NVIDIA Dev Blog

Every byte moved has a cost. As model checkpoints grow to hundreds of gigabytes or even a terabyte, that cost adds up quickly. To make things even worse, moving... Every byte moved has a cost. As model checkpoints grow to hundreds of gigabytes or even a terabyte, that cost adds up quickly. To make things even worse, moving these model weights around the cluster is extremely common. For instance, a cold start may pull weights from remote storage into GPU memory; autoscaling and rolling updates must populate each new replica; and RL post-training continuously… Source

NVIDIA OptiX ray tracing engine is an application framework for achieving optimal ray tracing performance on the GPU. Applications using OptiX can fail in ways... NVIDIA OptiX ray tracing engine is an application framework for achieving optimal ray tracing performance on the GPU. Applications using OptiX can fail in ways that are difficult to diagnose: an invalid API argument, a black frame, or a GPU-side bug buried under thousands of concurrent threads. Debugging facilities in the NVIDIA OptiX Toolkit (OTK) can help. OTK is a GitHub repository… Source

Customization is what enables developers to take a general model and tailor it to use cases, domains, languages, and more. However, customization comes with a... Customization is what enables developers to take a general model and tailor it to use cases, domains, languages, and more. However, customization comes with a few challenges. It requires infrastructure, technical expertise, and software specific to the workflow, as well as resources such as GPUs and the ability to use them effectively. It also depends on specialized domain knowledge: What… Source

A TensorRT engine build can take seconds to many minutes. Large strongly typed models, deep tactic search, and a cold timing cache on a brand-new GPU SKU can... A TensorRT engine build can take seconds to many minutes. Large strongly typed models, deep tactic search, and a cold timing cache on a brand-new GPU SKU can leave developers, end users, or AI agents staring at a frozen terminal with no idea whether to wait, retry, or kill the process. Most NVIDIA TensorRT integrations report nothing during a build or provide no way to abort early. Source

Frontier model pre-training has converged on mixture of experts (MoE), which is fundamentally changing what limits large-scale AI training. As compute per token... Frontier model pre-training has converged on mixture of experts (MoE), which is fundamentally changing what limits large-scale AI training. As compute per token falls, communication increasingly determines how efficiently models scale across thousands of GPUs. NVIDIA GB300 NVL72 set a world record for pre-training DeepSeek-V3 671B at 1,648 TFLOPs per GPU, showing how advances across the entire AI… Source

Agentic AI shifts more of the critical execution path onto the CPU. Agents operate in sandboxes to execute code, invoke tools, retrieve context, interact with... Agentic AI shifts more of the critical execution path onto the CPU. Agents operate in sandboxes to execute code, invoke tools, retrieve context, interact with databases, and analyze results before returning information to the model. As these loops run concurrently across an AI factory, CPU performance increasingly shapes both per-agent responsiveness and overall factory throughput. Source

What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale.... What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale. These factories are now tasked with powering agentic workflows that reason, plan, use tools, verify intermediate results, and execute complex multistep tasks across vast contexts. Agentic workloads are not defined by a single prompt… Source

The demand for AI continues to accelerate. Workloads are getting larger, models are becoming more complex, and there is mounting pressure to deploy AI compute... The demand for AI continues to accelerate. Workloads are getting larger, models are becoming more complex, and there is mounting pressure to deploy AI compute infrastructure faster than ever. AI factories—data center-scale systems that continuously convert data and energy into intelligence—are being deployed to meet this insatiable demand. This AI factory approach to the data center has… Source

Developers building 3D, design, simulation, robotics, and industrial digital twin applications need ways to bring physical AI capabilities into the tools and... Developers building 3D, design, simulation, robotics, and industrial digital twin applications need ways to bring physical AI capabilities into the tools and services they already use. Many of these workflows already depend on OpenUSD scenes, simulation-ready (SimReady) assets, Blender-based workflows, CAD pipelines, or domain-specific app stacks. The challenge is how to provide applications and… Source

Capcom's RE ENGINE team set out to bring path tracing into two shipping titles at once, Resident Evil Requiem and PRAGMATA, each with a different visual... Source

A video analytics AI agent that can perceive, reason, and act based on massive amounts of video footage must be integrated with existing workflows and... A video analytics AI agent that can perceive, reason, and act based on massive amounts of video footage must be integrated with existing workflows and applications to be useful. These include content management systems, messaging platforms, databases, ticket queue, and escalation paths. This integration is challenging because video systems, enterprise knowledge bases… Source

