NVIDIA GTC Taipei at COMPUTEX

Advancing Video Analytics With AI Agents

Speakers pending

The clip states that artificial intelligence has made extraordinary progress over the last 10 years and introduces Nvidia's new DGX product, called DGX Spark, noting it has 20 CPU cores. The source provides only these high-level claims and a short musical intro, offering no additional hardware, software, or benchmark details. As a result, engineers can register the product name and a CPU-core count but cannot assess system capability, performance, or suitability for specific video-analytics or AI workloads from this transcript alone.

Building Digital Twins of Foxconn’s Robotic Factories

Speakers pending

Foxconn addresses rising demand for NVIDIA accelerated computing by building robotic factories using NVIDIA Omniverse as a digital twin and AI Factory planners. Omniverse ingests facility and equipment data from Siemens Teamcenter X and Autodesk Revit to optimize floor layout, line configurations, and camera placement, and serves as a robot gym to simulate and train NVIDIA Isaac and Metropolis AI applications. Virtual integration reduces the enormous cost of physical change orders during construction and produces runtimes that are deployed to line computers and jets for production. The process creates a closed loop where robots running Isaac build NVIDIA AI supercomputers that are then used to train Foxconn robots, with the digital twin serving as the source of truth for equipment layout verification.

Cadence Cuts Chip Verification From Weeks to Hours With AI Engineers and NVIDIA OpenShell

Speakers pending

Chip verification faces extreme scale and timing constraints—trillions of transistors, 3D circuits and picosecond synchronization—so physical prototypes are impractical and RTL must be verified in simulation. Cadence and Nvidia built an AI-driven verification stack: Codex orchestrates Cadence ChipStack, which launches an RTL verification loop powered by Neumitron and secured by Nvidia OpenShell, calling specialist sub-agents for RTL generation, testbench creation, regression and debug. The agents run hundreds of simulations with Cadence Xcelium and perform formal verification with Jasper to find design flaws and enable fixes. Reported outcome: what once took weeks now takes hours, with verification cycles over 40× faster.

Digital Humans Transform Industries

Speakers pending

The talk frames the problem of making computers interact with people in humanlike, low-latency ways across industries (customer service, healthcare, advertising, design). The presented approach is a stacked set of AI capabilities — multilingual speech recognition and synthesis, LLMs for conversational intent, generative animation tied to a 3D face mesh, and neural rendering (real-time path-traced subsurface scattering) — packaged as deployable microservices (NIMs) in Nvidia Ace and served on a global accelerated network. Evidence includes application examples: phone-scan-based AI interior design that returns multiple photorealistic options, a conversational healthcare agent safely answering a drug-allergy question, and EMA, a Japanese virtual model/brand ambassador. Outcome: a production-oriented microservice architecture that enables real-time digital humans with regional low-latency delivery (100+ regions), though it requires orchestration across multiple specialized models and heavy real-time rendering compute.

Digital Twins Power Taiwan's Industrial AI Transformation

Speakers pending

Taiwanese manufacturers are building digital twins on NVIDIA Omniverse to move manufacturing design, validation, and operations into software-defined workflows. Workflows include converting 2D CAD into full 3D fab layouts (TSMC/MEAI), simulating multi-floor piping, power and cooling for data centers, and virtual commissioning of production lines and solder-paste dispensing. Companies use the twins as robot gyms to develop, train, and test manipulators, AMRs, humanoids, and vision agents, and connect twins to IoT for realtime interactive dashboards; reported outcomes include months saved in design and millions saved by reduced downtime and defects. The talk focuses on applied simulation and integration examples rather than quantitative metrics or formal guardrails.

During last night's dinner, Jensen took a quick break for a very important TVBS shoutout.

Speakers pending

Audience concern: whether AI is a short-lived bubble. Jensen counters that the industry has decades of runway because intelligence (and automating cognitive tasks) is fundamentally useful. He cites expanding compute capacity driven by partners (TSMC, SK Hynix, Amkor) as part of enabling that automation. Outcome: sustained multi‑year growth opportunity and a clear personal recommendation to engage and upskill rather than withdraw.

Early Preview of NVIDIA RTX Spark at Computex

Speakers pending

NVIDIA announced the RTX Spark super chip to enable local, high-capacity AI agents, advanced content creation, and high-performance gaming on slim Windows laptops and compact desktops. The system combines an NVIDIA Blackwell RTX GPU, an NVIDIA Grace CPU, up to 128 GB of unified memory, and the full NVIDIA AI/RTX software stack to run large models quickly and keep processing local for privacy. Demonstrations showed an agent (Open Claw) controlling ComfyUI to generate images and video from a sketch, real-time creative previews via DLSS 4.5 Ray Reconstruction in Blender, and games running at 1600p with ray tracing and DLSS 4.5. New laptops and compact desktops with RTX Spark will begin shipping this fall; the talk emphasizes local-model capabilities, large-memory workloads, and multi-API gaming support but provides demo examples rather than formal benchmarks.

