Articles by Category: Technical_deep_dives

Stop chunking tables: How we built an agentic GraphRAG for financial disclosures with Docling

2026-07-22 07:01
🚀 Exciting advancements in financial data retrieval! A recent article explores the development of an agentic GraphRAG designed to enhance question-answering over SEC filings. Traditional methods often struggle with structured data like XBRL and HTML, leading to inaccuracies. The new approach utilizes Docling to create aligned knowledge graphs, improving navigation and synthesis of information. This multi-stage framework enhances the accuracy of answers while providing clearer insights into...
Vincent Caldeira

Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI

2026-07-21 15:00
🚀 The NVIDIA Rubin GPU architecture is transforming AI capabilities by enabling continuous, large-scale intelligence production. These advancements support complex tasks with sustained reasoning, requiring efficient processing and low latency. The Rubin platform features enhanced Tensor Cores and a powerful memory subsystem, achieving significant improvements in energy efficiency and performance. #NVIDIA #AI #TechInnovation #GPUs #AgenticAI
Tanya Lenz

NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI

2026-07-21 15:00
NVIDIA’s new Vera CPU features the Olympus core, designed for optimal single-thread performance in Agentic AI. This CPU shifts critical execution tasks to enhance responsiveness and throughput in AI operations. It focuses on strong single-thread performance, memory bandwidth, and predictable latency. Olympus was co-designed with the entire Vera Rubin platform to maximize efficiency across AI infrastructure workloads. For more technical insights, check out the NVIDIA Vera CPU white paper....
Michelle Horton

Setting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72

2026-07-21 15:00
NVIDIA has set a new world record for pre-training with its GB300 NVL72, achieving 1,648 TFLOPs per GPU while training the DeepSeek-V3 model, which has 671B parameters. 💻 This advancement highlights the shift to mixture of experts (MoE) architectures, which allows for more efficient computation by activating only a subset of parameters per token. 🔍 However, the need for extensive communication between GPUs poses challenges, as delays can affect throughput during training. 📈 #NVIDIA #AI...
Kirthi Devleker

Operationalize AI agents with OpenShift and Kubernetes primitives

2026-07-21 07:16
Unlock the potential of AI agents with Red Hat OpenShift and Kubernetes! 🌐 In the latest article, key strategies on operationalizing agentic AI in production-ready AIOps systems are discussed. This includes swapping models, updating prompts, and injecting institutional knowledge without downtime. Kubernetes primitives like ConfigMaps and PersistentVolumeClaims help streamline operations, allowing for quick updates and zero-downtime rollouts. Discover how to enhance your AI infrastructure...
Tony Kay, Ishu Verma

NVIDIA NVLink: The Scale-Up Network for AI Factories

2026-07-20 15:46
NVIDIA NVLink is revolutionizing AI factories by addressing the increasing demand for complex AI workloads. As models grow larger, the need for high-bandwidth, low-latency GPU communication becomes essential. NVLink facilitates this, allowing multiple accelerators to operate as a cohesive unit. This technology is crucial for efficient AI inference and training, ensuring resilience and high performance in data centers. #NVIDIA #AI #DataCenter #TechInnovation #NVLink 🤖💻🔗
Elizabeth Goodman

Integrate NVIDIA Omniverse RTX Sensor Simulation Into Existing Apps

2026-07-20 15:00
Developers in 3D design, simulation, and robotics can enhance their applications with NVIDIA Omniverse RTX Sensor Simulation. The new ovrtx library allows for generating sensor outputs from OpenUSD content using a lightweight SDK. This integration supports seamless workflows, enabling teams to visualize data and validate systems effectively. Learn more about optimizing your applications with NVIDIA's tools! 🌐🔧 #NVIDIA #Omniverse #3DDesign #Simulation #AI
Tanya Lenz

ML-KEM vs Diffie-Hellman: The Showdown – IPsec Series, Part 3

2026-07-20 14:00
In the latest IPsec Series article, the comparison between Diffie-Hellman and ML-KEM highlights key aspects of both key exchange methods. Diffie-Hellman (X25519) is small and quick but faces quantum vulnerabilities. In contrast, ML-KEM offers quantum safety but with larger sizes that can cause fragmentation in messages. The article suggests a hybrid approach, utilizing both methods to enhance security. 🔒📡 Read more to explore their performance across various metrics! #CyberSecurity...
Julio Gomez

