How NVIDIA Uses JAX

Frontier model pre-training has converged on mixture of experts (MoE), which is fundamentally changing what limits large-scale AI training. As compute per token…

Large language model (LLM) training workloads increasingly run into GPU memory limits before compute is fully used. Model weights, gradients, optimizer states…

Mixture-of-experts (MoE) models have quickly become a foundational component of modern, large-scale AI systems. They are widely adopted because they enable…

Pre-training frontier LLMs comes down to throughput. When training spans trillions of tokens across thousands of accelerators, every percentage point of step…

NVIDIA CUDA 13.3 brings new capabilities and performance optimizations to developers across the CUDA ecosystem. The launch of NVIDIA CUDA Tile programming in…

For decades, computational chemistry has faced a tug-of-war between accuracy and speed. Ab initio methods like density functional theory (DFT) provide high…

Proteins rarely function in isolation as individual monomers. Most biological processes are governed by proteins interacting with other proteins…

Computer-aided engineering (CAE) is shifting from human-driven workflows toward AI-driven ones, including physics foundation models that generalize across…

CUDA 13.2 arrives with a major update: NVIDIA CUDA Tile is now supported on devices of compute capability 8.X architectures (NVIDIA Ampere and NVIDIA Ada)…

The article discusses the integration of the NVSHMEM communication library into the Accelerated Linear Algebra (XLA) compiler to optimize long-context model training in JAX.

The article discusses the NVIDIA Rubin platform, which introduces six new chips designed to create a powerful AI supercomputer.

The article discusses the NVIDIA ALCHEMI Toolkit-Ops, a specialized toolkit designed to accelerate AI-powered atomistic simulations in chemistry and materials science.

The article discusses Autodesk Research's development of the Accelerated Lattice Boltzmann (XLB) library, which enhances computational fluid dynamics (CFD) performance using NVIDIA's Warp and GH200...

The article discusses how NVIDIA's hardware innovations, particularly the Blackwell architecture and NVFP4 precision, along with their open source contributions, are driving advancements in AI.

The article discusses the introduction of Wheel Variants, a new Python packaging standard aimed at improving the installation and packaging workflows for CUDA-accelerated Python packages.

The article discusses techniques for optimizing low-latency communication in inference workloads using JAX and XLA, particularly focusing on the decode phase of large language models (LLMs).

NVIDIA cuQuantum is an SDK designed to accelerate quantum computing emulations significantly. The latest update, cuQuantum 25.

NVIDIA has announced world-record inference performance for the DeepSeek-R1 model using the Blackwell architecture, achieving over 250 tokens per second per user and a maximum throughput of over 30...

The article discusses the introduction of Gemma 3, a range of lightweight, multimodal, and multilingual models optimized for performance in AI applications.

New research from the University of Washington demonstrates how deep learning can enhance AI weather models, allowing for more accurate predictions and extending forecast capabilities up to 23 days...

The article discusses how to build a zero-copy AI sensor processing pipeline using OpenCV within the NVIDIA Holoscan SDK.

The article discusses the enhancements made in NVIDIA's cuDNN 9 library, focusing on the acceleration of Transformers through the implementation of Scaled Dot Product Attention (SDPA).

The article discusses NVIDIA cuQuantum 23. 10, an SDK designed to accelerate quantum circuit simulations using NVIDIA Tensor Core GPUs.

The article discusses the latest features of the NVIDIA NeMo framework and the performance enhancements brought by the NVIDIA H200 GPUs, which significantly improve the training of large language m...

The article discusses how NVIDIA Holoscan is being utilized to accelerate ptychography workflows at the Diamond Light Source, a leading synchrotron facility.

This article provides a comprehensive guide on deploying AI models in Python using the PyTriton interface with NVIDIA Triton Inference Server.

The article discusses how to efficiently scale large language model (LLM) training across a large GPU cluster using the open-source frameworks Alpa and Ray.

The article discusses RAPIDS RAFT, a library designed to optimize machine learning and data analytics on GPUs by providing reusable computational patterns.

The article discusses the use of NVIDIA BioNeMo Service for building generative AI pipelines aimed at drug discovery.

The article discusses the increasing computational demands for AI processing at the edge and introduces the NVIDIA Holoscan SDK v0.

The article discusses NVIDIA's BioNeMo service, a framework for training and serving biomolecular large language models (LLMs) designed for predicting protein structures and properties.

NVIDIA has announced significant updates to its AI software suite, including JAX, NVIDIA CV-CUDA, and NVIDIA RAPIDS, aimed at accelerating AI research, computer vision, and data science.

The article discusses the improved interoperability between NVIDIA Vision Programming Interface (VPI) and PyTorch, focusing on how VPI can enhance object detection and tracking in computer vision a...

The article introduces NVIDIA Warp, a Python framework designed for writing differentiable graphics and physics simulations on the GPU.

This article discusses the implementation of an end-to-end pipeline utilizing zero-copy techniques for efficient data transfer across various machine learning frameworks.

The article discusses the significance of tensor methods in modern machine learning, particularly their application in NVIDIA's AI algorithms.

This article discusses the importance of efficient memory layouts and memory pools in machine learning frameworks to enhance interoperability and performance.

The article discusses cuCIM, a new RAPIDS library designed for accelerated n-dimensional image processing and image I/O on GPUs.