FAQ topic09 / 12

Edge AI and NVIDIA Jetson FAQs

Edge AI runs inference next to the cameras and sensors instead of in the cloud. These FAQs cover NVIDIA Jetson modules and software, GPUs, NPUs and other accelerators, TOPS and TFLOPS, and how AI is applied in industry.

Search this topic

86 answers Search all FAQs

Topics 12

AGRICULTURE · EDGE AI

How Is AI Used in Agriculture?

AI in agriculture uses machine learning on images and sensor data to make farming decisions faster and more precisely. Common uses include detecting and spraying weeds, steering autonomous tractors, monitoring crop and livestock health, forecasting yields, and sorting and grading produce, often running on rugged edge computers mounted on the machines themselves.

6 MIN

EDGE AI · MANUFACTURING

How Is AI Used in Manufacturing?

AI in manufacturing means using machine learning models on factory data to inspect products, predict machine failures, guide robots, improve processes and keep people safe. Most of it runs on edge computers beside the line, where cameras and sensors produce the data and decisions are needed in milliseconds. Planning and forecasting models usually run in the cloud.

6 MIN

EDGE AI · COMPUTING

What Is Edge AI?

Edge AI means running trained artificial intelligence models on computers close to where data is created, such as beside cameras, machines or vehicles, rather than in the cloud. Decisions happen locally in milliseconds, less data crosses the network, sensitive data stays on site, and the system keeps working if the connection drops.

7 MIN

AI PERFORMANCE · METRICS

What Is a Teraflop (TFLOPS), and How Is It Different From TOPS?

A teraflop is one trillion floating-point operations per second, and TFLOPS counts them: a measure of how fast a processor handles decimal maths at a stated precision such as FP32. TOPS counts trillions of operations per second, usually integer maths at INT8. GPUs quote TFLOPS for graphics and general compute, while NPUs and AI accelerators quote TOPS for inference.

6 MIN

INDUSTRIAL AUTOMATION · IIOT

What Is Smart Manufacturing?

Smart manufacturing is the use of connected machines, sensors, software and AI to monitor production as it happens and adjust it automatically. Data from the shop floor flows to edge computers and business systems, which spot faults, predict failures, check quality and balance output against demand, so factories respond faster with less waste.

6 MIN

ROBOTICS · SOFTWARE

What Is ROS (Robot Operating System)?

ROS (Robot Operating System) is a free, open-source set of software libraries and tools for building robot applications. Despite the name, it is not an operating system: it runs on Linux and handles communication between the parts of a robot, such as sensors, navigation and motor control. ROS 2 is the current version.

6 MIN

ROBOTICS · EDGE AI

What Is an Autonomous Mobile Robot?

An autonomous mobile robot (AMR) is a robot that moves through a building or site on its own. It builds a map with lidar and cameras, works out where it is, plans a route and steers around people and obstacles, so it needs no tape, wires or fixed tracks. AMRs move goods, tow carts, clean, inspect and deliver.

6 MIN

PROCESSORS · SILICON

FPGA vs ASIC: What Is the Difference?

An FPGA (field-programmable gate array) is a chip whose logic can be reconfigured after manufacture, so one part can serve many jobs and be updated in the field. An ASIC (application-specific integrated circuit) is designed and fabricated for one job. ASICs are faster, smaller and more efficient at volume, while FPGAs avoid the high upfront cost and long lead time.

6 MIN

EDGE AI · NVIDIA SOFTWARE

What Is NVIDIA Holoscan?

NVIDIA Holoscan is NVIDIA’s sensor-processing platform for real-time AI on streaming data at the edge. It combines the open-source Holoscan SDK, which builds low-latency pipelines from capture to inference to display, with the Holoscan Sensor Bridge for bringing camera and sensor data over Ethernet straight into GPU memory, on Jetson, IGX and x86 GPU systems.

6 MIN

VISION · CAMERA INTERFACES

What Is MIPI CSI-2?

MIPI CSI-2 (Camera Serial Interface 2) is the standard high-speed link between a camera’s image sensor and a processor, defined by the MIPI Alliance. It carries image data over one to four differential lanes on a short flex cable, and it is the native camera input on NVIDIA Jetson, NXP i.MX, Raspberry Pi and most smartphone processors.