Earth-2 Goes Down to Street Level

Speakers pending

Weather forecasting faces large path uncertainties for storms and city-scale impacts. Nvidia Earth-2 is a digital twin that fuses generative AI, physics simulations, and computer graphics; its Cordi model was trained on WRF numerical simulations to produce higher-resolution weather fields. Cordi upsamples from ~25 km to ~2 km (12× higher resolution) and is reported to run ~1,000× faster and ~3,000× more energy-efficiently than traditional physical simulators; Taiwan's Central Weather Administration is already using the models for more precise typhoon landfall forecasts. The stated next frontier is hyper-local forecasts down to tens of meters using high-resolution physics models (PALM) to capture urban airflow and street-level effects such as downwash, indicating ongoing work and current limitations at that scale.

Foxconn Builds Robotics for Healthcare With NVIDIA Physical AI

Speakers pending

Hospitals face clinician shortages and rising patient demand, with Taiwan's nurse shortage highlighted as urgent. Foxcon is building physical AI for digital hospitals using NVIDIA's stack: DGX for model training, Isaac for healthcare robotics development and simulation, and Jetson Orurin with NVIDIA holoskin for edge execution. Example Neurobot — powered by the Fox Brain LLM and trained virtually in a digital twin ward — navigates, delivers specimens and medications, and interacts with staff and patients, freeing nurses for clinical work. Early real-world deployment is reported as positive, but the source provides no quantitative metrics or detailed failure-mode analysis.

GenAI Powered by NVIDIA Grace Hopper | Supercharges 3D Video Calling and 5G Communications

Speakers pending

Problem: 2D video limits depth, presence, and natural interaction in remote communications, and specialized hardware restricts access. Approach: NVIDIA Aerial streams standard 2D camera captures from phones/laptops/webcams over 5G to cloud services; NVIDIA Maxine-3D, running on the NVIDIA Grace Hopper Superchip, converts those 2D streams into 3D visualizations and avatars. Evidence: The demo elements include low-latency 5G streaming, 2D-to-3D conversion, dynamic virtual camera angles, enhanced eye contact, text-driven avatar stylization, and multilingual avatar speaking. Outcome/limitation: The system promises immersive 3D conferencing on any device without specialized client hardware or software, while depending on cloud GPU compute (Grace Hopper) and low-latency 5G connectivity.

Highlights from NVIDIA Keynote and GTC Taipei 2025

Speakers pending

NVIDIA framed the problem as building full‑stack infrastructure to power next‑generation "AI factories" and agentic AI. Their approach combines new interconnect and system designs (NVLink Fusion, Blackwell family, upcoming GB300), enterprise servers (RTX Pro Servers with CX8), and application blueprints (Isaac Group Dreams and Omniverse) to enable semi‑custom racks, digital twins and robotics. Evidence from the keynote: NVLink Fusion announced, Blackwell in production, GB300 scheduled for Q3, RTX Pro Servers launched, and a Foxconn/Taiwan government AI factory supercomputer partnership. Outcome: a platform and ecosystem intended to accelerate deployment of enterprise AI and physical AI; limitation: some components (GB300, broader factory deployments) are forthcoming rather than immediately available.

How Robot Brains Dream and Explore Unseen Worlds

Speakers pending

Industrial robotics face a shortage of large-scale real and synthetic data for training because human demonstrations do not scale. Groot Dreams is a blueprint built on the NVIDIA Cosmos physical-AI foundation model that amplifies data by fine-tuning on teleoperated human demonstrations and then synthesizing many novel future-world “dream” videos from a prompt (image + new instructions). Cosmos internally reasons about and filters generated dreams; Groot Dreams converts the selected 2D dream videos into 3D action trajectories that are used to train robot models. The result is the ability to teach a robot a wide variety of new actions with minimal manual captures, though the pipeline requires converting pixel dreams into actionable trajectories and relies on quality selection of generated dreams.

Introducing Generative Physical AI

Speakers pending

The talk frames a transition to robotics driven by "generative physical AI": models that understand instructions and autonomously perform complex real-world tasks. The approach couples multimodal LLM capabilities (perception, instruction understanding, planning) with reinforcement learning that uses physics feedback inside high-fidelity simulation "robot gyms." NVIDIA Omniverse is presented as the operating system for these simulations, and a three-tier compute stack (AI supercomputers for training, Jetson Orin/Thor for runtime, Omniverse for simulation) closes the train→simulate→deploy loop; simulation scale (millions of trial-and-error acts) is used to refine behaviors and reduce the sim-to-real gap.