Agentic AI changes the infrastructure pattern for AI factories. One request can trigger many model calls, tool calls, memory lookups, policy checks, storage... Agentic AI changes the infrastructure pattern for AI factories. One request can trigger many model calls, tool calls, memory lookups, policy checks, storage accesses, and network transfers before a final answer is produced. As more agents run at once and carry context across steps, users, tools, services, and sessions, infrastructure must move, protect, retrieve, and reuse data fast enough to keep… Source

Developers building video analytics applications across large spaces must track the same object as it moves between camera views. Single-camera 2D tracking... Developers building video analytics applications across large spaces must track the same object as it moves between camera views. Single-camera 2D tracking lacks reliable depth information and typically loses track of the object when it leaves the frame, limiting applications such as warehouse safety, retail analytics, and smart-building monitoring. Current 3D tracking methods require manual… Source

OpenUSD is an open, extensible framework that provides a common scene description language for physical AI. It enables teams to bring CAD data, simulation... OpenUSD is an open, extensible framework that provides a common scene description language for physical AI. It enables teams to bring CAD data, simulation assets, and real-world telemetry into a shared, physically accurate view of the world. Until now, building a USD implementation has typically required adapting a large existing codebase— even for teams that need a specific memory footprint… Source

For over fifteen years, x86 CPUs have shipped with a dedicated hardware instruction for carryless multiplication. It’s a small but stubborn primitive that... For over fifteen years, x86 CPUs have shipped with a dedicated hardware instruction for carryless multiplication. It’s a small but stubborn primitive that sits underneath authenticated encryption, error-correcting codes, and modern zero-knowledge proofs. Until now, NVIDIA GPUs lacked native support for this operation. NVIDIA CUDA 13.3 closes that gap with , a new PTX instruction available on… Source

The NVIDIA Nemotron Model Reasoning Challenge invited the Kaggle community to explore a focused question: What techniques can improve reasoning accuracy when... The NVIDIA Nemotron Model Reasoning Challenge invited the Kaggle community to explore a focused question: What techniques can improve reasoning accuracy when everyone starts from the same open model, benchmark, infrastructure and evaluation constraints? The response was massive. By the close of the competition, more than 5,000 active participants across 4,000 teams had generated thousands of… Source

What if autonomous coding AI agents could push your vision reasoning models above 90% accuracy with almost no manual effort? When adapting vision reasoning... What if autonomous coding AI agents could push your vision reasoning models above 90% accuracy with almost no manual effort? When adapting vision reasoning models to production video tasks, developers often lose days to data formatting, container setup, training scripts, baseline evaluation, and hyperparameter sweeps before they even know whether post-training improves accuracy. Source

Coding AI agents are becoming practical operators for long-running machine learning (ML) workflows. They can inspect repositories, set up runtimes, resolve... Coding AI agents are becoming practical operators for long-running machine learning (ML) workflows. They can inspect repositories, set up runtimes, resolve build issues, launch experiments, monitor execution, analyze metrics, and summarize results. For reinforcement learning (RL) research, this matters because meaningful metrics often appear only after the essential experiment infrastructure… Source

Useful quantum computers will require fault tolerant logical operations. Researchers are actively exploring many different quantum error correction (QEC) codes... Useful quantum computers will require fault tolerant logical operations. Researchers are actively exploring many different quantum error correction (QEC) codes to enable this, improving the Logical Error Rates (LER) of Quantum Processing Units (QPUs). While it is well understood how to run logical operations with surface codes (which belong to the topological code family) via lattice surgery… Source

Across science, engineering, and finance, many of the most important risks come from low-likelihood, high-impact events. Estimating the probability of these... Across science, engineering, and finance, many of the most important risks come from low-likelihood, high-impact events. Estimating the probability of these events with brute-force Monte Carlo sampling—running a model repeatedly with randomly drawn inputs to estimate the probability of rare outcomes—can require an excessive volume of model iterations, especially when each sample comes from an… Source

Robotics foundation models have made remarkable progress. Today's best systems can follow natural language instructions to pick, place, sort, and manipulate a... Robotics foundation models have made remarkable progress. Today’s best systems can follow natural language instructions to pick, place, sort, and manipulate a wide variety of objects. But as these models grow more capable, evaluating them rigorously has become one of the field’s hardest unsolved problems. In this blog post, we introduce the key problems and our method for addressing them. Source