Escape Analysis in Go – Stack vs. Heap Allocations Explained

2026-07-20 10:15
🔍 Escape analysis in Go is a key optimization technique that helps the compiler decide if values should be allocated on the stack or the heap. When a value "escapes," it means it needs to persist beyond the function's lifetime, leading to heap allocation. Understanding this process is crucial for optimizing performance. While Go abstracts memory management, knowing how to influence escape analysis can improve your code's efficiency. #GoLang #Programming #EscapeAnalysis #MemoryManagement...
Dominika Stankiewicz

Does “rtk” skill really cut agent tokens by 60–90%? We tested it

2026-07-20 10:11
🚀 In the latest article, the effectiveness of the “rtk” skill in reducing Claude Code token usage is examined. The claimed savings of 60-90% were tested but resulted in only a 7.6% increase in costs at low reasoning effort, with no changes at high effort levels. The task quality remained consistent across tests. This study continues the exploration of token-saving add-ons for coding agents, previously assessing the "caveman skill." #TechTesting #CodingEfficiency #TokenSavings #AIResearch...
Denis Shiryaev

Architect an open blueprint for cloud-native AI agents

2026-07-20 07:01
Explore the architecture for cloud-native AI agents! This article outlines a blueprint for safe and efficient operation of autonomous agents, focusing on their distinct components and deployment patterns. Key aspects include the separation of lifecycles, security requirements, and the role of Red Hat AI in hosting these agents. The solution emphasizes open interfaces and interchangeable components to enhance flexibility and security. 🔍💻🌐 #AI #CloudComputing #RedHat #TechInnovation...
Fatih E. Nar, Sally O'Malley, Adel Zaalouk

PyTorch distributed is changing and TorchComms is why

2026-07-20 03:16
🚀 PyTorch is evolving with the introduction of TorchComms, a new backend layer designed to replace the old c10d. This change aims to enhance training and inference systems by providing a communicator-centric architecture. 🔄 Users can transition without rewriting existing code, allowing for a smoother migration path while maintaining core functionalities like ProcessGroup and DeviceMesh. 📊 Initial results show significant improvements in startup times and inference efficiency, particularly at...
Arkadip Maitra

What 429 chaos experiments taught us about Kubernetes operator resilience

2026-07-20 03:16
🚀 In the latest chaos experiments on Kubernetes operators, 429 tests were conducted across 22 operators. Most operators passed standard tests like PodKill, but deeper tests like DeploymentScaleZero revealed critical gaps in resilience. Some operators failed to recover from network partitions and other failures, indicating issues with their reconciliation logic. In addition, SDK-level testing uncovered bugs that CLI tests missed, highlighting the importance of comprehensive testing strategies....
Ugo Giordano

Self-healing GPU nodes in Kubernetes: What we learned building the EKS node monitoring agent

2026-07-19 13:00
🚀 Managing Kubernetes at scale often involves dealing with frequent node failures, especially with GPUs. To address this, the EKS Node Monitoring Agent was developed to automate node detection and replacement, minimizing human intervention. This open-source tool allows Karpenter to replace failing nodes automatically, enhancing efficiency. AWS has also introduced Amazon EKS Auto Mode, which automates various cluster operations, including automatic node repair. This means less focus on...
Sajjan Gundapuneedi

From intent to enforcement: Lessons from operating Kubernetes controllers at scale

2026-07-17 12:00
Kubernetes controllers are essential for maintaining the platform’s declarative model. They observe system states and ensure intent is met, even as conditions change. However, operating at scale introduces challenges, particularly in maintaining correctness while dealing with cache lag and object churn. In Amazon EKS, two key controllers address these issues: the Network Policy Controller and the VPC Resource Controller. The former manages traffic within the cluster, while the latter assigns...
Sri Saran Balaji Vellore Rajakumar

Benchmark Red Hat Data Grid in OpenShift 4 using Hyperfoil

2026-07-17 07:15
🚀 Discover how to benchmark Red Hat Data Grid in OpenShift 4 with Hyperfoil! This article explains the process of benchmarking using the Red Hat Data Grid operator and Helm charts. Hyperfoil is an open-source tool designed for distributed performance testing, allowing users to measure memory footprint, latency, and throughput. Learn how to install the Hyperfoil operator and create custom benchmarking resources. Adjust JVM settings to enhance performance and manage your cluster effectively....
Francisco De Melo Junior, Alexander Barbosa Ayala