Introducing NVIDIA Cosmos 3: The Open Model That Thinks, Generates, and Acts

Speakers pending

Physical AI faces an unbounded, unpredictable real world and real-world data cannot scale; the talk reframes this by treating compute as data. NVIDIA Cosmos is presented as an open omnimodel for physical AI built on a new mixture-of-transformers architecture: an autoregressive transformer that reasons, plans, and instructs, and a diffusion transformer that generates what comes next from pixels, action, sound, and language. Post-training across embodiments converts Cosmos into an action-conditioned world model that can describe scenes, generate physics-accurate synthetic video, close the loop for policy training and evaluation, and produce actions for robots; the approach trades data collection for compute-generated synthetic data and computer-generated teachers.

Meet the Featured Startups at GTC Taipei 2025

Speakers pending

At GTC Taipei 2025 NVIDIA highlighted Inception startups applying GPU-accelerated stacks and simulation to domain-specific problems across healthcare, manufacturing, transportation, robotics, and sustainability. Examples include a drug-discovery pipeline that uses RAPIDS, CUDA, BioNeMo and TensorRT to screen millions of plant-based compounds and claims an 80% reduction in traditional timelines (years to months); sensor-fusion and simulation-driven 3D tracking for rail collision warning; AI-driven textile digital twins; energy-optimized humanoid robots on Jetson Orin NX; and a standardized, traceable recycling ecosystem with automated sorting. Evidence in the talk centers on applied software/hardware stacks, data scale, simulation for edge-case training, and targeted hardware/actuator choices; outcomes reported are speedups, improved reliability, lower energy and cost, and better material recovery. Limitations are implicit: claims focus on platform components and outcomes without published evaluation protocols, quantitative failure-mode analysis, or detailed deployment metrics beyond cited speedups and functional capabilities.

NVIDIA AI & Omniverse: Pegatron Digitalizes AI Smart Factory w/ Isaac Sim, Metropolis, and Omniverse

Speakers pending

Manufacturers must become software-defined to produce high-quality products faster and cheaper. Pegatron builds a cloud-native Omniverse digital twin that unifies disparate 3D and CAD datasets to give planners and suppliers a real-time view, and uses that twin as a training ground and data factory for perception AIs. NVIDIA stack components—Isaac Sim for robot simulation, Omniverse Replicator for synthetic PCB-A defect data, Metropolis for AOI workflows, TAO for training, and DeepStream for inference—form a pipeline that produced AOI at 99.8% accuracy with 4× throughput. Outcomes include faster factory bring-up, fewer change orders, continuous operational optimization, higher line throughput, and lower costs.

NVIDIA and Data Center Partners Review Highlights at COMPUTEX 2025

Speakers pending

Computex 2025 framed the modern computer as a unified, AI-driven data center — an "AI factory" — solving the problem of scaling inference and unstructured-data throughput across systems. Nvidia's approach centers on the Grace Blackwell platform and the new Grace Flagwell Ultra in full production, claiming 1.5x inference performance through dual-chip connectivity, TSMC POWSL packaging, and an NV link spine. The NV link spine provides 130 TB/s all-to-all GPU bandwidth across 72 Flagwell Ultra GPUs to present a single large virtual GPU, and Nvidia paired a purpose-built NVIDIA AI data platform plus AIQ to semantically query unstructured data. Availability signals focus on production deployment (MGX components from 60+ suppliers, RTX Pro servers, liquid cooling and high-power architectures), though transcript details on evaluation methodology and operational guardrails are limited and should be validated before integration.

NVIDIA at Computex - G-SYNC For Notebooks, New Monitors, New Features

Speakers pending

At Computex NVIDIA announced a new G-SYNC product wave: seven partner monitors (including IPS-based 34" ultrawide gaming displays) and G-SYNC support coming to notebooks. They also introduced two feature changes to the G-SYNC stack: a Window mode to enable G-SYNC for windowed games, and an explicit vsync on / vsync off option to let players trade off latency vs synchronisation. The net outcome is broader hardware support (desktop monitors + notebooks) and configurable latency behavior at very high frame rates; limitations implied are that adoption depends on partner hardware and correct integration into notebook/display stacks.

NVIDIA at Computex 2015 - Overview

Speakers pending

NVIDIA showcased solutions aimed at higher-fidelity VR and realtime visual effects: they launched the Maxwell-based GeForce GTX 980 Ti and highlighted DirectX 12 features such as sparse textures. The company demonstrated these capabilities through hands-on VR experiences and demos that included realtime fluid simulations. They also introduced new G-SYNC modules and G-SYNC monitors and announced G-SYNC support for notebooks. The demonstrations indicate improved VR/display integration, but the transcript contains no quantitative performance or availability details.

