NVIDIA AI Updates: June 17, 2026

1. NVIDIA Blackwell Sweeps MLPerf Training 6.0

NVIDIA. Blackwell delivered the fastest time to train on every benchmark in MLPerf Training 6.0, and was the only platform to submit results across all seven benchmarks in the suite. The submissions scaled to 8,192 GPUs, the largest Blackwell-based entry to date, underscoring the platform’s performance, scale, and reliability for frontier model training. Source

2. NVIDIA XR AI Brings Agents to AR Glasses

NVIDIA XR AI is a new developer library for building multimodal AI agents that can perceive, reason, and act in the flow of work on AR glasses and XR devices. It combines video perception, enterprise data retrieval, reasoning models, and tool orchestration to deliver hands-free, contextual assistance in real time. Source

3. NVIDIA ACE Game Agent SDK Ships with Unreal Engine 5 Plugins

NVIDIA ACE Game Agent SDK is a lightweight C/C++ agentic framework for native in-game integration, exposing Agent, Chat, and RAG APIs to build responsive AI-powered NPCs. The accompanying Unreal Engine 5 plugins add local, on-device automatic speech recognition, small language models for dialogue, and text-to-speech, all optimized to run on RTX hardware without cloud dependencies. Source

4. NVIDIA Details Transaction Foundation Models for Financial Intelligence

NVIDIA. A new recipe applies transformer architectures to financial transaction sequences, learning behavioral patterns that improve fraud detection and credit scoring. The workflow uses GPU-accelerated data processing and custom tokenization via CUDA-X cuDF and cuML, pretrains a compact Llama-based decoder-only model with NeMo AutoModel, and extracts embeddings to augment downstream classifiers. Source

5. NVIDIA Publishes Guidance on Low-Precision Transformer Training

NVIDIA. A new technical post walks through how to optimize transformer-based models for low-precision training, covering techniques to preserve accuracy while reducing memory and compute costs. The guidance targets developers training large models more efficiently on NVIDIA GPUs. Source

6. NVIDIA Outlines World-Action Models for Generalist Robotics

NVIDIA. World-Action Models leverage pretrained video and world-model backbones to predict both future visual states and robot actions simultaneously, shifting language grounding to the video generation stage. The post references NVIDIA frameworks and models including DreamZero, Cosmos Policy, the Cosmos world foundation model, and GR00T N1 as examples of this emerging approach to generalist robot control. Source

7. NVIDIA Boosts MoE Training Throughput with Advanced Fusion Kernels

NVIDIA. New fused MLP kernels built with CuTe DSL combine GEMM, activation functions such as SwiGLU and GeGLU, and quantization into single kernels for mixture-of-experts training. The kernels deliver a 1.3x to 2x kernel-level speedup and up to 8 percent end-to-end gains on DeepSeek-V3 pre-training by eliminating memory overhead and CPU synchronization bottlenecks. Source