Q&A: How Capcom Brought Path Tracing to RE ENGINE Across PRAGMATA and Resident Evil Requiem

2026-07-16 22:59
Capcom's RE ENGINE team has successfully integrated path tracing into Resident Evil Requiem and PRAGMATA, enhancing their visual experiences. Over two years, they developed a game-oriented path tracer optimized for DLSS, improving direct lighting and reducing visual gaps between gameplay and cutscenes. 🌟 This transition reflects a commitment to immersive gameplay, leveraging advanced NVIDIA technologies for realistic lighting and reflections. #Capcom #REENGINE #PathTracing #Gaming #ResidentEvil
Michelle Horton

Why smarter AI caching sometimes makes everything slower

2026-07-16 16:00
Caching plays a vital role in AI systems, especially in Retrieval-Augmented Generation (RAG) pipelines. While initial implementations performed well, real production workloads revealed issues like tail latency and increased costs. Traditional caching methods, such as Redis, initially improved response times but faltered with semantic queries. When queries varied in wording, Redis struggled, leading to redundant data and higher bills. The shift to semantic caching with vector databases aimed...
Zziwa Raymond Ian

Layered sandboxing for AI agents: OpenShift and OpenShell

2026-07-16 07:01
AI coding agents require capabilities that can expose them to various security risks. Red Hat OpenShift and NVIDIA OpenShell are two solutions that address these challenges by providing layered sandboxing. OpenShell wraps agents in a policy-enforced sandbox, controlling outbound connections and enforcing file system isolation. Meanwhile, OpenShift sandboxed containers use lightweight VMs to isolate workloads, protecting the host from kernel exploits. Together, they mitigate different threat...
Eitan Geiger

How obs-mcp boosts AI-native OpenShift observability

2026-07-16 07:01
Unlock the future of observability with the Model Context Protocol (MCP) for Red Hat OpenShift! 🚀 The obs-mcp component enables AI models to access diverse data sources, enhancing decision-making in complex environments. It automates data correlations, reducing manual workflows for engineers during incidents. Key features include: - Structured interfaces for efficient querying. - Toolsets for metrics, logs, and traces. - Guardrails for safe data access. Explore how obs-mcp transforms...
Vanessa Martini, Ivan Necas, Saswata Mukherjee

Develop Lightweight USD Runtimes Faster with AI Agents

2026-07-15 21:57
Unlock the potential of lightweight USD runtimes with AI agents! 🌐 OpenUSD is a framework that allows teams to integrate CAD data, simulations, and real-world data into a unified view. Traditionally, creating a USD implementation required extensive code adaptation. Now, with nanousd-labs, developers can generate runtimes directly from the USD Core Specification. This method streamlines the implementation process, making it faster and more efficient for physical AI applications. Explore how AI...
Michelle Horton

Building Faster Cryptography with Carryless Multiplication in NVIDIA CUDA 13.3

2026-07-15 17:37
NVIDIA CUDA 13.3 introduces native support for carryless multiplication with the new clmad PTX instruction, benefiting NVIDIA Ampere and newer GPUs. This enhancement significantly boosts cryptographic workloads. For instance, GHASH achieves a throughput of ~6.3 TB/s, making it up to 18.8x faster than previous methods. These advancements support critical applications like AES-GCM for secure communications and zero-knowledge proofs. 🔒💻⚡️ #NVIDIA #CUDA #Cryptography #TechInnovation #GPUs
Michelle Horton

“The database is the product”: What breaks when memory devices scale

2026-07-15 12:00
Imagine relying on an AI device to capture your meeting notes, only to find it struggles to retrieve them when you need them most. This highlights a critical issue with AI hardware—data retrieval can be the weak link, even when transcription is flawless. 🤖💼 The article examines the challenges faced by Plaud, a leading AI notetaker. While their architecture seemed sound initially, separating structured metadata and unstructured content led to delays at scale. This demonstrates that what works...
Ed Huang

Red Hat build of Agent Sandbox: Isolated workload management with Kubernetes

2026-07-15 07:01
🚀 The Red Hat build of Agent Sandbox simplifies workload management in isolated environments. Developers can request sandboxes without needing to understand the underlying technology, allowing them to focus on application development. Key features include lifecycle management, hardware-assisted isolation with Kata Containers, and near-instant provisioning through warm pools. This system is not limited to AI workloads and supports various applications requiring complete isolation. #RedHat...
Emanuele Giuseppe Esposito, Jens Freimann