6 MIN

GPUS AND AI · NVIDIA

What Is NVIDIA Blackwell?

NVIDIA Blackwell is the GPU architecture NVIDIA announced in March 2024 to follow Hopper and Ada Lovelace. It powers B200 and GB200 data centre accelerators, RTX PRO Blackwell workstation and embedded GPUs, GeForce RTX 50 cards and the Jetson Thor edge AI modules, and adds fifth-generation Tensor Cores with 4-bit (FP4) AI.

6 MIN

OPERATING SYSTEMS · EMBEDDED LINUX

What Is the Yocto Project?

The Yocto Project is an open-source set of tools for building a custom Linux operating system for one embedded device. Instead of installing a ready-made distribution, you describe the hardware, software and settings you need in layers and recipes, and the BitBake build engine produces a small, repeatable image for that device.

6 MIN

MACHINE VISION · QUALITY CONTROL

What Is Automated Visual Inspection?

Automated visual inspection (AVI) uses cameras, controlled lighting and software to check products for defects without a human inspector. Each part is imaged on the line, the software compares it with what a good part should look like, and faulty items are rejected and logged. It is faster and more consistent than manual checks.

7 MIN

EDGE COMPUTING · SERVERS

What Is an Edge Server?

An edge server is a server installed close to where data is created or used, such as a factory, depot, shop or cell site, instead of in a central data centre or the cloud. It processes data locally for low latency and lower bandwidth. In web delivery, the term also means a CDN server that caches content near users.

7 MIN

DEFENCE · SIGNALS INTELLIGENCE

What Is SIGINT?

SIGINT, or signals intelligence, is intelligence gathered by intercepting and analysing signals. It covers communications intelligence (COMINT), taken from messages between people, and electronic intelligence (ELINT), taken from non-communication emissions such as radar. GCHQ in the UK and the NSA in the US lead it, alongside military teams on aircraft, ships and ground stations.

7 MIN

VISION · SOFTWARE

What Is OpenCV?

OpenCV (Open Source Computer Vision Library) is a free, open-source library of more than 2,500 algorithms for image processing and computer vision. Developers call it from C++ or Python to read camera feeds, clean up and measure images, find features and run trained neural networks, on anything from a laptop to an NVIDIA Jetson.

6 MIN

MARINE · AUTONOMOUS SYSTEMS

What Is an Unmanned Surface Vehicle (USV)?

An unmanned surface vehicle (USV) is a boat that operates on the water with no crew on board. It is either driven remotely from a control station or navigates autonomously using onboard sensors and computers. Also called uncrewed surface vessels or sea drones, USVs carry out survey, inspection, environmental monitoring, security patrol and naval missions.

7 MIN

EDGE AI · TRAINING DATA

What Is Synthetic Data Generation?

Synthetic data generation is the creation of artificial training data, such as rendered images, simulated sensor readings or generated text, that mimics real data. It fills gaps where real data is scarce, costly to label, private or unsafe to collect, and it comes with exact labels. Most vision and robotics teams combine it with real data.

6 MIN

EDGE AI · ROBOT LEARNING

What Is Reinforcement Learning in Robotics?

Reinforcement learning is a type of machine learning in which an agent learns by trial and error, taking actions and receiving rewards or penalties. In robotics, a control policy is usually trained this way in GPU-accelerated simulation, across thousands of virtual robots at once, then transferred to the real machine.

6 MIN

EDGE AI · ROBOTICS

What Is Physical AI?

Physical AI is artificial intelligence that senses the real world, reasons about it and acts in it through a machine such as a robot, vehicle or camera system. Its models are trained on GPU servers, tested in physics simulation and then run on a computer inside the machine, such as an NVIDIA Jetson module.

6 MIN

EDGE AI · ROBOT LEARNING

What Is Imitation Learning?

Imitation learning is a way of training a robot or AI agent by showing it how a task is done rather than programming it. An expert, usually a person, demonstrates the task, the system records what it sees and what the expert does, and a model learns to map each situation to the right action.