NVIDIA at COMPUTEX 2023 Highlights

Speakers pending

Computex 2023 presented a full-stack effort to enable large generative-AI models and industrial digitalization: NVIDIA announced the DGX GH200 (connecting 256 Grace Hopper Superchips into a massive “1X of flop GPU”), an MGX open modular server design, and the Spectrum X ethernet platform aimed at hyperscale AI. Complementing hardware, NVIDIA highlighted Omniverse and generative AI services for factory digitization, a WPP Omniverse ad content engine, and ACE for games to animate virtual characters. Evidence came from Jensen’s 90-minute keynote, partner demos, SoftBank’s planned Grace Hopper–powered data center, and a MediaTek in-vehicle AI partnership; the transcript does not provide performance benchmarks, cost, or deployment timelines. The announcements create an integrated path from chips to datacenter networking to application-level services, but operational details and metrics remain unspecified in the source.

NVIDIA at COMPUTEX 2024 Highlights

Speakers pending

NVIDIA framed COMPUTEX 2024 around preparing for the next industrial digitization wave by delivering full‑stack platforms that span consumer AI PCs to industrial robotics and digital twins. The company unveiled a roadmap including the Blackwell Ultra GPU and the Reuben platform, which will combine new GPUs, an Arm‑based CPU called Vera, and advanced networking, delivered on a one‑year cadence. Evidence presented at the keynote and show floor included >200 RTX AI laptops, >500 AI apps/games, over 60 partner demonstrations, and an NVIDIA AI Summit with 30 sessions and 1,000+ attendees. Outcome: an explicit hardware‑software platform roadmap and ecosystem activation; limitation: major platform innovations are staged on a multi‑quarter (one‑year) release rhythm rather than immediate availability.

NVIDIA at COMPUTEX 2024 Partner Highlights

Speakers pending

Computex 2024 emphasized building "AI factories" and data centers using NVIDIA's Blackwell architecture, new GB200 MV2 platform(s), MGX modular reference architecture, Spectrum X Ethernet, and inference microservices. The GB200 MV2 is presented as a mainstream LLM and data-processing accelerator targeted at inference, retrieval-augmented generation, and analytics workloads. MGX is positioned for scale-out servers (over 100 Blackwell systems / designs reported) while NVIDIA Inference Microservices (NIMS) provides pre-trained models as optimized containers for cloud, data center, or workstation deployment (adopted by 150+ partners). The announcements focus on system-level hardware + networking + software stacks, but the transcript contains no specific performance numbers, benchmarks, or operational guardrails.

NVIDIA Blackwell: The Journey From Die to Data Center

Speakers pending

The talk traces Blackwell from wafer fabrication to rack-scale AI systems: hundreds of lithography steps build roughly 200 billion transistors on a 12in wafer, which are scribed into dies, tested, packaged, and stress-tested. A chip-on-wafer-on-substrate flow bonds 32 Blackwell dies and 128 HBM stacks onto a custom silicon interposer, producing the B200 Super Chip after molding and curing. System integration uses automated PCB assembly, custom liquid cooling, ConnectX7 SuperNICs and BlueField‑3 DPUs, and Nvidia's MVLink fabric (14.4 TB/s per switch tray, 130 TB/s backplane) to present up to 72 Blackwells (144 dies) as a single giant virtual GPU. The result is a 1.8-ton, rack-scale AI supercomputer assembled from 1.2 million components and validated with high-temperature stress tests, with reliability and thermal management emphasized as production guardrails.

NVIDIA CEO Jensen Huang at Meet-a-Claw in Taipei

Speakers pending

Engineers face the challenge of learning how to operate and compose open-source large language model (LLM) agents. The approach described is Open Claw, an open-source project whose harness functions as an "operating system" around a large language model to structure usage. Evidence in the transcript: the harness is explicitly called useful and good for learning, and an event demonstrates what open-source agents can do so participants can build their own. Outcome: attendees can study the harness, observe many different agent capabilities, and then create custom agents; the transcript does not provide technical metrics, performance data, or production guardrails.

NVIDIA CEO Jensen Huang at Taipei’s Raohe St. Night Market

Speakers pending

Problem: The supplied transcript contains only a single colloquial utterance, preventing extraction of any technical content. System/approach: I attempted a transcript-grounded scan for architecture, inputs/outputs, evaluation, metrics, guardrails, trade-offs, failure modes, and deployment details. Evidence: The only line in the transcript is "Gotta go." and metadata names the event and speaker (NVIDIA CEO Jensen Huang at Taipei’s Raohe St. Night Market). Outcome/limitation: No engineering-usable information can be derived from this input; a fuller transcript or additional materials are required for a technical briefing.

