2026-10-21 12:30
Unlocking the potential of protein design! 🧬 The article explores RFD4-Proteina, an innovative model co-developed by NVIDIA and the University of Washington. It utilizes reward-guided inference-time search to navigate the vast possibilities of protein sequences. This tool aids researchers in designing effective protein binder candidates for oncology, specifically targeting PD-L1. RFD4-Proteina is accessible as an open model or via the NIM microservice for quick deployment. #ProteinDesign...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-21 09:52
🚀 NVIDIA Sentry is a new security framework designed to monitor AI agents and enforce access policies. It ensures that agents operate within their defined boundaries, even if they can bypass security controls. Built on NVIDIA DOCA and running on BlueField-4 DPUs, Sentry provides host-isolated runtime protection for AI systems. This framework establishes a separate hardware trust domain, continuously evaluating agent behavior for security and integrity. #NVIDIA #AI #CyberSecurity...
Source: Nvidia Developer Blog
Ofir Arkin
2026-10-21 09:50
NVIDIA's OpenShell Reference Architecture enhances security for autonomous AI agents operating at scale. These agents can perform complex tasks but may pose risks by accessing sensitive systems. OpenShell ensures safety through governance that manages agent activities. Key components include a control plane, a supervisor for policy checks, a sandbox for isolation, and a policy prover for verification. This structure helps maintain secure operations while allowing agents to adapt and perform...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-21 09:38
🚀 Building an AI factory digital twin involves creating a cohesive model from site plans, equipment, and engineering needs. The NVIDIA Omniverse DSX Blueprint offers essential assets and workflows to aid in this process. With the help of Frontier AI models, developers can effectively assemble and revise designs. This agent-assisted workflow enables teams to explore layouts and prepare for engineering reviews. #AIFactory #DigitalTwin #NVIDIA #Engineering #Simulation
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-21 09:35
NVIDIA's DSX POD streamlines AI factory integration by shifting much of the setup to a controlled factory environment. 🏭 This approach allows operators to receive pre-integrated AI infrastructure, reducing onsite complexity and accelerating productivity. The DSX POD aligns with NVIDIA's Vera Rubin platform, enhancing deployment efficiency through standardized processes. Learn how this innovation supports a faster path to productive AI capacity. ⚡️🔗 #NVIDIA #AIFactory #TechInnovation...
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-19 17:00
🚀 NVIDIA CUDA Tile (cuTile) is a Python-based DSL designed for optimizing GPU performance. It simplifies tile-level GPU programming by using high-level primitives. Jane Street, an early adopter, implemented Mamba-2 using cuTile, improving their sequence modeling. The article discusses the three custom kernels that enhance the Mamba-2 forward pass. For details on setup and architecture trade-offs, check out the full article! #NVIDIA #MachineLearning #CUDA #DeepLearning #AI
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-16 17:00
🚀 AI agents are transforming software development! Developers now guide AI to write code instead of doing it all manually. NVIDIA RTX Spark allows coding, testing, and deployment on a single machine, enhancing efficiency. With options for Windows and Linux, developers can work seamlessly without code migrations. Equipped with powerful NVIDIA hardware, RTX Spark delivers exceptional performance for AI tasks. #AIDevelopment #NVIDIA #RTXSpark #SoftwareEngineering #Innovation
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-15 18:00
🚀 Meta has launched Muse Spark Open, a new coding-optimized 816B mixture-of-experts (MoE) model. With 42B active parameters and a 1M token context window, it enhances agentic coding workflows by improving accuracy and understanding across complex codebases. This model supports multimodal tasks, processing text, images, and audio seamlessly, making it ideal for extensive coding projects. #Meta #AI #Coding #MuseSpark #Innovation
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-15 15:00
NVIDIA has introduced the Nemotron 3.5 Super VL, designed for complex enterprise workflows. This model employs a customizable multimodal architecture, featuring 120 billion parameters. It efficiently manages tasks by activating 12 billion parameters per token. On a single GPU, it achieves up to 5x higher throughput for agentic tasks compared to similar models. #NVIDIA #MachineLearning #AI #EnterpriseSolutions #Innovation 🚀💻
Source: Nvidia Developer Blog
Michelle Horton
2026-10-14 16:10
🚀 AI agents are transforming chip design workflows, but reliability remains a challenge. Errors can propagate through stages, impacting final outputs. 🔍 NVIDIA NeMo Gym is addressing these issues by evaluating agents against verified results. This helps teams trace errors and improve automation trustworthiness. 📈 A case study with Siemens Fuse EDA shows a rise in reliability scores from 55% to ~95% by refining agent skills. Learn how to enhance your chip design process with AI tools!...
