News Hub

NVIDIA presents mixture-of-experts training methods for scaling biological foundation models efficiently.

NVIDIA released NV-Reason-CT, a vision language model for 3D CT scan analysis with chain-of-thought reasoning for radiologists.

NodeWright manages Kubernetes node configuration including kernel settings, packages, storage, and GPU workload tuning.

AI coding agent patches can pass local tests but fail in live serving with real model loads.

NVIDIA Confidential Computing enables private LLM inference by processing sensitive data inside trusted environments.

NVIDIA Topograph optimizes GPU workload placement to reduce power consumption in AI factories.

NVIDIA DLSS 5 adds 3D-guided neural rendering for game developers to enhance lighting and materials.

GPU acceleration of ROS 2 nodes requires optimizing message passing to avoid CPU serialization bottlenecks.

NVIDIA TensorRT now supports multi-device inference, allowing a single network to execute across multiple GPUs using NCCL.

AI agent evaluation should measure task completion across sequential tool calls and recovery from failures, not just output quality.

NVIDIA Earth-2 enables weather-sensitive industries to make timely decisions using local observations and AI predictions.

NVIDIA released AIPerf, a benchmarking tool for measuring LLM inference performance at scale.

AI agents can inspect 3D scenes, author simulation data in OpenUSD, add physics properties, and validate digital twins for physical systems.

TensorRT Edge-LLM completed MLPerf Edge Agentic Benchmark 6.4x faster on Jetson AGX Thor than baseline.

cuTile Rust enables safe GPU kernel authoring in Rust using tile-based operations.

Mixture-of-Experts models activate only a subset of parameters per token while retaining full model capacity.

NVIDIA NVLink 6 provides multi-layer resiliency for large-scale AI training clusters.

NVIDIA Groq 3 LPX uses deterministic execution for power-efficient AI inference on Vera Rubin hardware.

NVIDIA Transformer Engine accelerates dropless mixture of experts model training in JAX.

Full-stack NIM optimizations enable 2.5x more concurrent users on Nemotron 3 Ultra with preserved interactivity.