Unreal Engine 5 Documentation has been moved to the Epic Developer Community

GPU Lightmass (GPULM) is a light-baking solution that precomputes complex light interactions from lights which have their mobility set to Stationary or Static, and stores that data in generated lightmap textures that are applied to scene geometry. This system of baking lighting into textures is similar to the CPU-based Lightmass Global Illumination system. However, using the GPU to generate and build lighting data means that we can leverage the latest ray tracing capabilities with DirectX 12 (DX12) and Microsoft's DXR framework.

GPULM significantly reduces the time it takes to calculate, build, and generate lighting data for complex scenes, achieving speeds comparable to those of distributed builds using Swarm with the CPU-based Lightmass. Additionally, GPULM offers new workflows with interactivity that enables you to edit the scene, then recalculate and rebuild lighting on-the-fly. This workflow is not possible with the CPU-based Lightmass system.

Enabling GPU Lightmass

To enable GPU Lightmass in your projects:

1. Open the Plugins tab from the Editor > Plugins menu. Under the Built-in > Editor category, locate and enable GPU Lightmass .

2. Open the Project Settings window from the Editor > Project Settings menu:

   1. Under the Engine > Rendering category, enable the following:

      1. Ray Tracing > Ray Tracing

      2. Virtual Textures > Enable Virtual Texture Support

      3. Virtual Textures > Enable Virtual Texture Lightmaps

   2. Under the Platforms > Windows category, set the following:

      1. Targeted RHIs > Default RHI: DirectX 12

3. Restart the editor for these changes to take effect.

Additional Requirements for Setting Up GPU Lightmass

The following are additional suggestions to best work with GPU Lightmass in your projects.

Reducing GPU Timeout Detection and Recovery Crashes

In situations where you have a complex scene that can put the GPU under heavy load can cause a timeout delay (or TDR) and subsequent crash to occur within Windows 10. This type of crash impacts not only Unreal Engine but anything on your machine using the GPU. You will typically see a message similar to this:

It's possible to prevent, or reduce, occurance of these types of crashes by extending the time it takes for the GPU to timeout, giving it time to potentially recover without causing the engine to close.

See the Movie Render Queue's step for setting up your system configuration for details on how to adjust the timeout delay and recovery within Windows 10.

Disabling Real-Time Ray Tracing Features

GPU Lightmass leverages Microsoft's DXR API for ray tracing, which requires DirectX 12. While ray tracing is required for GPU Lightmass to work, it does not require the additional ray tracing features that are enabled by default, such as ray-traced shadows, ambient occlusion, and reflections. Instead, it's best to disable all of these features using the following console command:

r.RayTracing.ForceAllRayTracingEffects 0 

Alternatively, you can set this in your project so that all ray tracing features are disabled when it loads by adding them to project's DefaultEngine.ini configuration file located in your project's Config folder. Add the following under the [/Script/Engine.RendererSettings] section:

[/Script/Engine.RendererSettings]
r.RayTracing.ForceAllRayTracingEffects=0

Configuring GPU Memory

GPULM requires that there be enough GPU memory available to account for its overhead. With that in mind, the following considerations affect the success of using GPULM to bake complex scenes:

• There must be enough GPU memory to store the entire scene in memory at the lowest level of detail (LOD) mesh, that is, the highest quality LOD mesh.

  GPULM does not currently take LODs into consideration while baking. See the Limitations of GPU Lightmass section of this page for more details.

• The Virtual Texture system can require significant memory use during light builds. This largely depends on the complexity of your scene and its size.

• There must be enough CPU memory to store all generated lightmaps in RAM. The GPU can swap out lightmaps to the CPU's RAM, but lightmaps are not saved to disk until the entire light bake is completed.

• DX12 generally uses more GPU memory than DX11. If you have a scene that was pushing the limits of your GPU memory in DX11, you may find it difficult to use in DX12 with additional overhead from ray tracing and virtual texture requirements without some sacrifices.

• GPULM has its own memory usage requirements for optional settings, like Irradiance Cache .

Using Virtual Texture Lightmaps for In-Editor Previewing

Enabling the Virtual Texture system with Virtual Texture Lightmaps allows for lightmaps to be generated and stored as virtual textures. This has the added benefit of allowing for light builds to update in the level viewport in real time. It also enables editing as the scene is being built instead of having to cancel or wait for the build to complete.