Red Hat OpenShift 4.22: What dynamic plugin developers need to know

2026-07-14 13:03
🚀 Red Hat OpenShift 4.22 is here! Dynamic plugin developers must migrate to new dependencies to ensure compatibility with the updated web console. Key changes include updates to React, React Router, and Redux versions. Refer to the release notes for migration examples and use the Claude skill for assistance. Explore more resources to ease your transition! #RedHat #OpenShift #WebDevelopment #DynamicPlugins #SoftwareUpdate
Jackson Lee

Post-mortem GPU crash debugging with LLMs

2026-07-14 12:00
🚀 Exciting advancements in GPU crash debugging! AMD has introduced the Radeon GPU Detective (RGD) MCP Server, which connects LLM agents to its crash analysis pipeline. This tool simplifies post-mortem debugging, allowing developers to analyze GPU crash dumps with a single prompt. By using structured access to crash data, LLMs can quickly identify root causes and suggest fixes, enhancing efficiency for developers. Discover how this innovation transforms the debugging process! #GPUDebugging...

Simplify your performance monitoring with the pmlogger PUSH model

2026-07-14 03:01
Red Hat Enterprise Linux (RHEL) 10.2 introduces the pmlogger PUSH model, making performance monitoring simpler. With the new PUSH model, remote clients can send data to a central collector system without additional configuration, streamlining the process for users. 📈 The Performance Co-Pilot (PCP) and Grafana continue to support data collection and visualization, enhancing system monitoring capabilities. For more details, check out the full article. 📊💻 #RedHat #PerformanceMonitoring #RHEL10...
Christian Horn

NVIDIA Ising Decoding Cuts Color Code Logical Error Rates by Over 300X

2026-07-13 19:00
🔍 Researchers are advancing quantum error correction (QEC) codes for fault-tolerant logical operations in quantum computers. Color Codes, though previously overlooked due to decoding challenges, are now back in focus. NVIDIA's Ising Decoder has improved logical error rates by over 347x, making these codes more viable for practical use. The Ising Decoding pipeline enhances efficiency and accuracy for QPU developers, paving the way for better quantum computing capabilities. #QuantumComputing...
Elizabeth Goodman

Extreme Event Likelihoods with Guided Generative Models

2026-07-13 15:00
🌍 Understanding low-likelihood, high-impact events is crucial in fields like science, engineering, and finance. Traditional methods like Monte Carlo sampling can be inefficient for estimating these risks. 🔍 A new paper explores using guided diffusion models to improve probability estimates for rare climate events, specifically tropical storms. This approach addresses the challenge of oversampling in guided models. ⚙️ By using tools like NVIDIA cBottle, researchers can compute odds ratios that...
Elizabeth Goodman

Why killing pods is not enough: Testing operator reconciliation with operator-chaos

2026-07-11 03:01
🔍 Exploring operator-chaos: An open-source framework for testing Kubernetes operators. This article highlights limitations of traditional chaos tools that focus mainly on pod recovery. Operator-chaos addresses critical gaps by testing operator reconciliation through 20 failure injection types across four categories. Key features include: - Four injection modes from CLI to action interceptors. - Knowledge models for tailored failure testing. - Structured verdicts to assess operator behavior...
Ugo Giordano

Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading

2026-07-10 18:17
🚀 Large language model (LLM) training faces GPU memory limits that hinder compute efficiency. Key components like model weights and gradients compete for high-bandwidth memory (HBM). 📊 The article discusses host offloading in JAX, which alleviates HBM pressure by moving activations to pinned host memory during training. This method is particularly effective on NVIDIA Grace Blackwell systems, thanks to their high-speed connectivity. 🔍 Performance tests using MaxText workloads demonstrate the...
Tanya Lenz

Kernel Fusion in NVIDIA CUDA: Optimizing Memory Traffic and Launch Overhead

2026-07-10 16:41
Discover how kernel fusion can optimize GPU performance in NVIDIA CUDA! 🚀 This technique improves memory bandwidth and reduces kernel launch overhead by combining multiple GPU operations into a single kernel. This means less time wasted on memory transfers and faster execution. CUDA Graphs also play a role, capturing sequences of operations for efficient execution, though they don't fuse kernel bodies. Both methods can work together for enhanced performance. Learn more about applying these...
Michelle Horton