NVIDIA CEO Jensen Huang Keynote at COMPUTEX 2024

Speakers pending

CPU single‑thread scaling has slowed while total compute demand is exploding, creating "computation inflation". NVIDIA's approach is accelerated computing: GPU + CUDA domain libraries, integrated software stacks (cuDNN, TensorRT, Triton) and an end‑to‑end hardware platform (Blackwell GPUs, Grace CPU, NVLink, NIC/switch, DGX/MGX systems) to reduce training/inference cost and energy. Evidence presented includes library accelerations (cuDF accelerating pandas), a digital twin/weather demo (Earth‑2), packaged inference containers (NIMs) and production Blackwell hardware with NVLink switching and decompression engines. Outcome: very large speedups and energy savings but requires explicit software refactoring, system integration, and new operational practices (tail latency, RAS, secure AI).

NVIDIA CEO Jensen Huang Keynote at COMPUTEX 2025

Speakers pending

Computing for modern AI workloads requires rethinking hardware, networks, software stacks, and the operational metrics used to evaluate systems. NVIDIA's response in the keynote is an end-to-end approach: new chip generations and rack-scale nodes (Grace Blackwell GB200/GB300), a high‑bandwidth MVLink fabric and MVLink Fusion for semi‑custom integration, server and workstation SKUs (RTX Pro, DGX Spark, DGX Station), plus software stacks (CUDA X libraries, Nemo/IQ, Omniverse, Isaac) and model blueprints. The talk gives concrete evidence (GB300: 1.5x inference, higher HBM and networking; MVLink spines with hundreds of TB/s all‑to‑all; IQ and Nemo retrieval speed/quality claims) and describes outcomes and limits: dramatic inference-time scaling and agent/robot training capability but very high power, liquid cooling, packaging and interconnect complexity and single‑rack scaling constraints.

NVIDIA Data Center Partners Recap | GTC Taipei 2026 Recap

Speakers pending

The talk frames a fundamental shift from generative models to agentic AI that plans, uses tools, and acts autonomously at extremely high frequency. NVIDIA responded with Vera Rubin (a pod-scale supercomputer end-to-end optimized from silicon to data‑center floor) and a Vera CPU designed for agentic workflows, claiming 10× agent throughput vs Blackwell and 1.8× CPU speed on agentic workloads. For operators, DSX (NVIDIA's data‑center optimization platform) addresses the hard power constraint by increasing safe GPU density and improving throughput-per-watt; evidence cited includes a tripling of GitHub commits in early 2026 and stated economic uplift metrics, while limits remain the fixed power budgets and the need for large-scale manufacturing co-design to deliver the systems.

NVIDIA DSX Powers Gigawatt‑Scale AI Factories at Maximum Efficiency

Speakers pending

Large, gigawatt‑scale AI factories require end‑to‑end co‑design of chips, racks, networks, power, cooling and grid interactions because compute directly drives revenue. Nvidia DSX is presented as a blueprint composed of a digital‑twin design stage (DSX Sim Omniverse), an operations/control plane (DSX OS), power/thermal optimization (DSX Max LPS with AI agents and in‑rack smoothing), and grid‑aware demand shaping (DSX Flex). The transcript reports simulation and validation before deployment, operating provisioning and continuous remediation after power‑on, hot liquid cooling at 45°C to save water/energy, and dynamic power allocation to recover stranded watts — all intended to increase GPU density and reduce cost per token. Limitation: claims of efficiency are architectural and operational; concrete numeric improvements beyond current overprovisioning (up to 40%) are not fully quantified in the source.

NVIDIA Foundational Technology Montage I GTC Taipei and COMPUTEX 2026 Edition

Speakers pending

NVIDIA built CUDA 20 years ago as a single architecture for accelerated computing to address the need for general-purpose hardware acceleration. That architecture is backed by a broad ecosystem of roughly a thousand CUDA‑X libraries that target domain-specific problems and are described as tools for agents. The talk lists concrete library examples—cuLitho, cuOpt, cuDSS, AIQ, Aerial, Warp, Parabricks—showing coverage across lithography, optimization, sparse solvers, document research, AI RAN, differentiable physics, and genomics. The stated foundation of this ecosystem is algorithms, which are presented as the core enabling element for breakthroughs across science and engineering.

NVIDIA GTC Live 2026 Pregame Show in Taipei | Full Replay

Speakers pending

NVIDIA GTC Live at COMPUTEX Taipei 2025 Pregame

Speakers pending

The transcript describes an industry response to rapidly escalating AI compute and system-integration demands: major suppliers (Nvidia, TSMC, Dell, ASUS, Quanta, Foxconn, Trend Micro, Delta, Techman) converge on a stack that pairs advanced-process silicon and packaging with modular server architectures, on-device vision/AI in robotics, and physics-accurate digital twins (Omniverse/Isaac). Evidence includes long-term foundry collaboration (TSMC since 1997, > six million 12-inch wafers delivered to Nvidia as stated), Dell/ASUS/QCT production of GB200/HGX-class systems (Dell first to ship GB200 with NVL72), Trend Micro training a cybersecurity LM (Cybertron) with NeMo, and factory productivity gains from Omniverse-driven digital twins (labor cost -50%, productivity +30%, changeover time /2, unit-per-hour +73%, cycle time -25%). Outcome: this integrated approach shortens iteration and deployment loops and enables new physical-AI applications, but remains constrained by manufacturing yield for very large dies, advanced-node and packaging complexity, power/thermal engineering, and the need for higher-fidelity sim-to-real (dynamics, materials) to reduce trial-and-error.