Enabling Multiple GPUs for Building Lighting

Multiple GPUs to compute lighting for your project is supported when using a NVIDIA SLI-based GPU that also supports ray tracing. Multi-GPU support is enabled by doing the following:

• GPUs must be connected with NVLink bridges and SLI must be enabled in the NVIDIA Control Panel.

• In the Unreal Engine DefaultEngine.ini file located in your [Engine Install Path]/Engine/Config folder, under the [/Script/Engine.RendererSettings] section by adding r.AllowMultiGPUInEditor=1 .

• Pass the command line argument -MaxGPUCount=[n] — where n represents the number of GPUs available through SLI — when launching the editor. For example, -MaxGPUCount=2 would use two GPUs while in multi-GPU mode to compute lighting.

  Optionally, you can create a shortcut of the UE4Editor.exe and in the Properties settings, set the add -MaxGPUCount=[n] to the Target line.

Once the editor starts, you can confirm you are using Multi-GPU mode by opening the Output Log and searching the log for message:

 LogD3D12RHI: Enabling multi-GPU with 2 nodes

Performance when using multiple GPUs can decrease the build completion times on average. For example, in medium-sized test scenes using two RTX-2080TIs, build times were found to be about 1.7x faster on average when volumetric lightmaps are not heavily used. Improved speeds for build completion are dependent on a number of factors, such as scene size and complexity, engine feature multi-GPU support, and number of GPUs being utilized.

GPU Lightmass multi-GPU support is ideally suited for architectural and virtual production-type environments that rely on a single area. Large game-sized levels might encounter memory and virtual texture limitations depending on their complexity and the amount of VRAM available on the GPU. It is possible for architectural and virtual production environments to reach the same limitations as game-size levels if they have a lot of complexity to their design and set up. Multi-GPU limitations:

• Volumetric Lightmap calculation is not yet supported. Scenes that have a significant Volumetric Lightmap computation will see less overall performance gains.

• Irradiance cache is generated locally on each GPU. Depending on the content, you may have slight shading differences between different baked tiles. This can be addressed by increasing the Irradiance Cache Quality , or by disabling Use Irradiance Caching entirely.

• Texture encoding and denoising are run on the CPU, and are not affected by multi-GPU.

Using GPU Lightmass

GPULM has its own panel which can accessed through the Level Editor's toolbar under the Build dropdown menu by selecting GPU Lightmass .

The GPULM panel can be docked within the editor, similarly to other panels.

Once you have configured your settings, press Build Lighting to start a bake.

GPU Lightmass Baking Modes

GPULM includes two modes with which to bake lighting: Full Bake and Bake What You See (BWYS) .

• Full Bake mode renders the full lightmap resolution for every object in the scene as lighting is being calculated and built. The results aren't visible until the build completes.

• Bake What You See mode offers greater flexibility in that you only build lighting for what is visible within the viewport instead of the entire scene. It also enables interactivity during the light building process to recalculate lighting as changes are being made, when the viewport's real-time mode is enabled.

Both of these baking modes are view-dependent, meaning that they first prioritize objects within the view before moving onto other objects in the scene.

In Full Bake mode, GPULM uses the following flow:

• Every object within the scene is considered, and objects within the current view are prioritized.

• Lighting is baked in texture space at an object's full lightmap resolution, and completed lightmap tiles are sent to the Virtual Texture system to update the display.

• Once all objects have been calculated and a lightmap rendered, encoding takes place and saves the data automatically.

Bake What You See mode offers a non-destructive way of interactively working within a scene with lighting results being calculated (and recalculated) in real-time. It encourages on-the-fly adjustments in areas of your scene where lighting is actively being worked on. Because of this interactivity, the BWYS workflow for light builds differs from that of Full Bake as follows:

• The Runtime Virtual Texture system determines visible tiles at the mipmap level necessary to resolve on screen.

• Lighting is baked in texture space at the resolved mipmap resolution of each tile within the camera view.

• Once all on-screen tiles have completed, BWYS waits indefinitely for the scene or viewport camera to update.

• Pressing the Save button starts encoding the final lightmaps. A completed lighting build is not produced for the scene if you choose to save the results; only parts of the scene which were visible on screen have their lighting data ready to be stored and saved.

• Pressing Save does not end the light bake, it only saves any encoded lightmap data that has been generated up to this point. Instead, press Cancel to stop the process.

Interactive Baking

Interactive baking is readily available for both Full Bake and BWYS modes, prioritizing visible virtual texture tiles that have been mapped to the current camera view; moving the camera will re-prioritize tiles for the current view. While interactive baking is used for both light baking modes, it is most useful used in conjunction with BWYS mode.

