenomenon called 'quad overdraw,' where the GPU wastes resources on overdraw, which Nanite exacerbates. Without Nanite, the overdraw was more controlled, leading to better performance.

What does 'quad overdraw' mean and how does it affect performance?

Quad overdraw refers to the GPU wasting resources on rendering pixels (quads) where multiple objects or triangles overlap. This inefficiency can lead to significant performance loss, especially in complex scenes with lots of small triangles, as seen with Nanite.

How do LODs help reduce performance issues caused by overdraw?

LODs (Level of Detail) help reduce overdraw by using lower-polygon models for objects further away from the camera, thus improving GPU efficiency. As objects move farther away, their geometry becomes less demanding on the GPU, reducing overdraw and enhancing performance.

What is the significance of temporal aliasing in the context of overdraw?

Temporal aliasing, such as pixel crawl and shimmer, is a visual issue caused by overdraw. Optimizing overdraw can reduce these aliasing effects, improving both performance and visual quality by decreasing the need for flawed anti-aliasing techniques.

What is the difference between Nanite and traditional LOD systems in terms of shadow mapping?

Nanite uses virtual shadow maps, which are slower on traditional meshes and can have casting issues at certain distances. Traditional shadow maps are faster on simpler meshes, but when combined with Nanite’s overdraw, they can overwhelm the GPU.

Why is the use of auto LODs in Unreal Engine problematic for certain meshes?

Auto LODs in Unreal Engine are problematic because they may not effectively manage overdraw for meshes with low polygon counts. For example, a 450,000-triangle mesh with only four LODs might not adequately address overdraw, leading to wasted GPU power and poor performance.

What optimization technique does the video suggest to improve performance with Nanite?

The video suggests combining meshes that are likely to be loaded into memory together, applying efficient LOD management, and reducing draw calls. This could be done by merging meshes intelligently, ensuring they share similar LODs and distances, thus minimizing overdraw.

How does AI-based optimization compare to traditional LOD methods in terms of performance?

AI-based optimization has the potential to offer better performance than traditional LOD methods by analyzing photogrammetry scans and topology to create efficient LODs that reduce overdraw. This approach would avoid the manual labor and inefficiency of traditional methods and could outperform Nanite in certain cases.