Tech • IA • Crypto
NVIDIA Keynote Live at SIGGRAPH 2026
9.4/10TL;DR
NVIDIA unveiled DLSS 5 and the Cosmos AI platform at SIGGRAPH 2026, signaling a shift toward AI-augmented graphics, simulation, and robotics.
KEY POINTS
A “reset” in computer graphics
NVIDIA framed the current moment as a generational inflection point where graphics, AI, and simulation converge. The industry is moving beyond traditional rendering pipelines toward hybrid systems combining physics-based computation with learned models. This shift is positioned as foundational for future applications including digital twins, robotics, and virtual worlds.
DLSS 5 introduces generative rendering
The new DLSS 5 system extends earlier upscaling and reconstruction techniques into real-time generative enhancement. Instead of only reconstructing pixels, it uses AI to enrich rendered frames, adding realism such as improved lighting, materials, subsurface scattering, and reflections. The approach combines deterministic rendering with probabilistic generation to achieve higher fidelity.
Balancing realism with artistic control
A central challenge addressed by DLSS 5 is preserving artistic intent. The system uses renderer data like normals, albedo, and motion vectors to constrain AI outputs, ensuring consistency in character identity and scene composition. Developers retain control through adjustable parameters such as structure and tone intensity, as well as masking tools for selective enhancement.
Real-time constraints drive new AI design
DLSS 5 operates within strict performance limits, handling 4K frames in under 16 milliseconds. To achieve this, large generative models are distilled into compact, task-specific systems optimized for real-time inference. The result is a model that enhances visuals without introducing latency, maintaining responsiveness required for games and interactive applications.
A third pillar beyond reconstruction and simulation
NVIDIA positioned DLSS 5 as introducing “generation” as a third core category in graphics pipelines, alongside reconstruction and function approximation. This adds a new scaling axis: image quality can now improve not only with more compute in rendering, but also through better-trained AI models, even without increasing runtime cost.
AI accelerates scientific simulation
Advances were also highlighted in physics-based simulation, where AI models trained on large datasets can replicate complex systems such as climate modeling. Simulations that previously required tens of thousands of GPUs can now be approximated in seconds with comparable accuracy, enabling faster experimentation and broader accessibility.
Massive compression of simulation data
AI models can compress enormous datasets—for example, 200 terabytes of simulation data into ~200 megabytes—while retaining predictive accuracy within 1–2%. This enables rapid evaluation of new designs, such as aircraft aerodynamics, without rerunning expensive full-scale simulations.
Cosmos platform targets physical AI
NVIDIA introduced Cosmos, a “world foundation model” designed to generate physics-aware synthetic environments for robotics. The platform integrates vision, language, audio, and action into a unified architecture, enabling systems to simulate and reason about real-world interactions.
Compute becomes a source of data
Cosmos reframes data generation by allowing developers to create synthetic training scenarios at scale, rather than relying solely on real-world collection. This is critical for robotics, where real-world data is slow and costly to obtain. The approach supports simulation of diverse environments such as warehouses, kitchens, and driving scenarios.
Real-time simulation for robotics and autonomy
Demonstrations showed Cosmos models running on a single GPU, generating interactive simulations and controlling robots in real time. A system called Cosmos Dreams enables closed-loop simulation for autonomous vehicles, producing photorealistic environments that respond dynamically to control inputs.
Unified models for perception and action
Cosmos integrates multiple capabilities—world understanding, prediction, and control—into a single architecture. By using shared representations grounded in physical laws, the system supports tasks like forward dynamics, inverse dynamics, and policy learning across different robotic embodiments.
Open ecosystem and industry adoption
NVIDIA emphasized an open approach, releasing models, datasets, and training tools while collaborating with partners across industries. Early adoption spans robotics, autonomous vehicles, and video analytics, with millions of developers engaged in building on the platform.
CONCLUSION
The announcements signal a shift toward AI-driven pipelines where rendering, simulation, and physical systems are increasingly unified, redefining how digital and real-world environments are created and understood.