NVIDIA GTC Live pregame show at COMPUTEX 2026

Speakers pending

Speakers framed the NVIDIA GTC Live pregame at COMPUTEX 2026 as a central coordination point for a rapidly accelerating AI ecosystem: the problem is the fast-moving, global AI surge; the approach is to gather industry leaders and spotlight Taiwan’s role at COMPUTEX; evidence cited includes A-list leadership, a surprise guest, and the expectation of participation by ten trillion-dollar companies; the outcome is strong momentum and excitement, with the limitation that the industry has still "only scratched the surface."

NVIDIA GTC Taipei 2026 Keynote | Full Replay

Speakers pending

Agentic AI changes the computing problem: agents require low-latency, memory‑heavy, tool‑driven loops rather than traditional application bundles. NVIDIA's approach is extreme co‑design across chips, racks, networking, storage, and software — exemplified by the Vera Rubin multi‑rack pod (NVLink‑72, Vera CPUs, BlueField DPUs, Spectrum‑X) plus DSX/Omniverse for digital twin design, and a matched software stack (Open Shell runtime, CUDA‑X libraries with agent skills, NeMoTron and Cosmos models). Jensen Huang presented production evidence: Vera Rubin is in full production, NeMoTron 3 Ultra is open and reported as 5× faster and 30% cheaper than leading open models, and Vera CPU shows multi× speedups on real workloads; however, physical AI data remains the hardest limiting input and the overall stack demands careful orchestration and TCO planning (tokens-per-watt, time‑to‑first‑token, reliability).

NVIDIA HGX H100 | The most powerful end-to-end AI supercomputing platform

Speakers pending

The provided transcript contains only brief closing material: two occurrences of the phrase "thank you" and two music cues. There is no technical content, data, architecture, or claims about NVIDIA HGX H100 in the source. Because the source lacks substantive material, no engineering details, evaluations, metrics, or deployment guidance can be extracted. To produce a technical briefing or any actionable recommendations, a full technical transcript or presentation content is required.

NVIDIA Isaac AMR | End-to-End Autonomy Platform for Next Generation AMRs

Speakers pending

Manufacturing and warehousing operations are moving from manual forklifts and guided vehicles to fully autonomous mobile robots to raise productivity and improve worker safety. NVIDIA Isaac AMR is an integrated end-to-end platform built around NOVA Orin (a sensor suite plus on-board compute) and a cloud mapping service that produces detailed 3D voxel maps which can be sliced into 2D maps for different robots. On-robot Lidar or cost‑effective cameras provide navigation, Isaac Mission Control (using cuOpt) optimizes routes, and Isaac Sim + NVIDIA Omniverse let teams train and fully validate autonomy in realistic digital twins before real-world deployment. The result is faster, lower-cost migration to full autonomy with a simulation-based validation guardrail and an explicit trade-off between higher-performance Lidar and lower-cost camera solutions.

NVIDIA Keynote at COMPUTEX 2023

Speakers pending

Computex keynote frames two simultaneous industry transitions: accelerated computing (full‑stack, domain specific, data‑center scale) and generative AI. NVIDIA presents system‑level solutions from silicon (H100, Grace Hopper/GH200), interconnect (NVLink, Spectrum‑4 + BlueField‑3), software (TensorRT, NVIDIA AI Enterprise, NVIDIA AI Foundation) and application platforms (Omniverse, ACE, Isaac) to deploy LLMs, real‑time ray tracing, 3D video and factory/robot simulation. Evidence includes production H100/Hopper supply, GH200 single‑module memory coherency (~600 GB aggregated), Spectrum‑4 switch specs and multi‑pod GH200 exaFLOPS design; outcomes are higher per‑Watt throughput, larger single‑GPU memory footprints to reduce sharding, and cloud/enterprise deployment paths — limitations remain in full stack integration complexity, power and network requirements.

NVIDIA Keynote Song at COMPUTEX 2024

Speakers pending

The speaker, who says they are new at computers, presents a short, highly repetitive song-like message that repeatedly names two phrases: "Infinity On" and "Digital Twins." The approach is performative and promotional — calling the audience to sing along, addressing "computers," and claiming to "show you all our best." Evidence in the transcript is purely lyrical and branded; there are no technical specifications, metrics, or implementation details, so the content is non-actionable for engineering work.

