2026-11-13 22:13
🚀 New algorithms in NVIDIA's cuBLAS are set to enhance scientific simulations! The latest CUDA Toolkit 13.4.2 allows for high-precision dense matrix multiplications while utilizing faster integer Tensor Cores, ensuring speed does not compromise accuracy. Two new methods, Ozaki-II and 2M, significantly improve performance, achieving up to 6.6x speedup in real-world applications like quantum chemistry and materials science. Explore the advancements in double-precision performance! #NVIDIA...
Source: Nvidia Developer Blog
Cole Brower
2026-11-13 21:17
Imaging black holes presents significant challenges due to their distance and size. For instance, M87* has an observed ring measuring just 42 microarcseconds. 📷 The Event Horizon Telescope (EHT) uses 13 radio observatories worldwide, employing very-long-baseline interferometry (VLBI) to capture radio emissions from superheated gas. This method helps reconstruct detailed images of black holes. 🌌 The next-generation EHT (ngEHT) aims to enhance image quality and monitor black holes over time....
Source: Nvidia Developer Blog
Akhilesh Mishra
2026-11-13 18:57
Unlock the power of NVIDIA Deep Learning Accelerator (DLA) on Jetson AGX Orin! 🚀 This article outlines how DLA enhances AI applications at the edge by offloading inference from the GPU, improving power efficiency. The cuDLA API allows developers to manage DLA execution directly in Python, simplifying the integration process. 🐍 Learn about the inference lifecycle with a hands-on image classification example, covering device creation, model loading, and execution management. Explore how DLA and...
Source: Nvidia Developer Blog
Nikshay Shrivastava
2026-11-13 17:00
🚀 Exciting advancements in robotics! NVIDIA's SOMA Retargeter allows roboticists to adapt human motion for humanoid robots. It uses human demonstrations to teach robots natural movement while accommodating their unique structures. This open-source tool processes various human motion sources and provides consistent input for robot motion refinement. Learn more about the retargeting process and how it connects to policy learning! 🤖 #Robotics #NVIDIA #Innovation #MotionCapture #HumanoidRobots
Source: Nvidia Developer Blog
John Malaska
2026-11-13 17:00
🚀 Running large models like GLM-5.3 locally requires significant GPU memory. Two NVIDIA DGX Stations linked via ConnectX-8 RDMA can serve one model with tensor parallelism to split weights and computation. This setup allows for efficient processing of 753 billion parameters while maintaining all model weights accessible. Learn how to set up this system and measure performance under load! 📊 #NVIDIA #MachineLearning #AI #TensorParallelism #GLM5.3
Source: Nvidia Developer Blog
Fatima Fall
2026-11-11 17:00
🚀 Teams using NVIDIA Dynamo, Kubeflow, or Slurm need clear visibility into their AI workloads. Standard deployment bundles often don’t capture running artifacts accurately. 🔍 The NVIDIA AI Cluster Runtime (AICR) has introduced a qualified AI Bill of Materials (AIBOM) to enhance tracking on Kubernetes. This ensures reproducible deployment bundles with version-locked recipes. 🛠️ AICR offers a reliable configuration by binding upstream sources and verifying security settings, significantly...
Source: Nvidia Developer Blog
Mark Chmarny
2026-11-10 16:00
🚗 Autonomous vehicle (AV) developers are focusing on reasoning models that fit onboard compute and latency limits. NVIDIA's Alpamayo 2 Edge offers a vision-language-action model for real-time operations on DRIVE AGX Thor, enhancing driving performance through adaptive reasoning and reinforcement learning. Key tools like AlpaJudge help evaluate and improve model quality, ensuring effective learning and reasoning in AV systems. #AutonomousVehicles #NVIDIA #AI #MachineLearning #EdgeComputing
Source: Nvidia Developer Blog
Marco Pavone
2026-11-10 15:00
Introducing Agent Substrate, a new open-source solution aimed at optimizing Kubernetes for AI agents. 🤖✨ AI agents often remain idle, awaiting prompts or responses, leading to inefficient resource use. Traditional Kubernetes setups struggle with this, causing delays and memory issues. Agent Substrate, now part of the CNCF Sandbox, offers a zero-trust runtime designed to run millions of agents at higher efficiency. It utilizes a pool of worker Pods, allowing for quicker response times and...
