Knowledge catalogue
ReplaySSM stores an SSM state as a high‑precision checkpoint plus a buffer of recent inputs, cutting memory traffic by ~50 % and accelerating SSM decoding kernels on NVIDIA Blackwell Ultra B300 GPUs b
NVIDIA’s SimReady Foundation provides a five‑step workflow to convert CAD models into simulation‑ready OpenUSD assets for robotics. The process imports an ABB YuMi robot from STEP files, applies visua
The NVIDIA KGMON team secured second place in the KDD Cup 2026 Data Agents competition by building a constrained harness around a small fixed LLM, unifying heterogeneous data into an SQLite database w
NVIDIA cuPhoton is an open‑source CUDA‑X toolkit that keeps scientific image data on the GPU from ingestion through classification, providing modules such as xDataReader (direct FITS loading), xRep (a
NVIDIA cuOpt’s Multi‑GPU Primal‑Dual Hybrid Gradient (mPDLP) solver distributes linear programming problems across NVLink‑connected GPUs, cutting per‑GPU peak memory by up to six times and handling up
NVIDIA’s Seattle Robotics Lab and Isaac engineering team created reusable robotic systems to automate GB300 tester‑tray assembly, achieving over 95 % success on busbar tasks with a 160‑second cycle ti
NVIDIA DSX Air creates a node‑based digital twin that models an AI factory’s infrastructure and software interfaces, allowing agentic workflows to interrogate the virtual environment, run configuratio
In CUDA 13.1 “green contexts” let applications explicitly partition a GPU’s execution resources—such as Streaming Multiprocessors (SMs) and work‑queue slots—within one process. By assigning dedicated
DOCA GPUNetIO delivers a unified GPU‑directed networking (GDA‑KI) foundation that lets CUDA kernels drive Ethernet, RDMA, Verbs and DMA operations without the CPU entering the application path. It shi
Megatron‑LM PR #7262 introduces a workflow that records ordered per‑rank tensor fingerprints to pinpoint and correct nondeterminism in large‑scale pretraining. By implementing low‑overhead kernel twea
NVIDIA DOCA AI agent skills supply verified API signatures, hardware capability checks, and build constraints that let agents fully understand the DOCA library (Flow, GPUNetIO, PCC, RDMA) for BlueFiel
The Nvidia Deploy repository provides open‑source C++ samples that pair ONNX Runtime with the TensorRT RTX execution provider to accelerate local inference on Windows and Linux. It separates model exp
NVIDIA Nemotron 3.5 ASR was fine‑tuned on 133.7 h of Saudi Najdi and Hijazi speech using weighted replay with FLEURS data, duration bucketing, and selective encoder unfreezing; this reduced the target
Generative recommender systems model recommendation as high‑cardinality event streams, turning user actions, context, and candidate items into tokens for sequence modeling over user behavior. NVIDIA’s
NVIDIA has released its cuObject client and server libraries, providing RDMA‑based APIs that allow GPUs to access object storage directly without routing data through the host CPU. The xio‑sig consort
NVIDIA NeMo Relay supplies a common observability layer that captures ordered lifecycle events and structured trajectories—ATOF event streams, ATIF step‑by‑step footprints, and OpenTelemetry spans wit
NVIDIA TensorRT Model Connect is an open‑source library that delivers AI model reference implementations in C++ built on top of NVIDIA TensorRT, designed to make the inference stack accessible to deve
NVIDIA released VSS Blueprint 3.3, a framework that stitches together vision‑language models (Cosmos), LLMs (Nemotron) and retrieval‑augmented generation to convert live or recorded video into natural
NVIDIA OpenShell 0.1.0 is an open‑source runtime that enforces which systems and data an AI agent may access without requiring changes to the agent itself. Its components—Gateway, Supervisor, Sandbox,
NVIDIA’s DSX MaxLPS employs policy‑governed power sharing to dynamically reallocate power across GPU resources, enabling up to 40 % more GPUs within the same approved power budget. In a joint test at