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-13 21:33
🚀 Enterprise applications face rising demands for data processing. Agentic AI increases the need for handling complex queries efficiently. NVIDIA's cuDF enhances Apache Spark by utilizing GPU acceleration, improving performance without escalating costs. This allows for better total cost of ownership (TCO) in major workloads. The article discusses how AI agents can quickly rewrite user-defined functions (UDFs) into GPU-optimized versions, achieving speedups of over 22x in Spark jobs. Discover...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-13 14:19
🚀 Exciting news from NVIDIA! Today, they announced that the full NVIDIA Run:ai platform is now open source. This move allows customers, developers, and partners to access advanced AI workload and GPU orchestration capabilities. With this release, NVIDIA DSX AI Factory expands its offerings, enabling faster market solutions for evolving customer needs. The platform supports running AI workloads on Kubernetes efficiently, managing resources throughout the entire lifecycle. The open-source...
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-12 16:00
🚀 ReplaySSM rethinks state storage for hybrid models, combining state-space model (SSM) layers with attention layers for improved decoding speed. Key advancements include a checkpoint representation of SSM state, which reduces latency and addresses issues like state traffic and error accumulation. This collaboration between NVIDIA and Princeton's Dao AI Lab highlights significant benefits for models such as GLM-5.3-Flash and Kimi-K3. The full paper is available on arXiv, with code on GitHub....
Source: Nvidia Developer Blog
Michelle Horton
2026-10-08 20:57
🔧 Preparing CAD assets for robotics simulation involves more than just geometry conversion to OpenUSD. Developers need to configure materials, collision geometry, and physics properties to ensure accurate robot behavior. NVIDIA Omniverse libraries, aligned with SimReady Foundation specifications, streamline this process. Frontier AI models can assist by automating tasks and reducing manual preparation. This article outlines the five essential steps for preparing an ABB Robotics YuMi robot,...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-08 18:30
🚀 The NVIDIA KGMON team achieved second place in the KDD Cup 2026 Data Agents competition by streamlining their agent's harness for clearer, more reliable performance. The competition involved answering natural-language questions across diverse data sources, pushing the agents to inspect data and reason effectively. Key strategies included constraining the action space and ensuring every attempt was inspectable, which helped identify and refine failures. These principles can enhance...