6 MIN

COMPUTING · SERVERS AND CLUSTERS

What Is High Performance Computing (HPC)?

High performance computing (HPC) uses many processors working in parallel, usually a cluster of servers with GPUs joined by a fast network, to solve problems too large or too slow for one computer. It runs simulation, modelling, data analysis and AI training, and its speed is measured in floating-point operations per second (FLOPS).

7 MIN

EDGE AI · ROBOTICS

What Is Embodied AI?

Embodied AI is artificial intelligence that has a body. It perceives its surroundings through sensors, acts through motors and learns from the results of its own actions, rather than only processing text or images. Robots, drones, autonomous vehicles and humanoids are all forms of embodied AI.

6 MIN

EDGE AI · FUNDAMENTALS

What Is Deep Learning?

Deep learning is a branch of machine learning that uses artificial neural networks with many layers to learn patterns directly from raw data such as images, sound and text. Each layer builds on the one before, from edges to whole objects or from letters to meaning. It drives computer vision, speech recognition and large language models.

7 MIN

EDGE AI · VISION

What Is Computer Vision?

Computer vision is the branch of AI that lets computers interpret images and video: naming what is in a frame, locating objects, outlining them pixel by pixel and following them over time. Modern systems use deep learning models such as CNNs and vision transformers, and many run on edge computers next to the cameras they read.

6 MIN

EDGE AI · DEPLOYMENT

What Is AI Inference?

AI inference is the stage where a trained model is put to work: it takes new data, such as a camera frame, a sensor reading or a text prompt, and produces an output such as a label, a detection, a forecast or generated text. It runs every time the model is used, so latency, cost and power matter most.

6 MIN

EDGE AI · ROBOT FOUNDATION MODELS

What Is a VLA Model?

A VLA (vision-language-action) model is an AI model that takes camera images and a plain-language instruction and outputs robot actions, such as joint movements or gripper commands. It starts from a vision-language model trained on web data and is then trained on robot recordings, so one model handles seeing, understanding and acting.

6 MIN

EDGE AI · GENERATIVE AI

What Is a Vision Language Model (VLM)?

A vision language model (VLM) is an AI model that takes images or video together with a text prompt and answers in text. It joins a vision encoder to a large language model, so you can ask about a scene, caption or summarise footage, read documents or search video in plain language instead of training a fixed-class detector.

6 MIN

EDGE AI · GENERATIVE AI

What Is a Multimodal LLM?

A multimodal LLM is a large language model that accepts more than one type of input, such as text with images, audio or video, and reasons across them in a single model. Some can also reply in speech. Vision language models are the most common kind; others add audio so people can talk to machines.

6 MIN

EDGE AI · GENERATIVE AI

What Is a Local LLM?

A local LLM is a large language model that runs on hardware you control, such as a PC, server or edge AI computer, instead of a cloud service. Prompts and data stay on site and it works offline. The main hardware limit is memory: the model’s weights must fit in GPU or accelerator memory, usually after quantization.

7 MIN

EDGE AI · GENERATIVE AI

What Is a Large Language Model?

A large language model (LLM) is a neural network trained on huge amounts of text to predict the next word, or token, in a sequence. That one skill lets it answer questions, summarise, translate and write code. Most LLMs use the transformer architecture and have billions of parameters, so they run on GPUs or other AI accelerators.

7 MIN

EDGE AI · ROBOTICS

What Is a Humanoid Robot?

A humanoid robot is a robot built in roughly human form, usually with a head, a torso, two arms and two legs, so it can work in spaces and with tools designed for people. Modern humanoid robots use cameras, AI models and an onboard computer to walk, balance and handle objects.

6 MIN

EDGE AI · DEEP LEARNING

What Is a Convolutional Neural Network (CNN)?

A convolutional neural network (CNN) is a deep learning model designed for grid-shaped data such as images. Its layers slide small learned filters across the input to pick out edges, textures and shapes, then combine them to classify an image, detect objects or label pixels. CNNs remain the most widely deployed vision models on edge hardware.

6 MIN

EDGE AI · SOFTWARE

PyTorch vs TensorFlow: What Is the Difference?

