NVIDIA-NeMo
16 Tools · 93 Intel
Tools
NVIDIA's Python API for defining large language models and running optimized inference on NVIDIA GPUs.
World foundation models, datasets, and tooling packaged together for physical AI and robotics work.
Real-time full-duplex speech-to-speech model with persona control via text role prompts and voice conditioning.
Reference architecture for GPU-accelerated vision agents that search, summarize, and answer questions about video.
NVIDIA foundation model for generalist humanoid robots, with perception-to-action policies that transfer across bodies.
NVIDIA's security scanner that inspects AI agent skills for vulnerabilities and malicious patterns.
AI weather and climate prediction workflows get built and deployed through this open-source deep-learning framework.
Lightweight C++ inference engine that runs LLMs and vision-language models on edge devices.
Text normalization and inverse text normalization toolkit built for ASR and TTS pipelines.
Collection of logits processors that customize LLM generation, working with vLLM, Hugging Face, and TensorRT-LLM.
Rust proxy that routes LLM traffic across providers and translates OpenAI and Anthropic APIs.
NVIDIA GEAR Lab initiative applying world models to the robotics data problem in embodied AI training.
Intel
As context lengths grow, attention takes a larger share of inference time, so how attention is designed rather than merely implemented increasingly determines interactive long-context performance.
A triage framework for GPU clusters running AI workloads: which telemetry tool owns which failure domain, a worked gray-failure case where one degraded InfiniBand link stalls NCCL collectives and slows a multi-day training job, and a short alerting set tied to service objectives and remediations.
NVIDIA introduces two paths for writing GPU kernels in Rust: cuda-oxide, a custom rustc backend compiling SIMT kernels to PTX, and cutile-rs, a stable-Rust tile-based model already used in HuggingFace's Grout and mistral.rs, both enforcing memory safety at compile time.
NVIDIA walks through quantization-aware distillation for Nemotron 3.5 Lightning. Post-training quantization produces a low-precision student, then distillation against the frozen teacher recovers the lost accuracy, taking the checkpoint from 66 GB to 22 GB at up to four times the throughput.
NVIDIA and Hermes Agent authors show how NeMo Relay captures lifecycle events and trajectories, exports OpenTelemetry spans to Phoenix, and supports controlled harness evaluation, including a 108-run comparison of baseline and fixed revisions.
A walkthrough of how GPU confidential computing actually works — hardware-fused signing keys, remote attestation before any secret reaches the VM, encrypted interconnect — with benchmarks showing the security tax on large-model inference lands under 8%, and the scheduler and autotuning fixes needed
NVIDIA and Princeton's Dao AI Lab describe ReplaySSM, which stores SSM state as a checkpoint plus buffered inputs. It cuts state memory traffic, speeds decoding kernels on Blackwell Ultra, simplifies speculative rollback, and limits low-precision state error.
CUDA Python 1.0 ships with CUDA 13.3 under semantic-versioning commitments: cuda.core, cuda.compute, cuda.bindings, nvmath-python and cuda-pathfinder as one maintained foundation, with resource sharing between GPU libraries and platform features like green contexts and process checkpointing exposed
NVIDIA's guide to deploying compact reasoning models like Nemotron 3.5 Lightning and Qwen3.8-27B on Jetson edge hardware, showing NVFP4 quantization and model-specific speculative decoding configs deliver up to 6.28x decode throughput speedup over BF16.
NVIDIA lays out the OpenShell reference architecture for governing agent workloads. A gateway manages sandbox lifecycles and credential access, a supervisor isolates workloads, and a policy prover verifies permissions, with Vera CPU and BlueField hardware options.
NVIDIA details an agentic pipeline where Codex or Claude orchestrates specialized 'Hermes' subagents via NemoClaw to prep Blender scenes for robotics simulation: adding physics, sensors, and semantic labels through OpenUSD, then validating with SimReady before handoff to Isaac Sim.
NVIDIA details how Transformer Engine in JAX delivers a 10.4x throughput gain for dropless Mixture-of-Experts training, raising DeepSeek-V3 performance from 103 to 1,068 TFLOPS/GPU via grouped-GEMM kernels and NCCL EP-fused dispatch.
NVIDIA's Nemotron Model Reasoning Challenge asked Kagglers what improves reasoning accuracy when everyone starts from the same open model, benchmark, infrastructure and constraints. What 5,000-plus entrants found.
Agents now inspect code, run tests, read documents, query internal systems and operate for hours on a user's behalf, which is the productivity case and the access problem at once. How to govern them inside an enterprise.
NVIDIA's KGMON team details its second-place KDD Cup data agent: a unified SQLite interface, schema scouting before reasoning, a narrow toolset with call repair middleware, isolated document extraction, and trace inspection.
CUDA Toolkit 13.4 adds Windows on Arm support, preview support for NVIDIA's next-generation Rubin architecture, a rewritten Multi-Process Service with cgroup-integrated GPU memory limits, and CCCL 3.4 updates reaching 92% memory-bandwidth utilization on Blackwell.
NVIDIA Dynamo previews shadow engine recovery: a GPU Memory Service allocates weights independently of engine processes, so a fully initialized standby maps the same weights on the same GPU. On GLM-5.2 with B200 nodes, failover fell from a 283-second cold restart to 7.3 seconds.
NVIDIA published agent skills for the DOCA SDK (Flow, GPUNetIO, PCC, RDMA) that supply verified signatures and hardware checks. A 65-prompt eval and a BlueField-3 RDMA demo show how domain skills cut guesswork and handwritten code.
Two clusters built from identical H100 or GB200 NVL72 systems can differ materially in training throughput, with NVIDIA routinely seeing 8-12% gaps between partner deployments and its own reference architecture.
NVIDIA packaged BioNeMo's folding, genomics and chemistry models as agent-callable skills wired into Claude Science and NIM microservices. A worked comparison of OpenFold3 and Boltz-2 on a Seh1 complex shows interface confidence collapsing when alignment input is removed.
Throughput and interactivity trade against each other at fixed accuracy, and a surprising amount of that frontier is decided by architecture choices made before training: layer aspect ratios, dimensions aligned to GPU tile sizes, NVFP4-friendly linear layers, and how MoE parallelism is split across
NVIDIA details Sentry, which runs on BlueField-4 DPUs to evaluate agent requests against Open Policy Agent rules, introspect agent memory, check container and SBOM integrity, and contain violations in milliseconds, independent of host software.
NVIDIA engineers had a coding agent build a real-time endoscopic segmentation app on Holoscan and HoloHub, iterating against engineer-defined goals and benchmarking each round. An ablation compared documentation alone, CLI alone, and CLI plus skills plus docs.
NVIDIA details Muse Spark Open inference on Blackwell Ultra: 42B active of 816B parameters, top-8 routing over 256 experts, a 1.8B perception encoder, NVFP4 checkpoints, and a NIM container. Throughput numbers are NVIDIA's own claims.