Source: Nvidia Developer Blog
Ekin Karabulut
2026-11-10 15:00
🚀 NVIDIA has unveiled two open-source projects: NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), designed for managing Kubernetes on GPU infrastructure. NVKE offers a managed Kubernetes service with APIs for easy cluster management. NVKR focuses on creating operating system and container images, ensuring compatibility across software and GPU generations. These tools aim to streamline GPU capacity usage for AI Factory customers while maintaining control over their...
Source: Nvidia Developer Blog
Nathan Taber
2026-11-10 15:00
🚀 NVIDIA Snapshot is transforming GPU workload management! This new open-source tool, part of the NVIDIA DSX AI Factory Platform, allows for quick reuse of initialized GPU workloads across inference replicas. By capturing and restoring the full state of a workload, it reduces the cold start time from minutes to seconds. This improvement enhances scaling and recovery during high demand or failures. #NVIDIA #AI #GPU #Kubernetes #TechInnovation
Source: Nvidia Developer Blog
Dan Feigin
2026-11-09 19:09
Introducing NVIDIA Holoscan 5, a cutting-edge platform for robot arm control. It enables gravity compensation, allowing for intuitive guidance of robotic arms. This method relies on precise calculations of the robot's dynamics, essential for effective operation. Holoscan SDK simplifies integration, supporting low-latency control and high-throughput perception without rewriting existing code. Developers can easily add perception capabilities and inspect performance in real-time. Explore how...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-11-06 22:20
🚀 Developers are addressing inference latency in image generation with FLUX 3 Image Klein from Black Forest Labs. This model is optimized for NVIDIA Blackwell GPUs, enhancing efficiency in design tools and workflows. The tutorial explores how to configure FLUX 3 in NVIDIA TensorRT LLM VisualGen, utilizing optimizations to improve generation speeds significantly. Key results show up to 3.46x faster text-to-image and 4.45x faster image-to-image generation. Detailed deployment instructions and...
Source: Nvidia Developer Blog
Olivia Stoner
2026-11-06 18:03
🚀 In reinforcement learning (RL), the process of updating model weights between training and inference GPUs, known as refit, faces challenges due to differing computational requirements and large communication payloads. 🖥️ The Reshard collective offers a solution, enhancing efficiency in weight refit. An open-source implementation, NCCL M2N Reshard, is available on GitHub. 🔍 This advancement aims to improve communication in distributed systems as models continue to scale, addressing the...
Source: Nvidia Developer Blog
Tanya Lenz
2026-11-05 17:00
🔍 One underperforming GPU can impact entire distributed training jobs, often remaining undetected until workloads are running. 📊 It's crucial to maintain cluster readiness by validating throughout its operational lifecycle, especially after configuration changes and capacity growth. 💡 The NVIDIA DSX AI Factory Platform offers tools for configuring, certifying, and operating GPU-optimized Kubernetes infrastructure, ensuring efficiency and resiliency. For more details, check the full article!...
Source: Nvidia Developer Blog
Sowmyan Soman
2026-11-05 16:00
In the realm of AI, maximizing the utility of High Bandwidth Memory (HBM) is crucial. 🧠 NVIDIA highlights that every byte counts, as redundant memory usage can hinder performance. The focus is shifting from sheer memory capacity to the effective work each byte accomplishes. For instance, NVIDIA's latest technology shows significant improvements in throughput, achieving up to 30x the performance per megawatt compared to previous models. 📈 #NVIDIA #AI #HighBandwidthMemory #TechInnovation...
Source: Nvidia Developer Blog
Shara Tibken
2026-11-04 22:00
NVIDIA has introduced the AI Factory Operations Agent to enhance its DGX SuperPOD operations. This tool aims to reduce manual tasks for site reliability engineers (SREs) by consolidating performance monitoring into a single interface. Using specialized agents for different operational areas, it detects failures and offers actionable recommendations through a chat interface. SREs can interact naturally and track all actions for audit purposes. The framework is built on NVIDIA's blueprints,...
