V Programming Tutorials & Engineering Articles

To address the quadratic latency bottleneck of self-attention in high-resolution video diffusion models, developers implemented Sparse VideoGen (SVG) to dynamically route attention heads to highly structured spatial or temporal sparse masks. Translating this algorithmic sparsity into physical hardware speedups on TPUs required optimizing the Splash Attention kernel by bypassing empty memory tiles, restricting exact coordinate masking strictly to boundary tiles, and permuting token memory layouts into temporal-major order for contiguous access. By aligning these sparse masks with actual hardware tile execution, the combined optimizations significantly reduced wasted matrix operations and achieved up to a 1.69x end-to-end inference speedup for 1440p video generation.

The Oxide Cloud Computer, purpose-built rack-scale infrastructure, is now available to government agencies through Carahsoft and its reseller partners.

MTIA 300 is the first of Meta’s family of in-house training and inference accelerators optimized for training ranking and recommendation models. We’re sharing how MTIA 300’s built-in NIC chiplets a…

Teams customize their models to hit their targets for latency, speed, memory, and compute. With the open NVIDIA Nemotron family of models…

Meta’s Generative Ads Recommendation Model (GEM), the foundation model behind ads recommendations across Instagram and Facebook, now trains at LLM scale on several thousand of the latest-gene…

As agentic and long-context workloads become common, the context lengths increase and attention consumes a larger share of inference time (Figure 1).

What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale.

Developers building 3D, design, simulation, robotics, and industrial digital twin applications need ways to bring physical AI capabilities into the tools and…

We use clickhousectl to spin up multiple ClickHouse versions side by side and benchmark two recent performance improvements.

Training an LLM is the easy part. The hard part is designing experiments and evaluations that you can trust enough to know whether the new…

What if autonomous coding AI agents could push your vision reasoning models above 90% accuracy with almost no manual effort? When adapting vision reasoning…

AI performance comes down to three dimensions: Deployments must balance all three: High accuracy is wasted if responses are slow, and raw throughput means…

Using population-level analysis to debug tricky crashes in our data infrastructure.

Generative AI workloads are rapidly outgrowing the memory and compute budget of single GPUs. For inference developers building media generation pipelines…

Physical AI—robots working autonomously alongside people in factories, warehouses, hospitals, and homes—is arriving faster than most expected.

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

Quick glossary for readers new to VLA/WAM terminology VLA Vision-Language-Action model: a robot policy that starts from a pretrained VLM backbone and adapts it…

NVIDIA CompileIQ tackles one of the hardest problems in performance engineering: finding the compiler options that unlock the best performance for a specific…

The Google Tensor ML SDK is graduating to its Beta phase, allowing developers to build and deploy high-performance machine learning models directly onto the TPU of Google Pixel 10 devices. By integrating with LiteRT, Google's edge deployment framework, the SDK provides a unified workflow for developers to convert, compile, and run PyTorch or TFLite models with robust fallback options. Additionally, a new model garden offers over 100 classic and generative AI models, including Gemma 3, enabling low-latency, private features like speech recognition, computer vision, and text generation.

This post is the second of a three-part series. See also Model Quantization: Concepts, Methods, and Why It Matters and Model Quantization: Turn FP8 Checkpoints…

Creative and visualization teams today produce more assets, in more formats, with leaner teams. Generative AI can accelerate that work – compressing tasks that…

This is the story of how our Developer Productivity team extended and rolled out rubyfmt, our Rust-based zero-config, ultra-fast autoformatter across the world's largest Ruby codebase.

The development of socially acceptable nuclear reactors requires that they are safe, clean, efficient, economical, and sustainable.

In vision AI systems, model throughput continues to improve. The surrounding pipeline stages must keep pace, including decode, preprocessing, and GPU scheduling.

Meta continues to lead the industry in utilizing groundbreaking AI Recommendation Systems (RecSys) to deliver better experiences for people, and better results for advertisers. To reach the next fr…

In the rapidly evolving landscape of large language model (LLM) development, NVIDIA Megatron Core has emerged as the foundational framework for training massive…

In this post, we dive into one of the most critical workloads in modern AI: Flash Attention, where you’ll learn: Environment requirements: See the quickstart…

How multimodal post-training was engineered for production at Microsoft, system-level interventions that reshaped policy gradient stability, and Copilot agent training at scale.

The article discusses NVIDIA TensorRT LLM AutoDeploy, a beta feature that automates the inference optimization process for large language models (LLMs).

This article introduces Dynamic Context Parallelism (Dynamic-CP), a scheduling approach in NVIDIA Megatron Core designed to optimize training for variable-length sequences in large-scale models.

PinLanding is a multimodal AI pipeline developed by Pinterest to generate shopping collections from billions of products.

The article discusses the advancements in AI technologies and infrastructure that shaped the year 2025, focusing on NVIDIA's innovations in AI factories, physical AI, and model optimization.

The article discusses the Skip Softmax technique, a method for accelerating long-context inference in large language models (LLMs) using NVIDIA TensorRT-LLM.

The article highlights Alexey's favorite features introduced in ClickHouse throughout 2025, including lightweight updates, data lake support, and advancements in text and vector indexing.