Source: Nvidia Developer Blog
Jiwei Liu
2026-10-07 18:21
🔧 How can robots assemble complex systems like the GB300 tester trays? This challenge was taken on by NVIDIA's Seattle Robotics Lab, in collaboration with the Isaac engineering team. 🤖 The project focuses on two main tasks: busbar assembly and multi-connector insertion. Both require precision and skill, pushing the limits of robot learning and mechanical intelligence. 🔍 Overcoming these challenges has provided valuable insights into the effectiveness of end-to-end learning versus a modular...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-07 16:00
AI factories are highly complex, integrating various components like GPUs and CPUs with advanced orchestration and security. 🌐 Validating infrastructure and software before deployment is crucial. Digital twins offer a solution, providing a safe simulation environment for testing changes. This approach allows teams to model and validate configurations continuously, enhancing operational efficiency. ⚙️ #AIFactories #DigitalTwins #TechInnovation #AI #OperationalEfficiency
Source: Nvidia Developer Blog
Avi Alkobi
2026-10-07 15:45
NVIDIA's cuOpt is advancing decision optimization in supply chains and energy grids. 🌐 As challenges grow with more SKUs and constraints, cuOpt offers over 10x speed improvements on large-scale linear programming problems using GPU acceleration. Recent advancements with mPDLP allow for solving complex benchmarks, like the zib03 problem, significantly faster by leveraging multiple GPUs. ⚙️ #NVIDIA #cuOpt #SupplyChain #Optimization #MachineLearning
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-07 15:00
🚀 Observatories and high-throughput instruments now generate vast amounts of image data, but traditional CPU-based pipelines can't keep up. NVIDIA cuPhoton addresses this challenge by providing a GPU-native toolkit that accelerates the entire process—from data acquisition to decision-making—reducing analysis time from hours or months to mere seconds. This advancement allows researchers to interactively process data and extract insights more efficiently. #NVIDIA #ImageAnalysis #DataScience...
Source: Nvidia Developer Blog
Quynh L. Nguyen
2026-10-06 19:58
Discover how bitwise determinism enhances large-scale pretraining with NVIDIA Megatron Core. This approach simplifies debugging and validation, especially for models with trillions of parameters. It ensures reliable failure replay and efficient recovery from interruptions, crucial for maintaining productivity across thousands of GPUs. A case study highlights NVIDIA's efforts to minimize overhead while ensuring training efficiency. #NVIDIA #MegatronCore #AI #MachineLearning #BitwiseDeterminism...
Source: Nvidia Developer Blog
Ashwath Aithal
2026-10-06 19:07
NVIDIA's DOCA GPUNetIO is transforming GPU networking by enabling GPU applications to handle networking and data movement independently from the CPU. This approach reduces latency and enhances real-time responses in distributed applications. Key features include CPU control functions for GPU memory transport and CUDA functions for data manipulation. The framework has evolved to integrate with various NVIDIA technologies, streamlining operations across multiple platforms. 🔗💻⚡️ #NVIDIA...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-06 16:13
📢 Exciting news for GPU-accelerated Kubernetes clusters! The NVIDIA AI Cluster Runtime (AICR) v1.0 introduces version-locked, validated recipes for seamless GPU cluster configuration. These recipes ensure compatibility across various components, reducing deployment errors. With a stable compatibility contract, operators and integrators can confidently utilize AICR's public interfaces for efficient deployments. #NVIDIA #Kubernetes #AICR #GPU #TechUpdates
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-06 15:00
🚀 GPU applications are evolving with the introduction of green contexts, allowing better control over resource sharing. These contexts enable applications to select specific GPU execution resources, enhancing efficiency in concurrent workloads. Available since NVIDIA CUDA 12.4 and further accessible in CUDA 13.1, green contexts improve how workloads are distributed. This leads to reduced competition for compute units and minimizes serialization issues. #GPU #CUDA #GreenContexts...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-10-01 18:13
Unlock faster application development on NVIDIA BlueField with the new DOCA AI agent skills! 🚀 These skills provide AI agents with the domain-specific knowledge necessary for effective infrastructure development. By leveraging verified API signatures and build constraints, developers can enhance their workflows. Available on GitHub, these skills offer a structured approach for better performance across various DOCA scenarios. Discover how to build more stable code with DOCA agent skills...
Source: Nvidia Developer Blog
Tanya Lenz
2026-10-01 17:59
Enhance your local applications with AI using C++ and NVIDIA TensorRT! The article discusses the "Do Inference Now" (DIN) Deploy, an open-source collection of C++ samples. It utilizes ONNX Runtime and NVIDIA TensorRT for seamless model integration across Windows and Linux systems. Developers can export models from Hugging Face to ONNX format and implement them in native C++ applications without needing a model-specific runtime. The project supports various AI tasks, including automatic speech...