The NVIDIA Open Agent Safety Platform blends the open‑source OpenShell runtime—executing autonomous AI agents in kernel‑isolated sandboxes—with hardware‑based enforcement via NVIDIA Sentry on BlueFiel
Mixture‑of‑Experts (MoE) models scale capacity more efficiently than dense transformers by activating only a subset of expert subnetworks per token. NVIDIA’s Transformer Engine optimizes MoE training
SWE‑Serve evaluates AI coding agents on 53 inference‑engineering tasks derived from SGLang pull requests across six functional domains. Among 19 tasks that launch a live server, patches that pass loca
NV‑Reason‑CT is a purpose‑built vision‑language model for volumetric CT that uses a full 3‑D Vision Transformer encoder coupled with a Qwen 3.5‑4B language model to generate radiologist‑style chain‑of
NodeWright is an open‑source, Kubernetes‑native package manager that declaratively configures host operating systems on GPU clusters while preserving workloads; its operator cordons nodes, drains pend
NVIDIA Cluster Readiness Engine (NVCRE) is an open‑source Kubernetes controller that runs real distributed workloads across topology‑aware node groups to certify GPU cluster readiness before productio
ROS 2 nodes in NVIDIA Isaac ROS 5.0 now employ a CUDA buffer backend (via rosidl::Buffer) so that payloads remain GPU‑resident and can be transferred between nodes with zero‑copy when both endpoints s
NVIDIA Confidential Computing (CC) secures large‑language‑model inference by providing encrypted virtual machines, GPUs, and NVLink, enabling processing of sensitive data in trusted hardware. Running
**Summary:** NVIDIA Topograph discovers and normalizes cluster topology from cloud APIs or on‑prem fabric systems into a common model that is published as Kubernetes node labels, Slurm configuration,
NVIDIA’s DLSS 5 introduces a 3D‑Guided Neural Rendering system that adds realistic lighting and material detail to the renderer’s frame while preserving artistic intent; it operates one‑frame‑in/one‑f
Agent evaluation is moving from scoring individual function calls to judging full task completion within executable environments that track state over multi‑step tool use. Scoring occurs at two levels
NVIDIA TensorRT’s multi‑device inference lets a single TensorRT network span multiple GPUs using NCCL‑backed collectives, preserving all standard optimizations and supported from TensorRT 11.0 onward.
NVIDIA Earth‑2 is an AI‑driven data‑assimilation platform that ingests observations from proprietary or third‑party sources—such as satellites, radar, and in‑situ sensors—to generate more frequent wea
NVIDIA AIPerf is a benchmarking framework that supersedes GenAI‑Perf with a multiprocess architecture, eliminating client‑side bottlenecks during high‑concurrency LLM testing. It supports more than 15
Agentic AI workflows using NVIDIA Omniverse libraries automate preparation of Blender scenes for robotics simulation by adding semantic labels, sensor definitions, physics properties and validating th
TensorRT Edge‑LLM ran Qwen 3.6‑27B on a single NVIDIA Jetson AGX Thor and completed the MLPerf Edge Agentic benchmark in 24 min 36 s—52.33 tokens/s—which is 6.4× faster than the llama.cpp reference (2
cuTile Rust extends Rust’s ownership model to tile‑based GPU kernels by splitting mutable outputs into disjoint parts while preserving host‑side ownership across launches. An agent skill translates cu
Dense models activate all of their parameters for each token, whereas a Mixture‑of‑Experts (MoE) model loads all experts into memory but activates only a subset—e.g., Nemotron 3.5 Lightning has 30 B t
The NVIDIA Vera Rubin platform attains superior performance‑per‑watt by combining factory‑level DSX MaxLPS power shifting, rack‑level capacitors with Intelligent Power Smoothing, and the Groq 3 LPX de
NVIDIA NVLink 6 offers a natively lossless fabric for large‑scale AI factories by integrating lightweight Forward Error Correction, physical‑layer retries, UPHY recovery and credit‑based flow control
