Items tagged with “Nvidia”.
NVIDIA HGX B200 lowers embodied carbon intensity by 24% vs. HGX H100, while delivering higher AI performance and energy efficiency. This article reviews the PCF-backed improvements, new hardware features, and implications for developers and enterprises.
NVIDIA and Berkeley Lab unveil Huge Ensembles (HENS), an open-source AI tool that forecasts low-likelihood, high-impact weather events using 27,000 years of data, with ready-to-run options.
NVIDIA Dynamo offloads KV Cache from GPU memory to cost-efficient storage, enabling longer context windows, higher concurrency, and lower inference costs for large-scale LLMs and generative AI workloads.
A detailed look at seven battle-tested techniques used by Kaggle Grandmasters to solve large tabular datasets fast with GPU acceleration, from diversified baselines to advanced ensembling and pseudo-labeling.
A detailed look at speculative decoding for AI inference, including draft-target and EAGLE-3 methods, how they reduce latency, and how to deploy on NVIDIA GPUs with TensorRT.
RAPIDS 25.08 introduces a function- and line-level profiler for cuml.accel, a default streaming executor for the Polars GPU engine, expanded datatype and string support, a new Spectral Embedding algorithm in cuML, and zero-code-change accelerations for several estimators.
Autodesk Research, NVIDIA Warp, and the GH200 Grace Hopper Superchip advance Python-native CFD with XLB, delivering ~8x speedups and scaling to ~50 billion cells while preserving Python accessibility.
A detailed look at how NVIDIA Run:ai Model Streamer lowers cold-start times for LLM inference by streaming weights into GPU memory, with benchmarks across GP3, IO2, and S3 storage.
PyNvVideoCodec 2.0 brings major enhancements to Python-based GPU-accelerated video encoding, decoding, and transcoding, built on NVIDIA's Video Codec SDK and available via PyPI and NVIDIA NGC. This update combines C++-level speed with Python flexibility.
A self-paced NVIDIA Dev Blog workshop demonstrates assembling a multi-layered AI agent for automated report generation using NVIDIA Nemotron on OpenRouter, featuring LangGraph, ReAct-based components, and practical prompts.
Open-source Qwen3-Next 80B-A3B-Thinking and 80B-A3B-Instruct preview a hybrid MoE architecture designed for long context and efficient inference, with NVIDIA-backed deployment options and open tooling.
CUDA-accelerated VC-6 enables selective, multi-resolution decoding tailored for GPUs, dramatically reducing I/O and accelerating data-to-tensor paths in vision AI pipelines.
Explores quantization aware training (QAT) and distillation (QAD) as methods to recover accuracy in low-precision models, leveraging NVIDIA's TensorRT Model Optimizer and FP8/NVFP4/MXFP4 formats.
NVIDIA’s AI Kill Chain framework maps how adversaries compromise AI-powered applications, detailing stages from recon to impact and offering defensive priorities to interrupt each step.
NVIDIA’s RTX PRO 6000 Blackwell Server Edition dramatically speeds protein-structure inference, enabling end-to-end GPU-resident workflows with OpenFold and TensorRT—achieving up to 138x faster folding than AlphaFold2.
CUDA is being redistributed by major third-party platforms to embed directly in package feeds, simplifying installation and dependency management for GPU-enabled applications and speeding deployment across ecosystems.
NVIDIA NIM Operator 3.0.0 expands scalable AI inference on Kubernetes, enabling multi-LLM and multi-node deployments, KServe integration, and DRA support in technology preview, with Red Hat collaboration and NeMo Guardrails.
Explores how NVIDIA Rivermax and NEIO FastSocket deliver ultra-low latency, high-throughput networking for financial services, leveraging kernel bypass, zero-copy, and GPUDirect to move data directly from NICs to GPUs.
NVIDIA Spectrum-XGS Ethernet enables scale-across networking to unify dispersed data centers into a single AI factory, delivering significant NCCL bandwidth gains and predictable performance over long distances.
A new GPU memory swap (model hot-swapping) approach lets multiple models share GPUs beyond capacity, reducing costs while preserving responsiveness for large-scale LLM inference.
A deep dive into nvMatmulHeuristics, a CUDA GPU kernel meta-parameter optimizer that accelerates GEMM auto-tuning with CUTLASS 4.2, delivering near-optimal kernels in far less time.
CUDA Toolkit 13.0 for Jetson Thor unifies ARM toolchains, enables Unified Virtual Memory, advanced GPU sharing, OpenRM/dmabuf interoperability, NUMA support, and enhanced developer tools for edge AI workloads.
Unified CUDA toolkit for Arm on Jetson Thor with full memory coherence, multi-process GPU sharing, OpenRM/dmabuf interoperability, NUMA support, and better tooling across embedded and server-class targets.
Leverage GPU memory swap (model hot-swapping) to share GPUs across multiple LLMs, reduce idle GPU costs, and improve autoscaling while meeting SLAs.
Introduces nvMatmulHeuristics to quickly select a small set of high-potential GEMM kernel configurations for CUTLASS 4.2, drastically reducing auto-tuning time while approaching exhaustive-search performance.
