Publisher
Stories with reporting from NVIDIA Developer Blog, ranked by what is gaining ground now. Each story links to the original articles.
NVIDIA describes using mPDLP in cuOpt to scale decision optimization to 100 million variables and beyond. The work targets expanding supply-chain and energy-grid problems with growing numbers of items and constraints.
NVIDIA introduces DLSS 5 with 3D-Guided Neural Rendering and granular controls for game developers to add lifelike lighting and material detail. The update also covers NVIDIA ACE and new RTX Kit capabilities.
A report describes how robots are trained to assemble GB300 tester trays and what the process reveals about robot learning and mechanical intelligence. It also discusses the role of conventional engineering in the work.
CUDA Toolkit 13.4 adds Windows on Arm support and greater control over shared GPUs. The release extends the platforms and GPU-management options available to CUDA developers.
NVIDIA announced an open software platform and reference system design for adding safeguards to autonomous AI agents. The described design combines a host-CPU runtime with an out-of-band watchdog on a DPU; one source characterizes it as continuous in-silicon monitoring, while the other describes the watchdog as something an agent cannot see.
NVIDIA introduces cuPhoton for scientific image analysis.
NVIDIA describes Groq 3 LPX deterministic execution for power-efficient, interactive inference on NVIDIA Vera Rubin. The supplied information does not provide further technical details about the system.
NVIDIA introduces two approaches to writing GPU kernels in Rust, expanding its native GPU programming options alongside CUDA C++ and CUDA Python. The announcement positions Rust as an additional route for developers building CUDA applications.
AICR reaches version 1.0 as an open, stable, and verifiable configuration approach for GPU-accelerated Kubernetes clusters. It addresses compatibility across components that follow separate release cycles, including host kernels and GPU drivers.
NVIDIA introduces cuObject and the NVIDIA SCADA Server SDK to expand AI infrastructure access to file and object storage. The effort targets infrastructure engineers, storage developers, and cloud service providers that need secure access to high-capacity storage.
NVIDIA describes Green Contexts as a way to control how GPU applications share work when multiple independent components run within one process. The feature is aimed at applications balancing latency-sensitive operations with other GPU tasks.
NVIDIA introduces DOCA Agent Skills to help AI agents work with BlueField infrastructure software. The skills target a gap in general-purpose agents’ ability to handle specialized development tasks.
NV-Reason-CT Open is a 3D CT vision-language model designed to support radiologist chain-of-thought reasoning. The announcement concerns applying AI reasoning to volumetric CT, extending beyond 2D scans and other imaging tasks.
NVIDIA presents a confidential-computing approach for production AI inference that addresses sensitive user information and proprietary model context. It targets private inference across personal, enterprise, and regulated settings.
NVIDIA describes multi-device inference integration for TensorRT in NVIDIA Dynamo-Triton, targeting generative AI workloads that exceed a single GPU's compute or memory capacity. The approach is intended to simplify serving models across multiple GPUs.
NVIDIA presents DOCA GPUNetIO as a way to coordinate GPU-initiated networking across its software stack. The focus is making networking and data movement operate as GPU-controlled functions rather than host-driven services.
The piece presents NVIDIA OpenShell as a way to add runtime controls to AI agents that can write code, use tools, and continue working as new information arrives. It focuses on controlling agents that pursue assigned goals over time.
SWE-Serve evaluates AI coding-agent changes to inference-serving software in live serving conditions, where a model loads and handles requests. It highlights a gap between patches that pass local tests and those that work in production-like operation.
The piece describes using NVIDIA BioNeMo Inference Runtime for high-throughput biomolecular structure prediction at proteome scale. Its focus is efficiently processing an entire worklist through the prediction pipeline.
NVIDIA BioNeMo NIM microservices are made available for protein structure prediction in Claude Science.