NVIDIA NVLink Fusion: Building Semi-Custom AI Infrastructure

Speakers pending

NVIDIA Omniverse Foundational Technology Montage I COMPUTEX 2025 Edition

Speakers pending

NVIDIA Omniverse Foundational Technology Montage | COMPUTEX 2024 Edition

Speakers pending

NVIDIA RTX PRO 6000 Blackwell Server Edition

Speakers pending

Requested briefing source contains only repeated musical cues and the single-word vocalizations "Heat" (including "Heat up here"). The observable system is an audio excerpt with no technical narration or data. Evidence: the transcript shows five "[Music]" markers and four occurrences of the word "Heat" and no product details. Outcome/limitation: no architecture, inputs/outputs, metrics, guardrails, or deployment information can be extracted from this source.

NVIDIA Spectrum-X Platform | World’s first Ethernet fabric built for AI

Speakers pending

AI training generates very large network flows that cause congestion, higher latency, and bandwidth unfairness when run on traditional cloud Ethernet, reducing GPU utilization. NVIDIA's Spectrum-X is an end-to-end Ethernet fabric purpose-built for AI, based on the Spectrum-4 switch/Spectrum-4 ASIC working tightly with the BlueField-3 DPU and enhanced RoCE features. In real-world examples Spectrum-X shows up to 1.7× improvement in per-GPU performance, power efficiency, training time, and TCO, and 2.5× higher effective bandwidth for NCCL AllReduce, yielding lower jitter and shorter tail latency. The platform relies on full‑stack optimizations (software, libraries, OS) and RoCE controls (adaptive routing, performance isolation, congestion control) to deliver predictable, higher effective bandwidth across AI clusters.

NVIDIA Tech Tour at Computex 2015

Speakers pending

NVIDIA Vera Rubin Platform Ramping into Full Production | Built for the Era of Agents

Speakers pending

Agentic AI requires systems that support observation, reasoning, planning, tool use, large working and long‑term memory, and dynamic sub‑agent spawning; conventional single‑chip or single‑rack designs are insufficient. NVIDIA Vera Rubin is a multi‑rack pod‑scale supercomputer built for this agentic era: heterogeneous, liquid‑cooled racks that combine NVL72 superchips for high‑throughput token generation, Groq 3 LPX for ultra‑low latency, BlueField‑4 DPUs for memory/storage/security acceleration, and Spectrum‑X photonic switching. The build is evidenced by the production rollout: seven new 3 nm chips, CoWoS packaging, HBM4 from multiple vendors, a 6‑trillion‑transistor compute board, MGX third‑generation racks with 1.3M components, and five connected rack‑scale systems now in full production. Outcome: a purpose‑built pod‑scale platform that trades extreme power, cooling, and supply‑chain complexity for agentic capabilities and deployment resilience (hot‑swappable NVLink switches, cableless interconnects), with the practical limitation of very large manufacturing and operations scale.

NVIDIA—Our Story, Built on Taiwan

Speakers pending

Official Keynote Closing Video | GTC Taipei 2026

Speakers pending

The keynote frames a concrete problem: agent-based AI is moving from novelty to production but demands much more compute, memory correctness, and safety. The presented approach is an integrated stack (Vera/Rubin token optimization, BlueField memory, a faster agent-focused CPU, NVLink Fusion for ASIC interconnect, Nemo Tron Ultra, Nemo Clock and Open Shell guardrails, ESX power optimizations, and GPU/CPU partitioning) that targets inference throughput, memory fidelity, sandboxing, and power efficiency. Evidence in the talk consists of headline performance claims ("cheapest tokens," 10x faster inference, 5x faster work, 50% faster CPU) and feature claims (memory correctness, guardrails, sandboxing, automated code migration), but the transcript does not include benchmark methods or detailed validation. Outcome: the stack is presented as production-ready for large-scale agent workloads and global AI clouds, while the talk does not provide the underlying evaluation details or failure-mode analysis needed for engineering validation.

Overclocking at Computex 2015

Speakers pending

This transcript is an event announcement with no technical substance: the speaker says "I'll explain later" and notes that "Master Kindergarten" has a demo today. Contestants from different countries have been invited to attend at the venue (the invitation is repeated). Evidence is the repeated invitation lines; outcome/limitation is that no system, data, metrics, or evaluation details are provided and a follow-up explanation is required.

Shaping the Future of Work With Agentic AI

Speakers pending

Aentic AI demonstrates a retrieval-augmented workflow that turns sales planning from days into minutes. VAS built a sales research AI agent in three days using the NVIDIA IQ blueprint and an accelerated AI data platform. The pipeline uses Nemo Retriever to continuously extract, embed, and index data for fast semantic search, then an agent drafts an outline, queries CRM/multimodal knowledge bases and internal tools, and Llama Nematron converts the outline into a step-by-step sales plan. Outcome: rapid, production-oriented agent creation and per-employee specialization enabled by Vast's accelerated AI data platform.