Large language model (LLM) training workloads increasingly run into GPU memory limits before compute is fully used. Model weights, gradients, optimizer states,... Large language model (LLM) training workloads increasingly run into GPU memory limits before compute is fully used. Model weights, gradients, optimizer states, communication buffers, and intermediate activations all compete for GPU high-bandwidth memory (HBM). As model size, sequence length, and batch size grow, HBM capacity often becomes the primary scaling bottleneck. This post explains how… Source

There are many ways to optimize code for GPUs. In this post, you’ll learn how kernel fusion can improve memory bandwidth and reduce kernel launch overhead,... There are many ways to optimize code for GPUs. In this post, you’ll learn how kernel fusion can improve memory bandwidth and reduce kernel launch overhead, along with multiple ways to apply it in NVIDIA CUDA code. A common bottleneck when writing GPU code is that GPU compute is so fast that even high-bandwidth device memory doesn’t use the GPU kernel fully. Kernel fusion addresses this by… Source

AI performance comes down to three dimensions: Accuracy: How well the model reasons and produces outputs Throughput: How many tokens per second a... AI performance comes down to three dimensions: Deployments must balance all three: High accuracy is wasted if responses are slow, and raw throughput means little if each user’s experience is laggy. Practical systems therefore optimize accuracy, throughput, and interactivity together. This post focuses on throughput and interactivity, and how model-design choices shape both without… Source

Biomolecular structure prediction and co-folding with models like OpenFold3 are now mainstream, large-scale workloads powering drug discovery and protein... Biomolecular structure prediction and co-folding with models like OpenFold3 are now mainstream, large-scale workloads powering drug discovery and protein design. Increasingly, they’re driven end-to-end by AI agents. For an agent to run that pipeline well, every step needs to be fast and scalable: Multiple Sequence Alignment (MSA) generation, co-folding inference, serving, and multi-GPU scale-out. Source

Fine-tuning LLMs for financial natural language processing (NLP) is constrained by limited, imbalanced data. Real-world financial news overrepresents earnings... Fine-tuning LLMs for financial natural language processing (NLP) is constrained by limited, imbalanced data. Real-world financial news overrepresents earnings and stock movements, while rarer events such as credit-rating changes, product approvals, and labor issues are harder to capture at scale. Synthetic generation can help fill those gaps for trading research, risk modeling, and surveillance… Source

Molecular dynamics (MD) simulations are among the most demanding workloads in computational science. Using them, researchers can observe atomic behavior in... Source

Presto is an open source, distributed SQL engine for running fast, interactive queries on very large datasets. On NVIDIA GPUs, Presto delivers peak performance... Presto is an open source, distributed SQL engine for running fast, interactive queries on very large datasets. On NVIDIA GPUs, Presto delivers peak performance for analytical query workloads and provides low latency for users and agents. GPU-accelerated Presto brings low latency to your analytical workloads, keeping you and your agents unblocked and iterating as fast as possible. Source

Agentic systems often face a trade-off between accuracy and cost. The highest-performing proprietary frontier models and harnesses provide top accuracy but are... Agentic systems often face a trade-off between accuracy and cost. The highest-performing proprietary frontier models and harnesses provide top accuracy but are expensive. Fine-tuning offers one way to address this problem. Smaller or more efficient open models starting with lower accuracy are taught to perform better with specific agents. However, fine-tuning requires expertise and hardware for… Source

As more teams move from humanoid robot bring-up to task-specific skill development, the need for repeatable development workflows is growing. Building humanoids... As more teams move from humanoid robot bring-up to task-specific skill development, the need for repeatable development workflows is growing. Building humanoids remains complex, and today’s development pipelines are still highly fragmented. As a result, developers spend significant time configuring robotics infrastructure before they can focus on building robot capabilities. Source

Spectrum is one of the most valuable assets in wireless communications. Over the last 30 years, telecom operators in the US have spent more than $240B to... Spectrum is one of the most valuable assets in wireless communications. Over the last 30 years, telecom operators in the US have spent more than $240B to acquire wireless spectrum. A goal of a radio access network (RAN) system is to extract the maximum spectral efficiency (bits/second/Hertz) possible, which translates into more capacity, stronger network resilience with fewer dropped packets… Source