Source: Nvidia Developer Blog
Shara Tibken
2026-11-04 18:46
🚀 NVIDIA has released CUDA 13.5, enhancing performance and support for new platforms. Key updates include general availability for NVIDIA Vera Rubin and expanded Windows on Arm support for DGX Station. The new version introduces advanced features for GPU scheduling, including locality domain enhancements and improved NVIDIA Multi-Process Service (MPS) for better resource management. Explore the full capabilities of CUDA 13.5! #NVIDIA #CUDA #AI #AcceleratedComputing #TechUpdates
Source: Nvidia Developer Blog
Shara Tibken
2026-11-04 17:00
NVIDIA's Factory Operations Blueprint (FOX) aims to streamline factory management by integrating AI agents into existing systems. 🤖 A Siemens report highlights that unplanned downtime can cost large plants $253 million annually. FOX helps teams gather data efficiently, monitor real-time operations, and reduce delays. At the Siemens Electronics Factory, these agents enable quick responses to rush orders and optimize production planning, potentially saving costs by up to 20%. 💡 The integration...
Source: Nvidia Developer Blog
Samuel Ochoa
2026-11-04 16:00
🚀 Learn how to enhance your AI agents with NVIDIA NeMo! This article guides you through turning production evidence into prioritized issues and effective evaluations. It highlights the importance of optimizing prompts, tools, and workflows to improve agent performance. Follow a complete optimization cycle using NVIDIA NeMo Compass and Eval Author to identify and fix agent failures. Explore the full tutorial on GitHub for a hands-on experience! 🛠️💻 #NVIDIA #AIOptimization #MachineLearning...
Source: Nvidia Developer Blog
Patrick Moorhead
2026-11-02 23:39
🔍 In distributed AI workloads, a single error can lead to multiple symptoms, making diagnosis challenging. NVIDIA Attribution helps by analyzing workload evidence to identify failure patterns. 🛠️ This system correlates logs, metrics, and telemetry, providing a structured diagnosis that includes the root cause and affected nodes. 📊 With two components, the Attribution Service and Blueprint, it offers a clear process to scope workloads, collect observations, and extract significant events....
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-11-02 22:43
🚀 As distributed AI training expands, hardware and software failures can disrupt progress. Fast recovery is crucial to maintain efficiency. NVIDIA Recovery, part of the DSX OS, automates this process by detecting interruptions and managing restarts from checkpoints. This reduces manual effort and maintains productivity across training jobs. Learn how to implement this solution in your own cluster for improved performance. #AI #DistributedTraining #NVIDIA #Automation #TechInnovation
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-11-02 18:00
🚀 Configuring Kubernetes for NVIDIA Vera Rubin presents challenges due to the complexity of integrating drivers, kernels, and networking. The NVIDIA AI Cluster Runtime (AICR) streamlines this process by offering reusable configurations and validating clusters for seamless deployment. Additionally, the NVIDIA DSX AI Factory Platform supports efficient AI factory operations with modular software components. Key updates include new GPU drivers and NICs, which aid in planning for upgrades....
Source: Nvidia Developer Blog
Shara Tibken
2026-11-02 17:00
🌐 Developers in physical AI face challenges in adapting world models to their tasks and integrating them reliably. NVIDIA Cosmos offers a comprehensive solution with its platform for world model training and development. The Cosmos Framework provides essential tools for training, multimodal data processing, and inference. The article details how to utilize Cosmos for generating synchronized data for various applications and the deployment process to launch GPU-accelerated services. Key...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-11-02 16:00
🚀 NVIDIA's DSX AI Factory Platform streamlines the creation and operation of AI factories. It integrates modular software components and supports multi-tenant operations. Kubernetes users can now leverage the NVIDIA Workload Map, which standardizes workload descriptions, simplifying workload management across different frameworks. This innovation reduces the need for separate adapters, enhancing efficiency in scheduler and controller development. #NVIDIA #AI #Kubernetes #TechInnovation...
