NVIDIA AI Updates: September 23, 2026

1. Isaac ROS 5.0 Ships With Agent-Ready Docs and Reusable Skills

NVIDIA. Released at ROSCon in Toronto, Isaac ROS 5.0 adds support for ROS Lyrical and Ubuntu 24.04 and introduces Isaac skills for setup and manipulation as reusable workflows that both humans and coding agents can invoke. NVIDIA also contributed a standard data-handling interface to ROS Lyrical through the Open Source Robotics Alliance, giving the roughly 1.3 million ROS users a consistent way to accelerate workloads across hardware, with CUDA as the reference implementation. Agent-ready documentation is an explicit deliverable, not a side effect. Source

2. The Zero-Copy Migration Comes With an Agent Skill That Does the Refactor

NVIDIA. The rosidl::Buffer abstraction plus NVIDIA’s contributed CUDA buffer backend let ROS 2 nodes exchange GPU-resident payloads over zero-copy transport while keeping standard messages and node boundaries intact. Every node in Isaac ROS 5.0 has been moved onto it. NVIDIA published a migrate-node-to-rosidl-buffer agent skill that audits a CUDA-accelerated node, traces data movement, and plans a minimal interface-preserving refactor, demonstrated on a Depth Anything 3 TensorRT node and verified with Nsight Systems by confirming no payload-sized host-device transfers at the ROS boundary. Source

3. NVIDIA Published a Methodology for Measuring Confidential Computing Overhead

NVIDIA. Running inference inside memory-encrypted confidential VMs with confidential GPUs and encrypted NVLink changes the assumptions behind memory movement, timing, scheduling, and multi-GPU communication, so frameworks need CC-aware adaptations to avoid paying an unnecessary penalty. The post walks through how TensorRT LLM adapts, and gives a controlled methodology teams can rerun: pick a workload with long input context, extended generation, and low concurrency, because high request volume amortizes fixed encryption costs and hides the effect. The reference config is DeepSeek-R1-0528-NVFP4 at 32K input and 1K output on TP=8 with FP8 KV cache. Source

4. Topograph Feeds Cluster Topology to Schedulers That Would Otherwise Guess

NVIDIA. Topograph discovers cluster topology from cloud APIs or on-premises fabric systems, normalizes it into a common model, and publishes it as Kubernetes node labels, Slurm configuration, or Slinky ConfigMaps so workload managers can place jobs inside the right locality domains. It regenerates the view as the cluster changes rather than requiring manual maintenance. Supported clouds are Google Cloud, Lambda, Nebius, Nscale, and Oracle Cloud Infrastructure, with on-premises coverage for InfiniBand, Spectrum-X, and Multi-Node NVLink domains, and KAI Scheduler integration for topology-aware gang scheduling. Source

5. DLSS 5 Treats the Engine Frame as Ground Truth for Neural Rendering

NVIDIA. DLSS 5 introduces 3D-Guided Neural Rendering, which uses the engine’s rendered frame as a fixed foundation and adds lighting and material detail on top, running one frame in and one frame out with engine motion vectors for deterministic, temporally stable output on GeForce RTX 50 Series at up to 4K. Developers get model selection, Structure Intensity and Tone Intensity controls, and semantic plus engine-level masking. NVIDIA ACE adds Nemotron Speech 3.5 Streaming and Qwen3 TTS, and the In-Game Inferencing SDK picks up RTX Spark support, Gemma4 integration, a Stable Diffusion plugin, and llama.cpp performance work. Source

6. Southeast Asian Public Sector Deployments Are Running on Nemotron 3

NVIDIA. At AI Day Singapore, NVIDIA detailed public sector work moving from pilot to production across the region. Singapore’s HTX is researching with Nemotron 3 Super for complex reasoning and agentic workflows and Nemotron 3 Nano Omni for multimodal public safety applications grounded in operational data. NCS is using Nemotron models with the NVIDIA Blueprint for video search and summarization alongside humanoid robotics work, citing security, responsiveness, and data governance requirements. Source