PyTorch and TensorFlow are open-source frameworks for building and training neural networks. PyTorch, from Meta, runs like ordinary Python and leads in research and generative AI. TensorFlow, from Google, has long-established tools for production serving and mobile deployment. Both run on NVIDIA GPUs, and models from either export to ONNX and TensorRT for edge hardware.

7 MIN

EDGE AI · FUNDAMENTALS

Machine Learning vs Deep Learning: What Is the Difference?

Deep learning is a type of machine learning. Classic machine learning applies algorithms such as decision trees or regression to features a person picks out of the data. Deep learning uses neural networks with many layers that learn those features from raw images, audio or text. It needs more data and computing power, usually a GPU.

6 MIN

GRAPHICS · GPU CARDS

What Is a Graphics Card?

A graphics card, also called a video card, is an expansion card that carries a graphics processing unit (GPU) with its own memory, power circuits, cooling and display outputs. It plugs into a PCIe slot and draws images for the screens, and it also speeds up video, AI inference and other parallel computing work.

6 MIN

PROCESSORS · ARM

What Is the Qualcomm Hexagon NPU?

Qualcomm Hexagon is the processor block Qualcomm uses for AI inference in its Snapdragon and Dragonwing chips. It began as a digital signal processor (DSP) for audio, imaging and sensors, and has grown into a dedicated NPU. It works alongside the CPU and Adreno GPU, which together Qualcomm calls the Qualcomm AI Engine.

6 MIN

PROCESSORS · NPU

What Is Intel AI Boost?

Intel AI Boost is Intel’s name for the NPU, or neural processing unit, built into Intel Core Ultra processors. It runs AI inference such as image recognition, speech and background effects at low power, freeing the CPU and GPU. Software reaches it through Intel OpenVINO, Windows ML, DirectML and ONNX Runtime.

6 MIN

PROCESSORS · NPU

What Is AMD Ryzen AI?

AMD Ryzen AI is AMD’s name for the AI hardware and software in its Ryzen processors. At its centre is an NPU based on AMD’s XDNA architecture, built onto the processor beside the CPU cores and Radeon graphics. In our range, Ryzen 8000-series processors carry a 16 TOPS NPU and Ryzen AI 300-series processors a 50 TOPS NPU.

6 MIN

PROCESSORS · AI

NPU vs CPU: What Is the Difference?

A CPU is the general-purpose processor that runs the operating system, applications and any code you give it. An NPU, or neural processing unit, is built only for the matrix maths in AI models, so it runs inference using far less power than the CPU and leaves the CPU cores free. Many modern processors now include both on one chip.

6 MIN

AI PERFORMANCE · NPU

How Many TOPS Does an NPU Need?

There is no single figure: the TOPS an NPU needs depends on the model, image resolution, frame rate and number of camera streams. Check whether a rating is for the NPU alone or the whole platform, since vendors often add CPU and GPU TOPS together. The reliable test is to benchmark your own model on the target processor.

6 MIN

AI SOFTWARE · MODEL FORMATS

What Is ONNX?

ONNX, the Open Neural Network Exchange, is an open file format for AI models. A model trained in PyTorch or TensorFlow can be exported to ONNX and then run by different inference engines, such as ONNX Runtime, Intel OpenVINO and NVIDIA TensorRT, so the same model can move between hardware without being rebuilt.

6 MIN

GPU · ARCHITECTURE

CUDA Cores vs Tensor Cores: What Is the Difference?

CUDA cores are the general-purpose parallel processors in an NVIDIA GPU, each handling one floating-point or integer operation at a time. Tensor Cores are specialised blocks that multiply small matrices in a single step at reduced precision, which is the core maths of AI. A GPU uses CUDA cores for graphics and general compute, and Tensor Cores for AI.

6 MIN

EDGE AI · DEPLOYMENT

Edge AI vs Cloud AI: What Is the Difference?

Edge AI runs models on a computer beside the cameras or machines, so results arrive in milliseconds and the data stays on site. Cloud AI sends data to remote servers, which offer large scale but add delay, bandwidth costs and a dependence on the connection. Many systems combine both: they train in the cloud and infer at the edge.

6 MIN

AI SOFTWARE · MODELS

What Is INT8 Quantization?