Source: Nvidia Developer Blog
Alon Lavian
2026-10-30 21:35
🚀 Teams creating AI applications need a reliable way to transition from model configuration to inference endpoints. NVIDIA Dynamo facilitates this with its distributed inference framework for GPU workers. To successfully deploy on Kubernetes, various components like GPU drivers, resource allocation, and monitoring must align. The NVIDIA DSX AI Factory Platform simplifies this with the AI Cluster Runtime (AICR), providing validated deployment recipes. The article explores the infrastructure...
Source: Nvidia Developer Blog
Stefan Schimanski
2026-10-30 20:33
🚀 AI is evolving beyond data centers! Developers are creating edge agents for tasks like monitoring driver attention and aiding robots in their environment. NVIDIA DGX Spark offers a streamlined platform for developers to prototype and test models with just a few commands. The TensorRT Edge-LLM enhances this process with a lightweight C++ runtime for embedded workloads. In a recent tutorial, users can run NVIDIA Nemotron-3.5 Lightning on Spark, enabling efficient application prototyping. #AI...
Source: Nvidia Developer Blog
Shara Tibken
2026-10-30 16:00
🚀 Building autonomous agents just got easier with NVIDIA NemoClaw! To create these agents, developers need a model, a harness, and domain-specific tools. NemoClaw provides open blueprints that let you customize and control your stack. NemoClaw supports various harnesses like OpenClaw, Hermes Agent, and LangChain Deep Agents, enhancing capabilities and ensuring safe operation. Discover how to integrate these components for reliable execution in production! #NVIDIA #NemoClaw #AI...
Source: Nvidia Developer Blog
Nicola Sessions
2026-10-30 15:01
🚀 NVIDIA GPUDirect Storage (GDS) enhances GPU storage access in both on-premises and cloud settings. It offers a fast, efficient data path that boosts bandwidth, reduces latency, and lowers CPU and memory usage. To configure GDS, it's essential to understand your specific environment and choose the right settings. Detailed documentation is available to guide you through the process. #NVIDIA #GPUDirectStorage #DataAccess #TechUpdate #CloudComputing
Source: Nvidia Developer Blog
Phil Carns
2026-10-29 15:00
Unlocking the potential of Tensor Core performance can enhance machine learning interatomic potentials (MLIPs) for atomistic simulations. 🌌 These simulations are vital for understanding molecular and material behavior, but balancing cost and accuracy is crucial. MLIPs utilize neural networks to predict energies and forces efficiently, allowing for larger systems to be studied. However, numerical precision is essential for stability in these simulations. By leveraging Tensor Core FP32 methods,...
Source: Nvidia Developer Blog
Anoushka Bhutani
2026-10-27 20:50
🚀 Exciting advancements in GPU sharing with NVIDIA CUDA 13.4! Many AI workloads utilize only a portion of a GPU, leading to idle capacity. CUDA 13.4 introduces memory enforcement in time-sharing mode, allowing better resource management. Key features include exact memory allocation, guaranteed ranges, and a choice of compute modes for workloads. To implement these features, a new open-source Kubernetes operator, gpu-fractioning, is available on GitHub, enhancing GPU efficiency across...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-27 17:00
🚀 NVIDIA has launched RTX Spark, a Windows PC platform designed for AI development and gaming in compact systems. It features the NVIDIA Blackwell RTX GPU and Grace CPU, supporting up to 128 GB of unified memory. Key technologies include the DLSS suite for enhanced FPS and image quality, and NVIDIA ACE for responsive AI characters. Game developers can utilize the RTX Kit SDKs for improved rendering and the latest updates for Unreal Engine integration. Explore the future of gaming with NVIDIA...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-27 16:00
🌍 As AI continues to grow, understanding local rules is essential for effective communication. NVIDIA's Nemotron models, developed with partners in France and Belgium, aim to enhance AI's adaptability to local languages and cultural contexts. Post-training with localized data improves accuracy and relevance in specific applications. This approach allows developers to tailor AI interactions based on regional needs, ensuring better service for users. Explore how fine-tuning can make AI more...
