Nvidia Debuts DLSS 5 and Cosmos 3 Edge at SIGGRAPH

Nvidia Uses SIGGRAPH 2026 to Debut DLSS 5 and Put World Models on the Edge

A generative-AI rendering stage for games this fall, physics surrogates 10,000x faster than simulation, and a 4-billion-parameter Cosmos that runs on-device.

Jul 21, 2026Source: Nvidia

At its sponsored research keynote at SIGGRAPH 2026 in Los Angeles on Monday, July 20, Nvidia laid out its next act for graphics: generative AI inside the rendering pipeline, neural surrogates replacing physics simulation, and world models small enough to run on a robot.

DLSS 5: generation as a rendering stage

The headline for gamers is DLSS 5, shipping in fall 2026. Instead of only upscaling frames, it applies a one-step pixel-space diffusion transformer causally, frame by frame, as a "learned stage" after the renderer — enriching materials, lighting and detail while the game engine keeps authoring the world. Developers get three models and explicit controls: structure and tone intensity sliders, plus per-object and per-character adjustments to protect artistic intent. "The renderer keeps building the world exactly as the game has authored it," said Edward Liu, Nvidia's director of applied deep learning research. "The generation becomes the learned stage afterwards to enrich its appearance."

Physics in seconds, not days

Nvidia researcher Neil Ashton showed AI surrogate models producing answers on a single RTX 6000 that traditional solvers need hours or days for: an aircraft-airflow model predicting unseen geometries more than 10,000x faster at roughly 1-2% accuracy, trained on 2,000 simulations — 200 terabytes compressed into a ~200 MB checkpoint. Weather prediction, Nvidia claims, drops from "hours or days" to "seconds or minutes."

World models go on-device

The robotics push centers on Cosmos 3 Edge, a 4-billion-parameter world model that runs on Jetson Thor and RTX PRO hardware, with Agile Robots, Doosan Robotics, Siemens, Skild AI and others evaluating it. Alongside it: Cosmos Dreams, neural closed-loop simulators for training autonomous vehicles, and MotionBricks, a real-time motion model trained on 350,000+ clips and demoed live on a Unitree G1 humanoid.

The through-line

"Whether for games, cinema, robotics or factory digital twins, the goal is the same," Jensen Huang said in press materials: "create virtual worlds that behave with the fidelity and realism of the physical world."

Why It Matters

Nvidia is repositioning graphics from an end product into training infrastructure for physical AI — the same neural rendering that ships in games becomes the simulation engine for robots and digital twins. DLSS 5's per-object controls also mark the first serious attempt to reconcile generative AI with authorial intent in games, a fight every creative industry is about to have.