Source: Nvidia Developer Blog
Luca Spindler
2026-10-01 05:00
🚀 Fine-tuning NVIDIA Nemotron enhances automatic speech recognition (ASR) for Saudi Arabic dialects. Regional dialects often lack representation, impacting model performance. Fine-tuning helps improve recognition for specific dialects while considering local conditions. The NVIDIA NeMo framework offers a structured approach to adapt ASR models effectively. Techniques like minimal curation, replay mixing, and length bucketing are key to this process. #NVIDIA #SpeechRecognition #ArabicDialects...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-09-30 20:54
Unlock the potential of generative recommenders (GR) with NVIDIA Dynamo-Triton! GR systems enhance personalization by treating recommendations as sequence modeling over user behavior. This method captures user interactions and context effectively. Dynamo-Triton supports a streamlined HSTU GR inference workflow, achieving significant latency improvements. Benchmarks show up to 5.93x lower latency with GPU-backed KV caching. Learn how to transition your HSTU models from development to...
Source: Nvidia Developer Blog
Tanya Lenz
2026-09-30 19:13
🚀 NVIDIA is enhancing AI storage access with cuObject and the SCADA Server SDK. AI engineers and cloud providers need quick and secure access to large file and object storage. The demand for high-speed data transfer is growing, requiring RDMA-accelerated, zero-copy methods. NVIDIA's partnership with Google Cloud and Microsoft expands xio-sig, introducing cuObject for easier storage integration. The new SCADA SDK allows storage providers to create responsive servers for GPU requests, improving...
Source: Nvidia Developer Blog
Harish Arora
2026-09-30 16:00
🔍 Exploring agent behavior is crucial for optimizing performance. An agent might complete a task but still take an inefficient route, leading to increased latency and potential failures. 📊 The tutorial on NVIDIA NeMo Relay guides users through evaluating agent traces. It covers the setup of a Hermes Agent, running tasks, and inspecting event streams. 💻 You'll learn to combine task verification with trace evidence for better analysis. Check it out for insights on improving agent efficiency!...
Source: Nvidia Developer Blog
William Markito Oliveira
2026-09-29 19:10
🚀 Exciting insights from the article on NVIDIA TensorRT Model Connect! This open-source project focuses on making the NVIDIA inference stack accessible to developers. It emphasizes parallel work, model-family isolation, and GPU-backed validation. Key engineering choices include scalable tasks and modular AI outputs, ensuring reliability in software development. The approach allows for efficient exploration of candidate changes while maintaining quality control. Learn more about the lessons...
Source: Nvidia Developer Blog
Tanya Lenz
2026-09-29 18:35
Unlock the potential of visual AI agents with NVIDIA's VSS Blueprint 3.3! This new version simplifies the development and deployment of agents that process video data at scale. It integrates vision-language models and tools for efficient ingestion, event detection, and reporting. Cost reductions are notable, with faster application composition and improved runtime efficiency using Adaptive Efficient Video Sampling. Join us live on Oct. 1 at 9 a.m. PT to see a visual AI agent in action! 📅🎥...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-09-28 08:56
NVIDIA has introduced the Open Agent Safety Platform, addressing the need for safety in agentic AI. 🚀 The article draws parallels to the internet's evolution in the 90s, highlighting how security measures, like sandboxing, enabled innovation while protecting users. 🔒 Recent incidents show AI agents breaching safety measures, sparking a debate on the urgency of enhancing AI safety research. The call is clear: establish independent security controls to foster trust in AI agents. 🤖 #AISafety...
Source: Nvidia Developer Blog
Tanya Lenz
2026-09-28 08:55
NVIDIA has introduced OpenShell 0.1.0, an open-source runtime designed to manage AI agents effectively. This tool allows agents to write code, access resources, and adapt to new information while enforcing strict access controls. It helps prevent unauthorized actions, protecting sensitive data and operations. Organizations like Cadence, Slack, and Gecko Robotics are already integrating OpenShell into their processes for chip design, automation, and robotics. Learn more about how OpenShell...