INT8 quantization converts an AI model’s weights and calculations from 32-bit floating point (FP32) to 8-bit integers. The model becomes about a quarter of the size and runs faster on NPUs, GPUs and AI accelerators, usually with only a small loss of accuracy. It is a standard step when preparing a trained model for inference at the edge.

6 MIN

EDGE AI · PROCESSORS

NPU vs TPU: What Is the Difference?

NPU, or neural processing unit, is a general term for an AI engine built into a processor, used by Intel, AMD, Qualcomm, NXP, Rockchip and others. TPU, or tensor processing unit, is Google’s name for its own AI chips: Cloud TPUs in Google’s data centres, and the small Edge TPU for devices. Both speed up neural network maths.

7 MIN

EDGE AI · ACCELERATORS

What Is the Hailo-10H?

The Hailo-10H is an AI accelerator from Hailo for generative AI on edge devices. It delivers up to 40 TOPS (INT4) and carries 8 GB of LPDDR4 memory on an M.2 M key module, so small language models, vision language models and speech can run locally at a few watts, without a cloud connection or a large GPU.

6 MIN

GPU · NVIDIA

What Is the NVIDIA RTX 2000 Ada?

The NVIDIA RTX 2000 Ada is a low-profile professional GPU based on the Ada Lovelace architecture, with 16 GB of GDDR6 memory with ECC and a 70 W maximum power rating. It suits AI inference, vision and multi-display work in compact and rugged computers, and also comes as a single-slot 50 W RTX 2000E and an embedded MXM module.

6 MIN

EDGE AI · HARDWARE

What Is an AI Accelerator?

An AI accelerator is hardware designed to run neural networks faster and more efficiently than a general-purpose CPU. The main types are GPUs, NPUs built into processors, dedicated accelerator chips such as Hailo modules, and FPGAs. Each takes over the repetitive maths of AI inference, so the CPU stays free for the rest of the system.

6 MIN

GPU · INTEL

What Is Intel Arc? Intel’s GPUs for Edge and Embedded Systems

Intel Arc is Intel’s graphics brand. It covers discrete Intel Arc GPUs, the A-series and newer B-series cards and modules, and the Arc graphics built into Intel Core Ultra processors. Arc GPUs handle displays, video encoding and decoding, and AI inference, and they are supported by Intel’s OpenVINO toolkit for running AI models.

6 MIN

PROCESSORS · GPU

GPU vs CPU: What Is the Difference?

A CPU has a small number of powerful cores that handle many different tasks one after another, such as running the operating system and applications. A GPU has thousands of simpler cores that work in parallel, which makes it far faster for graphics, video and AI inference, where the same calculation is repeated across large amounts of data.

7 MIN

GPU · FORM FACTORS

What Is an MXM GPU?

An MXM GPU is a graphics processor on a Mobile PCI Express Module, a compact standard board that mounts flat on a carrier instead of standing in a PCIe slot. It lets laptops, embedded systems and rugged computers add discrete NVIDIA, AMD or Intel graphics in a small space, often cooled through the chassis rather than by its own fan.

6 MIN

GPU · EDGE AI

What Is the NVIDIA L4?

The NVIDIA L4 is a low-profile, single-slot GPU based on the Ada Lovelace architecture, with 24 GB of GDDR6 memory and a 72 W power rating. It is built for AI inference and video processing, and its low power and passive heatsink let it fit compact servers and rugged edge computers where larger GPUs will not.

6 MIN

ROBOTICS · HARDWARE

What Is a Robot Controller?

A robot controller is the computer that runs a robot: it reads sensors, makes decisions and commands the motors. Industrial arms use the manufacturer’s motion controller, while mobile robots and vision-guided systems increasingly use an edge AI computer as the robotics controller, running camera perception, planning and communication with drives over CAN or Ethernet.

6 MIN

AI VISION · BASICS

What Is Pose Estimation?

Pose estimation is a computer vision technique that finds the position of a person’s body in an image or video by locating keypoints such as the head, shoulders, elbows, hips and knees, and joining them into a skeleton. Tracked over time, it shows how someone is moving, which supports safety monitoring, ergonomics, gesture control and sports analysis.