NVIDIA outlines a SFT + QAT workflow to recover FP4 accuracy for gpt-oss fine-tuning, comparing MXFP4 and NVFP4, and detailing deployment and performance gains with 98% pass rates on select tasks.
Explains why small language models (SLMs) enable scalable, cost-efficient agentic AI, the role of heterogenous model ecosystems, and practical paths to adoption with NVIDIA NeMo and Nemotron Nano 2.
Guide to fine-tuning gpt-oss with SFT + QAT to recover FP4 accuracy while preserving efficiency, including upcasting to BF16, MXFP4, NVFP4, and deployment with TensorRT-LLM.
Explores how small language models enable cost-effective, flexible agentic AI alongside LLMs, with NVIDIA NeMo and Nemotron Nano 2.
A practical overview of NVIDIA Isaac for Healthcare's telesurgery workflow, its modular architecture, and why it accelerates development from simulation to clinical deployment with low latency and unified tooling.
A production-ready, modular telesurgery workflow from NVIDIA Isaac for Healthcare unifies simulation and clinical deployment across a low-latency, three-computer architecture. It covers video/sensor streaming, robot control, haptics, and simulation to support training and remote procedures.
A detailed look at how NVIDIA scaled a LangGraph-based AI-Q research agent from a single user to hundreds, using NeMo Agent Toolkit, load testing, and phased rollout with observability.
NVIDIA’s CUDA 13.0 introduces a shared memory register spilling optimization that redirects register spills from local memory to on‑chip shared memory when space is available, reducing latency and improving performance in register‑heavy kernels.
CUDA 13.0 introduces shared memory register spilling to reduce local memory pressure by directing spilled registers to on‑chip shared memory when space permits. Opt-in via PTX inline assembly after the function declaration; gains typically 5–10% in register‑pressure workloads.
Guidance on deploying and scaling LangGraph-based agents in production using the NeMo Agent Toolkit, load testing, and phased rollout for hundreds to thousands of users.
NVIDIA advances co-packaged optics through industry collaboration, detailing the Micro Ring Modulator, COUPE engines, and modular external laser sources that enable dense, energy-efficient data-center interconnects.
Overview of NVIDIA's co-packaged optics (CPO) platform, highlighting photonics/electronics integration, Micro Ring Modulator, COUPE-based engines, and ecosystem collaboration with TSMC and partners for high-density, energy-efficient data-center interconnects.
Jetson Thor delivers edge-scale AI, enabling fast, multi-model generative reasoning with a Blackwell GPU, MIG, FP4/FP8, and 128 GB memory for next-gen robotic platforms.
NVIDIA extends NVFP4 from inference to pretraining, enabling 4-bit precision to accelerate large-scale transformer pretraining while preserving FP8/BF16-like accuracy. Early results on a 12B model show stable convergence and strong downstream performance, with up to 7x GEMM speedups on Blackwell Ult
Jetson Thor combines edge AI compute, MIG virtualization, and multimodal sensors for flexible, real-time robotics at the edge, with FP4/FP8 acceleration and support for Isaac GR00T and large language/vision models.
NVFP4 is a 4-bit data format delivering FP16-level accuracy with the throughput and memory efficiency of 4-bit precision, extended to pretraining for large language models. This profile covers 12B-scale experiments, stability, and industry collaborations.
A detailed guide to identifying and addressing five frequent pandas bottlenecks, with CPU-focused tips and a GPU-powered drop-in accelerator that can dramatically speed up workloads.
NVIDIA Blackwell Ultra accelerates AI training and inference with dual-die integration, NVFP4 4-bit precision, 288 GB HBM3e, and high-bandwidth interconnects, enabling scalable AI factories at scale.
NVIDIA outlines how Blackwell architecture, CUDA-X libraries, and a thriving open-source ecosystem—from RAPIDS to Cosmos and Nemotron—accelerate AI development from prototype to production.
A developer-focused resource outlining five common pandas bottlenecks, practical CPU and GPU fixes, and drop-in GPU acceleration with cudf.pandas for scalable data workflows.
An in‑depth profile of NVIDIA Blackwell Ultra, its dual‑die NV‑HBI design, NVFP4 precision, 288 GB HBM3e per GPU, and system‑level interconnects powering AI factories and large‑scale inference.
Overview of how NVIDIA's Blackwell hardware, open source models, and a rich software stack accelerate AI—from data prep to deployment—through broad model/tool availability and end-to-end infrastructure.
cuPQC v0.4 adds expanded hash function support and Merkle-tree calculations, accelerating data integrity checks and security for high‑speed cryptographic workloads on GPUs.
NVIDIA HPC SDK v25.7 advances unified memory programming for GPU-accelerated HPC, automating CPU-GPU data movement and boosting developer productivity on coherent architectures like Grace Hopper.
NVIDIA explores how NVLink Fusion extends scale-up NVLink technologies to hyperscalers, enabling custom CPUs and XPUs with a rack-scale fabric for AI inference at scale.
cuPQC v0.4 expands hash function support and adds Merkle-tree calculations to accelerate data integrity, verification, and privacy-preserving cryptography on GPUs.