Source: Nvidia Developer Blog
Alex Watson
2026-09-28 01:00
Unlocking AI factory potential! 💡 NVIDIA's DSX MaxLPS enhances efficiency by reallocating power among GPUs, allowing up to 40% more GPUs within the same power budget. This approach addresses the common issue of unused capacity in static setups. A joint evaluation with Nscale at their Iceland data center showcases the benefits of dynamic power sharing. Learn how this can transform AI workloads! 🌍⚡️ #NVIDIA #AI #DataCenter #Efficiency #TechInnovation
Source: Nvidia Developer Blog
Sarah McKenney
2026-09-24 15:00
Exploring the advancements in language models, the article discusses the challenges of scaling dense architectures. Mixture-of-experts (MoE) architectures offer a solution by activating only a small number of subnetworks for each token, enhancing efficiency in large language models. However, effective implementation is crucial as it can impact GPU utilization and introduce additional challenges. NVIDIA's Transformer Engine helps optimize these processes, aiming to improve efficiency in...
Source: Nvidia Developer Blog
Michelle Horton
2026-09-18 19:04
Deploying a model is just the start; measuring its speed accurately is key. Traditional methods often fall short due to single-process limitations and concurrency issues. Enter NVIDIA AIPerf, a new benchmarking tool designed for scalability. AIPerf's multiprocess architecture allows for accurate load generation without bottlenecks. With support for 15+ endpoint types and customizable load patterns, it meets diverse benchmarking needs. AIPerf simplifies setup, making your testing process...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-09-16 23:20
Unlock the potential of AI in robotics simulation! 🤖✨ Learn how agentic AI workflows can prepare and validate digital twins for physical AI systems. These agents inspect 3D scenes, author relevant data in OpenUSD, and ensure everything meets SimReady requirements. Discover the challenges in transforming Blender scenes into simulation-ready environments and how specialized agents streamline the process using NVIDIA tools. #AI #Robotics #Simulation #3DModeling #DigitalTwins
Source: Nvidia Developer Blog
Tanya Lenz
2026-09-16 20:37
NVIDIA's TensorRT Edge-LLM has achieved a significant milestone in AI performance. 🌟 In the MLPerf Inference v6.1 Edge Agentic benchmark, it completed tasks on the Jetson AGX Thor 6.4x faster than previous submissions. This performance allows AI agents to efficiently manage complex conversations over multiple steps. Key methods used include NVFP4 quantization and tree-based multi-token prediction. 📈 The benchmark evaluates both performance and accuracy, ensuring AI agents operate effectively...
Source: Nvidia Developer Blog
Elizabeth Goodman
2026-09-16 16:28
🚀 Exciting advancements in GPU kernel authoring! The cuTile Rust (cutile-rs) system enhances the Rust programming language by integrating tile-based GPU kernel authoring. It ensures safe ownership while allowing programmers to manage lower-level control when needed. An AI agent skill translates existing CUDA Tile Python and Triton-TileIR kernels into cuTile Rust, achieving 99.5% performance accuracy. This includes 24 public operators and around 40 GPU kernels for various tasks. For those...
Source: Nvidia Developer Blog
Tanya Lenz
2026-09-15 17:00
Exploring model architectures: Dense vs. Mixture-of-Experts (MoE) 🤖 Dense models activate all parameters per token, ensuring consistency. In contrast, MoE models, like Nemotron 3.5 Lightning, only activate a subset for each token, enhancing capacity without using all parameters. ⚙️ The choice between these models impacts throughput, memory cost, and complexity, making it essential to consider deployment needs. Learn more about their performance and application scenarios! 📊 #MachineLearning...
Source: Nvidia Developer Blog
Elizabeth Goodman