NVIDIA RTX Spark: Neural rendering, game AI characters

1. DLSS 5 enhances lighting and materials

RTX Spark is a platform for thin laptops and compact desktops. According to the company, it offers up to 128GB of unified memory shared by the CPU and GPU, supporting both graphics processing and local AI execution.

DLSS 5 neural rendering, which uses AI to enhance image details, supports this platform starting with NBA 2K27. According to the company, the October model update is 2x faster than September's and halves added GPU overhead. Visual fidelity, or the level of visual detail, remains similar. The company reports a 20% reduction in video memory usage at 4K.

Neural rendering uses AI to enhance lighting and material details in images rendered by the engine. Developers control what changes using artist-designed geometry, texture data describing surface patterns, and lighting data. NVIDIA said Avalanche Software integrated these controls to preserve the game's artistic intent and style.

2. Unreal Engine and Arm-based development paths

The DLSS 4.5 plugin officially supports Windows on Arm as of version 8.9.0.

For Unreal Engine 5.8, it provides native programs and app build and packaging paths for the processor. Developers can use these paths to prepare projects for the new platform.

Supported games can insert new frames between engine-rendered frames to smooth motion. According to the company, RTX Spark's Dynamic Multi Frame Generation adds up to 5 frames per rendered frame and adjusts the multiplier automatically.

Debugging tools have also been adapted for Arm-based development environments. Nsight Graphics 2026.3 analyzes frames to inspect rendering errors and GPU performance. It supports native apps and emulated apps that translate code intended for other processors.

Nsight Systems 2026.5 displays CPU and GPU tasks on one timeline to help identify bottlenecks that slow processing. Its unified Windows installer supports native and emulated apps.

Nsight Aftermath SDK 2026.3 records the GPU's state at the time of a crash and can be integrated into native apps on Windows on Arm.

NVIDIA RTX Branch of Unreal Engine 5.8 provides a foundation for ray tracing, which calculates light paths, and neural graphics. This update focuses on code cleanup, stability, and compatibility improvements. Developers can download the source from GitHub to modify and integrate it.

3. Running interactive characters on-device

NVIDIA ACE supports the development of in-game AI characters. PUBG introduced an AI companion based on a first-generation small language model in public beta. Aniimo uses a second-generation model for a voice-responsive in-game companion and is scheduled for release in early 2027.

According to the company, the released Nemotron 3.5 Bolt E4B 4GB delivers about 40% higher average quality than the previous model on relevant benchmarks, or performance comparison tests. The company also says it generates up to about 80% fewer output tokens than comparable models. NVIDIA highlights a design that leaves memory headroom for graphics processing and voice models.

The model performs tasks by calling application APIs multiple times during conversations, supporting interactions that reflect the game state. Developers can access it through the In-Game Inferencing SDK and configure short replies for voice conversations.

Local execution also accounts for memory needed by the game and voice models. Here reasoning means optional extra thinking on top of the model's own inference, which runs either way: developers can enable it for complex requests, or disable it or cap the amount of thinking to meet latency targets. Disabling or capping it changes how deep an answer goes and how fast it arrives, not whether the model answers at all. This means choosing a project-appropriate balance between character response depth and latency.

One use case, SkyrimNET, is a platform created by MinLL for Skyrim modding, the practice of modifying the game. It gives non-player characters personal memories, voices, and autonomy, the ability to decide and act independently. It enables conversations and context-aware quests, supporting both cloud and local inference.

4. Rendering technologies and character motion

RTX Kit supports this platform with 8 technologies for texture compression, complex scene processing, and hair and skin rendering, among other tasks. Neural texture compression uses AI to reduce storage while maintaining visual fidelity similar to conventional block compression. Tools also remove lighting noise or handle indirect light, which bounces off surrounding surfaces.

MotionBricks is an open-source research project that generates motion between target poses in real time. Rather than requiring developers to author each transition, it generates movements based on player input or game events. This update adds a human motion model, trained model files, and export tools.

MotionBricks is provided as research code and models, and the Unreal Engine integration shown in the research demonstration is not included in this release.

RTX Spark's key advance is its broader toolset for testing game graphics and local AI characters together in one development environment.