6 MIN

EDGE AI · BASICS

What Is a Digital Twin?

A digital twin is a virtual model of a physical asset, system or place that is kept up to date with real data from sensors, cameras and scans. Engineers use it to see the current condition of the real thing, spot changes over time, and test decisions on the model before acting on the asset itself.

6 MIN

NVIDIA JETSON · PERFORMANCE

What Is Jetson Super Mode?

Jetson Super Mode is a set of higher-power operating modes NVIDIA added for Jetson Orin Nano and Orin NX modules with JetPack 6.2. Raising clock speeds and power limits lifts AI performance substantially, for example to up to 67 TOPS on Orin Nano 8GB and up to 157 TOPS on Orin NX 16GB, on the same hardware.

6 MIN

EDGE AI · BASICS

What Does TOPS Mean in AI?

TOPS stands for tera operations per second: the number of trillion simple operations an AI processor can perform each second. It is the headline figure used to compare AI accelerators such as NVIDIA Jetson modules and the NPUs in Intel and AMD processors. It is usually quoted for 8-bit integer (INT8) maths, and sometimes with sparsity, which doubles the figure.

6 MIN

EDGE AI · HARDWARE

iGPU vs dGPU: What Is the Difference?

An iGPU (integrated GPU) is built into the processor and shares the system memory, so it adds little power, heat or cost. A dGPU (discrete GPU) is a separate graphics processor with its own memory, offering far more performance but needing more power, cooling and space. For edge AI, iGPUs suit lighter inference and dGPUs suit heavy workloads.

7 MIN

EDGE AI · SOFTWARE

What Is OpenVINO?

OpenVINO is Intel’s open-source toolkit for optimising and running AI inference. It takes models trained in frameworks such as PyTorch, TensorFlow or ONNX, optimises them, and runs them efficiently on Intel hardware, including CPUs, integrated GPUs such as Iris Xe and Arc, discrete Arc GPUs and the NPUs in Core Ultra processors.

6 MIN

AI VISION · BASICS

What Is Object Tracking?

Object tracking is a computer vision task that follows an object from one video frame to the next, keeping track of where it is and where it is going. Detection finds objects in each frame; tracking links those detections over time, so a system can follow a person, vehicle or drone and act on its movement.

6 MIN

CONSTRUCTION · EDGE AI

How Is AI Used in Construction?

AI in construction uses machine learning on drawings, schedules, camera, LiDAR and machine data to make projects safer and more productive. Common uses include planning and estimating, detecting people near moving equipment, tracking progress against plans, monitoring stockpile volumes, guiding autonomous machinery and predicting when machines need maintenance.

6 MIN

AGRICULTURE · MACHINE VISION

What Is Precision Spraying?

Precision spraying applies herbicide or pesticide only where it is needed, instead of covering the whole field. Cameras on the sprayer boom capture the ground as the machine moves, a computer identifies weeds in real time, and individual nozzles fire on each target. It is also called spot spraying or targeted spraying.

6 MIN

SAFETY · AI VISION

What Is a Pedestrian Detection System?

A pedestrian detection system uses sensors on a vehicle, such as AI cameras, radar, LiDAR or worker-worn tags, to spot people nearby and warn the operator or slow and stop the vehicle. In cars it supports automatic emergency braking. On forklifts and AGVs it reduces collisions in busy warehouses and factories.

6 MIN

EDGE AI · BASICS

What Is AIoT?

AIoT, the Artificial Intelligence of Things, is the combination of AI with IoT devices. Sensors and cameras collect data, and AI models analyse it, often on an edge computer close to the devices, so the system can recognise what is happening and act on its own instead of only reporting data.

6 MIN

RETAIL · DATA

What Is Retail Analytics?

Retail analytics is collecting and analysing data from a retail business, such as sales, footfall, dwell time, queues and stock levels, to make better decisions about layout, staffing, promotions and inventory. In physical stores, edge AI on camera feeds now adds live insight into how shoppers move, wait and buy.

6 MIN

EDGE AI · BASICS

AI Inference vs Training: What Is the Difference?