Industrial machinery generates more alarms than technicians can triage. For each important alarm requiring follow-up, the technician pulls historical context,... Industrial machinery generates more alarms than technicians can triage. For each important alarm requiring follow-up, the technician pulls historical context, determines the correct procedure, checks whether a specialist signal confirms the failure mode, and writes up a recommendation. This process remains consistent, and is well-suited for an AI agent. This post discusses a per-alarm… Source

Agentic systems turn model reasoning into action through multi-step workflows that combine inference, tool use, code execution, retrieval, orchestration, and... Agentic systems turn model reasoning into action through multi-step workflows that combine inference, tool use, code execution, retrieval, orchestration, and result handling. As these systems scale across the AI factory, performance depends not only on GPU acceleration, but also on the CPU work that happens between model steps. Across the creation and deployment of an agentic system… Source

Training LLMs at massive scale brings unique infrastructure challenges, especially as jobs span thousands of GPUs and run for extended periods. The longer these... Training LLMs at massive scale brings unique infrastructure challenges, especially as jobs span thousands of GPUs and run for extended periods. The longer these jobs run, the greater the likelihood of encountering unscheduled interruptions or resource fluctuations. Even infrequent device unavailability can have outsized effects on tightly interconnected clusters, resulting in slowdowns for a given… Source

AI has transformed how organizations operate, driving unprecedented levels of productivity and innovation. However, AI adoption can be impeded by concerns... AI has transformed how organizations operate, driving unprecedented levels of productivity and innovation. However, AI adoption can be impeded by concerns surrounding data privacy, sovereignty and how to secure data while it is in use, or during inference and engagement with AI models. NVIDIA Confidential Computing (CC) was engineered to be a secure and performant solution for the era of agentic… Source

Reinforcement learning (RL) is central to aligning language models, from reinforcement learning with human feedback (RLHF) within AI assistants to newer... Reinforcement learning (RL) is central to aligning language models, from reinforcement learning with human feedback (RLHF) within AI assistants to newer reinforcement learning with verifiable rewards (RLVR) workflows for reasoning and agent tasks. RL is now becoming a practical technique for specialized AI where enterprises need more accurate agents for domain-specific workflows. Source

GPU-accelerated query engines are often constrained by memory and I/O bandwidth. NVIDIA hardware advances—including high bandwidth memory (HBM), NVIDIA... GPU-accelerated query engines are often constrained by memory and I/O bandwidth. NVIDIA hardware advances—including high bandwidth memory (HBM), NVIDIA NVLink-C2C, and dedicated decompression engines featured in NVIDIA GB200 NVL4—help remove these bottlenecks by increasing effective storage capacity, accelerating data movement between CPUs and GPUs, and speeding data access without consuming… Source

NVIDIA Ominverse NuRec is a neural reconstruction pipeline for building high-fidelity 3D representations of real-world environments from multisensor data such... NVIDIA Ominverse NuRec is a neural reconstruction pipeline for building high-fidelity 3D representations of real-world environments from multisensor data such as cameras and lidar. It is used to reconstruct dynamic scenes captured by autonomous vehicle (AV) and robotics platforms into simulation-ready digital environments that can be rendered, replayed, and analyzed inside NVIDIA Omniverse and… Source

AI agents are quickly moving beyond chat. They inspect code, run tests, read documents, search knowledge bases, query internal systems, and operate for hours on... AI agents are quickly moving beyond chat. They inspect code, run tests, read documents, search knowledge bases, query internal systems, and operate for hours on behalf of a user. This unlocks productivity, but can also give agents access to sensitive enterprise data and the ability to complete tasks and take action across business systems, making a secure, governed environment essential. Source

AI agents have changed a lot in the last two years. The first could only answer one question at a time. Then came multi-turn chat, where the model could keep... AI agents have changed a lot in the last two years. The first could only answer one question at a time. Then came multi-turn chat, where the model could keep some context across a session. Today, we have long-horizon agents. Systems that plan many steps, split work between sub-agents, keep context across a long task, and run tools in a safe sandbox. The NVIDIA AI-Q Blueprint is an open source… Source

As context windows grow longer, moving large model weights efficiently becomes critical to performance. A common way to address this is quantization, an... As context windows grow longer, moving large model weights efficiently becomes critical to performance. A common way to address this is quantization, an optimization technique that compresses model weights into a smaller data format. One quantization format is NVFP4, an innovative 4-bit floating point introduced with NVIDIA Blackwell architecture. That’s the approach behind our new Nemotron 3… Source