Overview of NVLink Fusion, enabling scalable, high-bandwidth AI inference across custom CPU/XPU configurations and rack-scale NVLink fabrics.
NVIDIA HPC SDK v25.7 advances unified memory programming to automate CPU-GPU data movement, boost developer productivity, and simplify porting of ocean models like NEMO to GPUs on Grace Hopper–based systems.
Learn how to deploy and stream Omniverse Kit apps at scale across cloud and on-prem environments with flexible options, templates, and real-world deployments.
NVIDIA NeMo-RL v0.3 adds Megatron-Core support to boost training throughput for large models, addressing DTensor limitations, with long-context training, MoE support, and simplified configuration.
Guidance for deploying and streaming Omniverse Kit Apps at scale with flexible options across cloud, on-prem, and DGX Cloud.
Overview of NeMo-RL v0.3 with Megatron-Core backend for post-training large models, detailing 6D/4D parallelism, GPU-optimized kernels, and simplified configuration to boost reinforcement learning throughput on models at scale.
NVIDIA Streaming Sortformer is an open, production-grade, low-latency speaker diarization model that sorts speakers by arrival order and integrates with NVIDIA NeMo and Riva for real-time transcription pipelines.
NVIDIA’s co-packaged optics (CPO) approach delivers dramatic power efficiency for large-scale AI data centers, with Quantum-X Photonics and Spectrum-X Photonics enabling high-bandwidth, low-latency networking.
Overview of NVIDIA's Gamescom 2025 updates: DLSS 4, RTX neural rendering, RTX Kit, NvRTX, ACE AI, and related tooling for scalable, AI-driven gaming experiences.
Dynamo 0.4 introduces disaggregated serving, SLO-based autoscaling, AIConfigurator, and enhanced observability to accelerate large-model inference at scale with improved efficiency.
NVIDIA Research introduces ProRL v2, the latest evolution of Prolonged Reinforcement Learning for LLMs. It explores thousands of extra RL steps, new stabilization techniques, and broad benchmarking to push sustained improvements beyond traditional RL schedules.
CUDA-QX 0.4 advances quantum error correction workflows with automated detector error-model generation, a tensor-network decoder, improved BP+OSD decoding, and a Generative Quantum Eigensolver (GQE) in the Solvers library, accelerating end-to-end QEC development.
NVIDIA and partners unveil Wheel Variants, a new format to tailor Python wheels to exact hardware, enabling smoother CUDA-accelerated package installs and scalable packaging workflows.
NVIDIA announces general access for Isaac Sim 5.0 and Isaac Lab 2.2 at SIGGRAPH 2025, delivering new robot models, enhanced sensor simulation, ROS 2 integration, and cloud-based deployment via NVIDIA Brev.
NVIDIA's Omniverse NuRec pairs 3DGUT with COLMAP to turn real-world data into photorealistic USD assets for Isaac Sim and CARLA, accelerating sim-to-real testing and robot training.
NVIDIA Cosmos Reason is an open, fully customizable reasoning vision-language model for physical AI and robotics. It enables step-by-step multimodal reasoning and boosts robotics performance through post-training refinements.
NVIDIA unveils updates to Omniverse libraries, Cosmos WFMs, and interoperability tools to enable physically accurate robot simulations, OpenUSD workflows, and accelerated AI-enabled robotics development.
NVIDIA Research explores World Foundation Models (WFMs) for synthetic data generation and data curation to accelerate robot and industrial AI training, detailing Cosmos Predict, Cosmos Transfer, and Cosmos Reason and related workflows.
cuDF introduces JIT-compiled transforms to fuse kernels, reduce intermediates, and boost ETL throughput on GPUs. The 25.08 release adds ternary and string operators, with trade-offs around compilation time and kernel caching.
As multimodal AI models advance from perception to reasoning and autonomous action, new cognitive attack surfaces emerge. These threats target how AI systems solve problems across modalities, not just how they receive inputs.
XGBoost 3.0 enables terabyte-scale GBDT training on a single GH200 Grace Hopper Superchip, streaming data from host RAM via external-memory Quantile DMatrix for faster, simpler scaling.
CUDA Toolkit 13.0 introduces tile-based programming, unified Arm tooling across servers and embedded devices, and notable improvements in Nsight Compute, fatbin compression, and deployment workflows—positioning CUDA for future GPUs and Tensor Core architectures.
NVIDIA and OpenAI optimize gpt-oss-120b and gpt-oss-20b for accelerated inference on Blackwell, achieving up to 1.5M tokens per second on GB200 NVL72, with Day 0 support across cloud to edge.
NVIDIA vGPU 19.0 takes full advantage of the RTX PRO Blackwell architecture, enabling MIG-enabled graphics and AI virtualization with up to 48 concurrent VMs per GPU and new tools for AI workflows in virtualized data centers.
CUDA kernels often suffer from bandwidth bottlenecks. Vectorized loads and stores in CUDA C++ can boost bandwidth utilization, reduce instruction counts, and lower latency—when alignment and data-type constraints are observed.