AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
NVIDIA Developer Blog
First-party releases and research from NVIDIA Developer Blog. Headlines and excerpts link to the original articles.
Latest stories
Newest first
NVIDIA released cuPotential, a C-API library of modular primitives for accelerating long-range electrostatics and dispersion in molecular dynamics, plus an OpenMM plugin that swaps in drop-in replacements for the nonbonded force, Langevin integrator and Monte Carlo barostat. Benchmarks show 2x to over 3x end-to-end speedups on 24,000- to 400,000-atom systems under NPT conditions.

NVIDIA's Vera Rubin NVL72 made its MLPerf Training v6.1 debut with the fastest time to train on six of seven benchmarks and up to 2x more performance per GPU than an optimized GB300 NVL72 baseline.

NVIDIA published provisional MLPerf Endpoints v1.0 results claiming Vera Rubin NVL72 delivers more than 10x the throughput per kilowatt of GB300 NVL72 at a shared minimum of 50 tokens per second per user. The results are pending MLCommons verification.

NVIDIA's cuPhoton.xBeam, a GPU-based FX correlator for radio interferometry, processed a 10 TB Event Horizon Telescope-style recording in 70 seconds on a 64-GPU Vera Rubin NVL72 system, about 460x faster than an x86 CPU baseline of 31,992 seconds. The tool ships in cuPhoton v0.1.3 with a command-line interface and Python API.

NVIDIA's cuda-bindings 13.3.0 adds a cuda.bindings.cudla module that exposes the cuDLA runtime API to Python, letting developers run DLA inference on Jetson AGX Orin in hybrid mode alongside GPU work on one CUDA stream. The same API also applies to RTX Spark.

NVIDIA released SOMA Retargeter v0.2, an open-source tool that adapts human motion captured on the SOMA skeleton to specific humanoid robots via scaling, Newton-based inverse kinematics, and foot-plant correction, with reference configurations for five robots.

NVIDIA published a tutorial for serving GLM-5.3 NVFP4, a 753-billion-parameter mixture-of-experts model, across two GB300 DGX Stations linked by ConnectX-8 RDMA using vLLM tensor parallelism, since its roughly 464 GB checkpoint exceeds a single GPU's 252 GB of HBM3e.

NVIDIA's open source AI Cluster Runtime (AICR) now ships a qualified AI bill of materials component built on Google's k8s-aibom, which inventories AI workloads running on Kubernetes by emitting a CycloneDX 1.6 document per recognized workload.

NVIDIA's Holoscan 5 tutorial builds a gravity-compensation loop for a Seeed Studio reBot Arm B601-DM using three custom operators: joint state read at 400 Hz, inverse dynamics via Pinocchio, and torque commands sent through motorbridge. Pinning the control partition to a CPU core cut p99.99 latency about 5x in C++ and 9.6x in Python.

NVIDIA's TensorRT LLM VisualGen runtime now applies NVFP4 quantization, sparse and quantized attention, step caching and multi-GPU parallelism to Black Forest Labs' FLUX 3 Image Klein, reaching up to 3.46x faster text-to-image and 4.45x faster image-to-image at 2048x2048 on a single B200.

NVIDIA released NCCL M2N Reshard, an open-source collective that transfers updated policy weights from training GPUs to inference GPUs during reinforcement learning post-training, cutting refit time by 2.7x to 5.7x in NeMo RL benchmarks on H100 and GB200 systems.

NVIDIA published a walkthrough of three open-source tools, licensed under Apache 2.0, for managing GPU clusters on Kubernetes: AICR for version-locked configuration recipes, NVCRE for workload-based certification, and NVSentinel for continuous health monitoring and automated remediation.

NVIDIA detailed five co-designed memory levers, from NVFP4 weight quantization to pod-level KV-cache storage, that it says let Vera Rubin NVL72 reach up to 30x the throughput per megawatt of GB300 NVL72 at the same 288 GB of HBM per GPU.

NVIDIA released an open-source, Apache 2.0 workload-level recovery service that watches training jobs across a cluster via logs, GPU metrics and scheduler signals, then isolates faulty nodes and restarts from checkpoints. It ships with Slurm integration, starts in a non-acting monitor mode, and logged roughly 8,550 anomaly detections over three months on one cluster.

NVIDIA's AI Cluster Runtime now ships a Preview Vera Rubin NVL72 recipe family that pins validated driver, OS, DRA and operator versions into a reproducible bundle. End-to-end validation on reference clusters passed deployment, conformance and a cross-node NCCL all-reduce gate at 744.71 GB/s bus bandwidth.

NVIDIA released Workload Map, an Apache 2.0 open source Kubernetes CRD and Go library that describes distributed workload structure once as YAML so schedulers and platforms can interpret JobSet, PyTorchJob, Ray, and other frameworks without per-framework adapters.

NVIDIA has introduced AI Cluster Runtime (AICR) recipes that pin and validate the full Kubernetes runtime stack behind NVIDIA Dynamo, including the Dynamo 1.4.2 platform chart and its DynamoGraphDeployment CRD at v1beta1. AICR also documents two cache-aware routing paths and a five-step deploy-and-validate flow.

NVIDIA introduced NemoClaw, open blueprints that assemble a production agent stack from a tuned harness, Nemotron models, NeMo optimization tools, and the OpenShell secure runtime, supporting OpenClaw, Hermes Agent, and LangChain Deep Agents. A LangChain-tuned Nemotron 3 Ultra profile reaches near-frontier performance at roughly 10x lower cost per run than leading closed models, NVIDIA says.

NVIDIA has released an open-source GPU Fractioning Operator, hosted under the KAI Scheduler GitHub organization, that enforces per-workload GPU memory limits on Kubernetes nodes using a new CUDA 13.4 capability; KAI Scheduler v0.18.0 is the first scheduler integrated with it.

NVIDIA announced RTX Spark, a Windows PC platform pairing a Blackwell RTX GPU with a Grace CPU and up to 128 GB of unified memory, alongside game-developer updates including DLSS 5, the Nemotron 3.5 Bolt E4B 4GB model and Nsight tool support. An October DLSS 5 model update runs 2x faster than September's and cuts 4K VRAM use by 20%.