NVIDIA RTX Spark: Blackwell + Grace, 128 GB Unified Memory

RTX Spark arrives as a new hardware foundation for game developers who want to run AI workloads locally. By combining a Blackwell RTX GPU with a Grace CPU and up to 128 GB of unified memory, the platform is designed to handle both graphics rendering and AI inference on the same machine. This memory setup preserves headroom for graphics rendering and local speech models, which matters for developers building AI characters that run on-device.

The software stack is equally central. DLSS 5 introduces 3D-Guided Neural Rendering as a final stage in the graphics pipeline, and Avalanche Software has already used its SDK controls to tune the integration for Hogwarts Legacy while maintaining the game's artistic intent. An October model update runs 2x faster than the September version and halves GPU overhead, with a 20% reduction in VRAM use at 4K, and it will ship by default in future DLSS 5 games. Separately, the DLSS 4.5 Unreal Engine plugin's version 8.9.0 adds official Windows on Arm support for UE 5.8, giving developers a supported path to bring DLSS-enabled projects to RTX Spark systems.

For AI characters, NVIDIA ACE introduces Nemotron 3.5 Bolt E4B 4GB, a small language model built to run locally on GeForce RTX and RTX Spark. The model delivers roughly 40% higher average quality against the prior model on relevant benchmarks and up to about 80% fewer output tokens versus peers. It supports multi-turn tool use, a 64k context window, and 35 languages. Developers can access it through the In-Game Inferencing SDK, and the SkyrimNET modding platform for The Elder Scrolls V: Skyrim offers a way to test it in a game. The overall picture is a platform play: NVIDIA is lining up hardware, SDKs, and middleware so that game studios can treat RTX Spark as a target for both neural rendering and on-device AI characters.