AI Model Co-Design: Hardware-Friendly LLM Design

2026-07-10 16:36
AI performance is determined by three key dimensions: 🔍 **Accuracy**: How well a model reasons and produces outputs. ⚡ **Throughput**: The number of tokens generated per second. ⏳ **Interactivity**: The responsiveness of the model, influenced by latency. Successful deployments must balance these dimensions. High accuracy is ineffective if responses are slow, and throughput without a good user experience falls short. This article explores how design choices impact throughput and interactivity...
Elizabeth Goodman

What running Kubernetes across millions of clusters taught AWS about zone failures

2026-07-10 15:00
AWS has learned valuable lessons about zone failures while managing Kubernetes across millions of clusters. They emphasize that not all failures are obvious; some zones may slow down, causing issues without disappearing completely. This often leads to cascading problems when automated systems react inappropriately during these gray areas. To combat this, AWS has built zonal resiliency into Amazon EKS, focusing on static stability—preserving capacity and routing around bad zones during...
Raghav Tripathi

Accelerating End-to-End Co-Folding Performance with NVIDIA BioNeMo Agent Toolkit

2026-07-10 13:00
Unlock the potential of biomolecular structure prediction with the NVIDIA BioNeMo Agent Toolkit! 🌟 This toolkit enhances workflows for drug discovery and protein design by speeding up key processes like Multiple Sequence Alignment and co-folding inference. 🧬 With advanced acceleration on NVIDIA B300 and H100 GPUs, researchers can now efficiently handle large compound libraries and predict complex molecular assemblies. Discover how these innovations are transforming the field! #Biotechnology...
Elizabeth Goodman

Troubleshoot Red Hat OpenShift Virtualization localnet with the netobserv command

2026-07-10 03:01
🚀 Troubleshooting network visibility in Red Hat OpenShift Virtualization is vital for VMs communicating over localnet. The article discusses using the `oc netobserv` command-line tool to troubleshoot four scenarios, focusing on VLANs, DHCP issues, MTU, and dropped packets. Key points include how to confirm VM VLAN placement, diagnose DHCP lease problems, and analyze MTU mismatches. For detailed insights and practical scenarios, check out the full article! 📊🖥️ #OpenShift #Virtualization...
Moyo Oyegunle

The “silent hallucination” loop: how our autonomous data pipeline poisoned its own vector store

2026-07-09 13:00
🚨 In a recent article, the author discusses challenges faced with an AI system for a fintech client. Initially, the system worked well, extracting data from PDFs. However, it soon generated inaccurate chatbot responses, citing outdated financial data and incorrect revenue attribution. The root cause was a faulty data ingestion process that mismanaged probabilistic outputs, leading to "hallucinations" in the database. Despite implementing a validation layer, issues persisted due to reliance on...
Emmanuel Akita

EvalHub: Capability and safety benchmarking for AI models

2026-07-09 07:01
EvalHub is a new service for evaluating large language models (LLMs) on Red Hat AI. It streamlines testing by running multiple benchmarks through scalable Kubernetes jobs. This service assesses models on three key aspects: capability, safety, and performance. By using a unified collection of benchmarks from various providers, it offers a comprehensive evaluation with a single submission. Key features include tracking experiment lineage and maintaining performance metrics as OCI artifacts....
William Caban Babilonia

Tune and troubleshoot Red Hat Data Grid cross-site replication

2026-07-09 07:01
🚀 Red Hat Data Grid enables efficient in-memory caching and cross-site replication across OpenShift clusters. The Data Grid operator facilitates this by deploying clusters and caches through user-configurable resources. By utilizing multi-site installations, data is safeguarded during outages in various locations. Key configurations include Route types and load-balancer services for secure communication. For detailed insights on deployment and troubleshooting, check the full article! #RedHat...
Francisco De Melo Junior, Alexander Barbosa Ayala

Running Low-Latency Analytical Workloads with GPU-Accelerated Presto on NVIDIA GB200 NVL72

2026-07-08 16:05
Unlock the power of analytical workloads with GPU-accelerated Presto! 🚀 Presto is an open-source SQL engine that runs fast queries on large datasets. By utilizing NVIDIA GPUs, it offers low latency, enhancing user experience and productivity. This post showcases multi-GPU execution on NVIDIA's DGX B200 and GB200 NVL72, highlighting the impact of NVIDIA GPUDirect Storage for high I/O throughput. 📊 Learn how GPU-accelerated Presto achieves peak performance using NVIDIA cuDF algorithms and...
Tanya Lenz