Talk to Your Supply Chain Data Using NIM

Speakers pending

Nvidia faces extremely large, interdependent supply-chain decision problems (a single Hopper DGX can contain ~40,000 parts from 150+ suppliers across many factories) that require fast replanning when disruptions occur. Their business operations team built an AI planner assembled from NIMs (Nvidia Inference Microservices): an LLM Nim for intent, a Nemo retriever RAG Nim for proprietary data access, and a Nvidia coop Nim for GPU-accelerated combinatorial optimization. Nvidia coop (a GPU-accelerated operations-research library with 23 world-record benchmarks) computes optimal routes in seconds and dynamically recalculates when constraints or requirements change. The result is real-time, natural-language driven scenario analysis and product reallocation that reduces replanning from overnight to seconds, enabling faster operational response to supply changes.

Techman AI Robot for Quality Inspection | Developed in Isaac Sim and powered by Omniverse

Speakers pending

Automated Optical Inspection (AOI) often needs access to product areas unreachable by fixed cameras, but programming articulated robot arms to capture hundreds of images is time-consuming and combinatorially complex. Techman built a digital twin in NVIDIA Omniverse using Isaac Sim to simulate and optimize robot poses and inspection sequences, and used Omniverse Replicator to synthesize training data for ML models. Simulation-driven optimization cut programming time by over 70% and reduced per-inspection cycle time by 20%; synthetic data dramatically improved model performance. The solution is deployed at the edge to run dozens of low-latency AI models during inspection, enabling faster, higher-quality AOI on production lines.

The AI Factory: How NVIDIA Builds the Engine for the New Industrial Revolution

Speakers pending

The supplied transcript is a short, fragmentary stage intro with heavy vocal repetitions and a music cue. The speaker utters an ambiguous performance-style phrase that appears to reference improved speed and efficiency (“made faster than a speed a light efficient”), but provides no technical context, metrics, or system description. Evidence in the snippet is limited to raw audio-text artifacts and a few unclear phrases, so no concrete engineering conclusions can be drawn from this source alone. The primary limitation is that the fragmentary, noisy transcript prevents reconstruction of any substantive technical content.

The Taipei 101 Tribute to NVIDIA GTC

Speakers pending

The supplied transcript contains a single phrase: "Europe is coming." There are no technical details, speaker attributions, architecture descriptions, inputs/outputs, evaluation methods, metrics, guardrails, deployment instructions, or evidence to analyze. Because the source contains only that short statement, no engineering conclusions about systems, performance, or trade-offs can be drawn. Outcome: further analysis requires the complete transcript or technical materials from the Taipei 101 tribute talk.

This is NVIDIA Alpamayo Thinking Out Loud

Speakers pending

The transcript lists a set of low-level driving decisions triggered by immediate roadway events: stationary lead vehicle blocking the lane, stop sign at an intersection, pedestrian occupying our lane, a vehicle cutting in from the left, and a stopped vehicle blocking on the right. The system response repertoire is limited to lateral nudges, slowing, stopping, and yielding mapped directly to these triggers. Evidence in the transcript shows clear input-to-action mappings (e.g., stationary lead vehicle -> nudge left; pedestrian in lane -> stop to yield). Outcome: these rules resolve short-range conflicts by lateral and longitudinal adjustments but do not describe sensing, timing, or higher-level planning context.

Tokens Turn Data Into Knowledge | Official Keynote Intro | GTC Taipei at COMPUTEX 2026

Speakers pending

The talk frames the core problem as large, complex data sets that hide actionable patterns. The proposed system-level approach is a 'new kind of factory' that generates tokens—the building blocks of AI—to turn raw data into knowledge, reasoning, and action. Evidence in the talk is descriptive: tokens reveal previously unseen patterns, enable digital mirrors of cities, help robots learn from humans, and support health and safety outcomes. The outcome is an industrialized, scalable path to intelligence, with the caveat that this is presented as the beginning of a large effort (here starting in Taipei) rather than a completed deployment.

Tokens, the Building Blocks of AI Breakthroughs | COMPUTEX 2025 | Official Keynote Intro

Speakers pending

WPP & NVIDIA Partner to Build Generative AI-Enabled Content Engine on Omniverse Cloud

Speakers pending

NVIDIA and WPP are building a generative AI–enabled content engine on NVIDIA AI and Omniverse Cloud to speed production of product-accurate, brand-compliant digital advertising assets. The pipeline begins by constructing a physically accurate digital twin of a product from industry-standard design data in Omniverse Cloud, then WPP artists compose virtual sets augmented by generative AI tools (e.g., Adobe, Getty Images) trained on fully licensed data via NVIDIA Picasso. The resulting outputs are photorealistic visual content and e-commerce experiences delivered at scale; the source does not provide numeric metrics, failure modes, or operational details for evaluation or iteration loops.