Shaders are GPU programs that process visual data—such as rays, pixels, geometry, and textures—to produce specific rendering effects. Shaders find necessary... Shaders are GPU programs that process visual data—such as rays, pixels, geometry, and textures—to produce specific rendering effects. Shaders find necessary data through a process called resource binding. CPU code orchestrates the creation of GPU resources such as textures and memory buffers and then carefully arranges for shader code to access them through a binding protocol. Source

Generative AI workloads are rapidly outgrowing the memory and compute budget of single GPUs. For inference developers building media generation pipelines, the... Generative AI workloads are rapidly outgrowing the memory and compute budget of single GPUs. For inference developers building media generation pipelines, the challenge is scaling across multiple devices without sacrificing the critical optimizations—like kernel fusions, memory planning, and quantization—that NVIDIA TensorRT delivers for production deployments. Multi-device inference support… Source

AI companions in games have long been constrained by scripted behavior trees and fixed dialogue. PUBG Ally is a different kind of system. Built by KRAFTON for... AI companions in games have long been constrained by scripted behavior trees and fixed dialogue. PUBG Ally is a different kind of system. Built by KRAFTON for PUBG: BATTLEGROUNDS, this AI teammate is powered by NVIDIA ACE and its suite of efficient models and tooling. PUBG Ally uses automatic speech recognition, a 2B-parameter small language model, and text-to-speech to understand player… Source

An increasingly common design pattern for autonomous vehicles (AVs), robotics, and spatial AI systems is bird's-eye-view (BEV) perception. BEV models project... An increasingly common design pattern for autonomous vehicles (AVs), robotics, and spatial AI systems is bird’s-eye-view (BEV) perception. BEV models project multicamera image features into a shared top-down grid, providing downstream perception and planning modules with a common spatial layout for reasoning about lanes, vehicles, pedestrians, and free space. A key operation in this pipeline… Source

Power can account for 40% of the operating expenses (OpEx) to run an AI factory. Each watt can be spent on overhead, data ingestion, training, or generating... Power can account for 40% of the operating expenses (OpEx) to run an AI factory. Each watt can be spent on overhead, data ingestion, training, or generating tokens for customers. And most sites are capped at a fixed power level provided by a regional provider. Under these conditions, performance per watt becomes a key efficiency metric that directly translates to token costs. Source

As AI systems move from single-turn interactions to coordinated multiagent workflows, low-latency inference becomes increasingly important. Autoregressive LLMs... As AI systems move from single-turn interactions to coordinated multiagent workflows, low-latency inference becomes increasingly important. Autoregressive LLMs generate tokens sequentially, which can limit GPU utilization and constrain throughput in latency-sensitive serving scenarios. Speculative decoding helps mitigate this bottleneck by using a lightweight model to draft future tokens… Source

AI scientists are emerging as a new interface for scientific computing. These agents can read papers, write code, generate hypotheses, call APIs, inspect files,... AI scientists are emerging as a new interface for scientific computing. These agents can read papers, write code, generate hypotheses, call APIs, inspect files, and iterate on results. But science isn’t software engineering. There is no test suite that turns green when a hypothesis is correct; discovery is iterative, uncertain, and grounded in the physical world. You can’t take a general coding… Source

Telecom operators are adopting AI across network operations, customer care, and back-office workflows, but most are still early in the journey to autonomy. In... Telecom operators are adopting AI across network operations, customer care, and back-office workflows, but most are still early in the journey to autonomy. In network operations, for example, automation typically sits in the Level 2–3 band of TM Forum’s autonomous networks levels taxonomy, streamlining execution of predefined solutions in selective network domains. Reaching Level 4–5 autonomy… Source

The NVIDIA CUDA Core Compute Libraries (CCCL) provides delightful and efficient abstractions for CUDA developers in C++ and Python. It features: Parallel... The NVIDIA CUDA Core Compute Libraries (CCCL) provides delightful and efficient abstractions for CUDA developers in C++ and Python. It features: This post introduces a new group of functionality in CCCL that provides modernized C++ abstractions for fundamental CUDA programming model concepts that make CUDA C++ development safer and more convenient. NVIDIA CCCL runtime is a… Source