Source: Nvidia Developer Blog
Shara Tibken
2026-10-27 15:00
Boosting inference performance on NVIDIA DGX Spark can be achieved by addressing key limitations. Recent evaluations of the Ling-3.0-flash model identified memory-bound issues during decoding. Performance can be improved through three techniques: 1️⃣ Humming MXFP4 MoE kernels for optimized expert computation. 2️⃣ Online FP8 quantization to reduce weight-read volume. 3️⃣ Multi-token prediction for efficient token generation. The article provides insights into how these optimizations work,...
Source: Nvidia Developer Blog
Yang Liu
2026-10-26 16:00
NVIDIA Minerva addresses the challenges of managing complex AI factories with hundreds of thousands of GPUs and intricate networks. 🎛️ The platform collects telemetry data across the entire system, uniting insights from different teams to identify issues quickly. This helps prevent lost production time caused by subtle failures that can disrupt GPU performance. ⏳ Conventional monitoring struggles with "gray failures," where problems are not easily detectable. Minerva uses AI to correlate...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-26 15:00
NVIDIA has introduced Mission Control 2.5, enhancing autonomous recovery for Kubernetes workloads. This update aims to improve cluster resiliency and uptime, essential for training large-scale AI models on multiple GPUs. Traditional methods struggle with the demands of modern AI operations, making proactive management vital. The autonomous recovery engine now supports Kubernetes, enabling self-healing systems that detect failures and resume jobs with minimal disruption. #NVIDIA #Kubernetes...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-23 17:27
🚀 The Global Performance Tuner in NVIDIA TensorRT 11.1 automates performance tuning for deep learning models. It customizes optimizations for various models, enhancing accuracy and efficiency. This second part of the series delves into the algorithm and design behind the tuner, explaining how different scheduling strategies are represented as build routes. For best practices on implementation, refer to the full article. #NVIDIA #TensorRT #DeepLearning #PerformanceTuning #AI
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-23 17:27
Deploying neural networks involves complex decisions beyond simply converting operators to GPU kernels. NVIDIA TensorRT uses scheduling to optimize performance through choices about graph transformations, fusion, and memory placement. However, no single strategy fits all models, as changes can impact inference performance and accuracy differently. This article introduces the Global Performance Tuner in TensorRT 11.1, aimed at enhancing performance tuning. It uses an analogy of a chef...
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-23 14:58
🚀 Teaching robots like Microduck to walk involves more than a simple walking policy. They must perceive their surroundings and adapt to obstacles. 🛠️ The process includes creating a simulation environment and using NVIDIA tools like RTX Spark and Isaac Lab for training and evaluation. 📊 This article outlines how to set up a simulated obstacle course and refine navigation policies for real-world deployment. #Robotics #ArtificialIntelligence #Simulation #NVIDIA #TechInnovation
Source: Nvidia Developer Blog
Gabriele Leone
2026-10-22 16:39
Unlocking the power of GPU memory in Python is crucial for handling large datasets. 📊 NVIDIA's GPUDirect Storage (GDS) enables direct data transfer from storage to GPU memory, enhancing efficiency and reducing CPU load. This technology supports a variety of major storage vendors. The article highlights the use of NVIDIA KvikIO, a library that simplifies high-performance file I/O in Python and C++. 🐍 Learn how to configure storage I/O and leverage GDS for accelerated data movement. #NVIDIA...
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-22 16:00
🚀 AI development is evolving! Developers are now running multiple models in parallel and experimenting with advanced architectures using NVIDIA DGX Spark. NVIDIA recently announced support for eight-node DGX Spark clusters, which can scale unified memory to 1 TB. This upgrade enables the use of larger models, including the GLM 5.3 (753B), for local training and inference. 🔧 The new NVIDIA Sync simplifies clustering and device management. A benchmark showcases impressive performance, reaching...
Source: Nvidia Developer Blog
Shara Tibken