Starting BWYS mode kicks off a lighting build that only completes when it is manually stopped. This mode only considers and builds lighting for objects within the camera view. Moving the camera around the scene, changing position or adding/removing objects from the scene, or changing properties of scene Actors cause lighting to be recalculated in real-time. Because this mode requires the lighting results to be manually saved, it is a non-destructive way to work if lighting has previously been built for the scene.

Using GPU Lightmass Speed Modes

One benefit of GPULM is that it leverages virtual texturing during light bakes. This enables lighting to be built and displayed in real-time in the editor viewport. It also works progressively if the scene is being edited at the same time and updates the scene lighting accordingly. The drawback to using in-editor previewing in real time is that there is additional overhead of re-rendering frames in the scenes to update the visible result. Disabling this overhead frees up the GPU to work more efficiently and quicker.

When the state of the Level Viewport's Realtime mode controls the two speeds at which GPULM operates. When Realtime mode is:

• On uses the slow mode because of the additional overhead of rendering frames in real-time.

• Off removes the overhead of rendering the scene in real-time which significantly increases the speed of lighting builds.

The Level Viewport Realtime mode can be toggled by using the:

• GPU Lightmass panel's realtime viewport checkbox to toggle the Realtime mode of the viewport.

• Level Viewport's dropdown menu to toggle Realtime mode.

• Keyboard shortcut Ctrl + R .

In addition to the Realtime mode toggle for the viewport, GPULM also includes settings to increase the speed of lighting builds using some built-in parameters.

• Realtime Workload Factor is a multiplier that can improve the speed of a lighting build when Realtime mode is toggled On in the viewport. Using too high a value can cause the editor to become unresponsive in scenes with lots of geometry.

• Non-Realtime Workload Factor is a multiplier that can improve speed of lighting builds when Realtime mode is toggled Off .

GPU Lightmass Settings

GPU Lightmass contains most of its settings within its own panel.

Under the World Settings panel, the only Lightmass setting that carries over from the CPU-based Lightmass system is the Volumetric Lightmap Detail Cell Size . It sets the size of a Volumetric Lightmap voxel at the highest density around geometry in the scene. Decreasing the size of the voxel increases the time it takes to build lighting and the amount of memory that gets used. In some cases, it can increase memory by up to a factor of eight.

Setting Up Lightmap UVs and their Resolutions

Baking lighting requires that each Static Mesh have its own Lightmap UV and that the UV charts (also called UV islands) are all contained within the 0-1 texture space and have no overlapping or wrapping pieces. This ensures that when lighting is calculated that no artefacts occur due to a bad lightmap UV.

Setting up a good Lightmap UV is the first step to getting a quality light bake for your object's geometry. Next, it's important to make sure that the lightmap texture applied to your geometry has enough resolution to capture all the necessary light and shadow detail. This means changing the lightmap resolution per object, as needed. You can do this in one of two ways:

• Set the Light Map Resolution of the object using the Static Mesh Editor .

• Select a Static Mesh Actor in the scene and, in the Details panel under the Lighting category, set a new lightmap resolution for this individual Actor using its Overridden Light Map Res property.

Controlling the Number of Light Bounces

The number of light bounces within any given scene is not controllable. Instead, GPULM uses a Russian Roulette algorithm, which takes into account various probabilities and weighting calculations to determine how many bounces any given ray being cast will take. This also means that light bounces unlikely to contribute to indirect lighting and illumination of the scene are more likely to be terminated.

Choosing a Denoising Option

GPULM relies on denoising methods provided by Intel's Open Image Denoise library to remove noise and smooth the results of the final lightmap renders.

There are three options to choose from when considering how you want your light bakes to be denoised. Use the Denoising dropdown to make your selection:

• None : No denoising is applied to the lighting render. This can be useful to determine a number of GI Samples to use for a given scene. For example, if you noticed artefacts being introduced by the denoiser, increasing the samples and possibly other settings, such as Irradiance Cache and First Bounce Ray Guiding, provide the denoiser with higher quality input.

• On Completion : Denoises the results of the light build on the CPU after rendering. The entirety of the scene's lightmaps are denoised when the light build is finished.

• During Interactive Preview : Denoises each virtual texture tile as it completes. This can be useful for seeing final results more quickly but is less efficient because it requires transferring each tile off the GPU to denoise on the CPU, and then back to the GPU to display.

The comparison below shows a completed light build with and without denoising applied.

Limitations of GPU Lightmass