NVIDIA FLARE separates persistent federation services from dynamically launched job workers, allowing each participating site to deploy federated learning jobs on Docker, Kubernetes, or Slurm without
**AI Knowledge‑Base Summary** NVIDIA’s Transformer Engine augments JAX for dropless Mixture‑of‑Experts (MoE) training, delivering a 10.4× throughput boost on NVIDIA GB200 GPUs and raising DeepSeek‑V3
NVIDIA built a Digital Supply Chain Intelligence command center on Palantir Foundry that unifies material, capacity, and qualitative signals to drive critical allocation decisions. The weekly mixed‑in
NVIDIA’s BioNeMo Inference Runtime (BioIR) accelerates biomolecular structure‑prediction models on GPUs while preserving the familiar PyTorch workflow by providing an end‑to‑end pipeline that handles
NVIDIA’s Nemotron 3 Ultra NIM 2.0.12 stack boosts a 4×B200 system’s throughput by up to 2.5 ×, achieving roughly 1,997 tokens per second at a target of 50 transactions‑per‑user through autotuned kerne
CUDA Toolkit 13.4 introduces Windows‑on‑Arm support and preview functional compatibility for the NVIDIA Rubin architecture (compute capability 107), enabling cross‑platform CUDA development on ARM sys
Encode‑Prefill‑Decode (EPD) disaggregation is an inference optimization in NVIDIA Dynamo that separates vision encoding from LLM prefill–decode stages for multimodal requests. On image‑heavy prompts,
NVIDIA introduced two Rust projects for GPU development in September 2026: `cuda‑oxide`, a nightly‑only backend that compiles SIMT‑style kernels to PTX via LLVM and uses DisjointSlice and launch contr
NVIDIA NemoClaw delivers a memory‑driven Chief of Staff agent that constructs a structured self‑model (people, projects, priorities) stored in Markdown pages, with an SQLite ledger recording obligatio
NVIDIA Jetson can run compact open‑source reasoning models such as Nemotron 3.5 Lightning (a mixture‑of‑experts architecture with 30 B total parameters but only ~3 B active per token) and dense Qwen 3
A centralized identity gateway pattern decouples session ownership from request enforcement across federated Kubernetes clusters, letting regional gateways validate user identities via a shared Redis‑
NVIDIA’s PAIR Virtual Inference Router distributes AI inference requests across multiple local machines—supporting Ollama and LM Studio without any agent or harness changes—to mitigate multi‑agent bot
Speculative decoding speeds LLM inference by letting a lightweight draft model propose multiple tokens that the larger target model verifies in parallel, reducing decoding iterations while preserving
NVIDIA’s modern CUDA Toolkit (via CCCL launch API and indexing model) eliminates out‑of‑bounds shared‑memory errors and replaces hand‑written kernels with CUB block‑level primitives, reducing per‐imag
NVIDIA and CrowdStrike evaluated an “agentic” cyber‑defense platform that fuses NVIDIA’s Nemotron open‑model family with Falcon telemetry for a continuous offense‑defense closed loop running at machin
Inference workloads fall into four use‑case buckets (Chatbots/Copilots, Agents, Content Generation, Translation), each with distinct token patterns that dictate the required GPU memory and compute. Si
NVIDIA’s BioNeMo Agent Toolkit integrates over a decade of life‑science models and libraries into agent-callable skills, and when combined with Claude Science and NVIDIA NIM microservices it allows AI
NVIDIA Omniverse NuRec reconstructs real‑world driving scenes through 3D Gaussian splatting to produce the Physical AI NuRec Dataset—over 1,500 neural‑reconstructed videos (~20 s each) captured from s
NVIDIA TensorRT Model Connect delivers a set of reference implementations for deploying open‑source models from Hugging Face checkpoints to native C++ inference in two steps: a Python CLI builds a dep
The Qwen3.8‑Flash‑Next 176 B multimodal MoE model uses a hybrid Gated DeltaNet / Qwen Sparse Attention (QSA) architecture that compresses long‑context history and aggregates tokens into micro‑blocks f