Training is how an AI model learns: it is fed large sets of labelled data and adjusted until it predicts accurately, which needs powerful GPUs. Inference is using the trained model on new data to make a decision, which is lighter per image and usually runs in real time on an edge computer beside the camera or machine.

6 MIN

IIOT · MAINTENANCE

What Is Predictive Maintenance?

Predictive maintenance uses data from sensors on machines, such as vibration, temperature and current, to predict when a part is likely to fail, so it can be repaired just before it breaks. It avoids both unplanned breakdowns and the wasted cost of replacing parts on a fixed schedule.

6 MIN

MACHINE VISION · BASICS

What Is Machine Vision?

Machine vision is the use of cameras, lighting and a computer to inspect, measure, identify or guide objects automatically. In industry it checks products for defects, reads codes, verifies assembly and guides robots, faster and more consistently than a person can, and today it increasingly runs AI models on edge computers beside the line.

7 MIN

EDGE AI · UAV & DRONES

How Does a Jetson Connect to a Flight Controller?

An NVIDIA Jetson connects to a drone flight controller such as a Pixhawk over a serial UART or Ethernet, and sometimes CAN, usually speaking MAVLink. The Jetson runs the AI and autonomy as the companion computer and sends commands, waypoints and position estimates; the flight controller flies the aircraft and returns telemetry.

6 MIN

EDGE AI · UAV AND DRONES

What Is a Drone Companion Computer?

A drone companion computer is a small AI computer that runs alongside the flight controller on a UAV. The flight controller flies the aircraft; the companion computer runs computer vision and AI to give the drone autonomy, such as obstacle avoidance, navigation and object detection. NVIDIA Jetson is a common choice for its AI performance per gram and per watt.

6 MIN

EDGE AI · SOFTWARE

What Is Linux for Tegra (L4T)?

Linux for Tegra, or L4T, is the NVIDIA operating system for Jetson edge AI modules: an Ubuntu-based Linux plus the board support package (bootloader, kernel and drivers) that makes Linux run on Jetson hardware. NVIDIA now calls it Jetson Linux, and it is the base layer of JetPack.

7 MIN

EDGE AI · COMPUTER VISION

What Is NVIDIA VPI?

NVIDIA VPI, the Vision Programming Interface, is NVIDIA’s library of computer-vision and image-processing algorithms. It lets the same vision operations run on different processors, such as the GPU, the CPU and dedicated vision hardware on Jetson, through one API, so pre-processing can move off the GPU and leave it free for AI inference.

6 MIN

EDGE AI · GPU COMPUTING

What Is CUDA?

CUDA is the NVIDIA platform for running general-purpose code on a GPU. It lets software use the thousands of cores in a GPU for parallel work, and it is the foundation that every other NVIDIA accelerated library, including the AI libraries, is built on. CUDA is what allows an application to use the GPU at all.

7 MIN

EDGE AI · INFERENCE

What Is TensorRT?

TensorRT is NVIDIA’s inference optimiser and runtime. It takes a trained AI model, usually as an ONNX file, and compiles it into an engine tuned for one specific NVIDIA GPU, fusing layers and using lower precision such as FP16 or INT8. Much of the real-time performance of a deployed edge AI application comes from running the model through TensorRT.

7 MIN

EDGE AI · DEEP LEARNING

What Is cuDNN?

cuDNN, the CUDA Deep Neural Network library, is NVIDIA’s library of GPU-accelerated building blocks for deep learning: optimised implementations of operations such as convolutions, attention, pooling and normalisation that neural networks are made of. It sits between CUDA and AI frameworks such as PyTorch and TensorFlow, which call it to run their layers on NVIDIA GPUs.

6 MIN

EDGE AI · NVIDIA JETSON

Jetson T5000 vs T4000: What Is the Difference?

The Jetson T5000 and T4000 are the two NVIDIA Jetson Thor modules, both built on the Blackwell GPU. The T5000 is the top variant, with up to 2,070 FP4 TFLOPS, 128 GB of memory and up to 130 W. The T4000 is the lower-power variant, with up to 1,200 FP4 TFLOPS, 64 GB and up to 70 W.

6 MIN

EDGE AI · SOFTWARE

What Is NVIDIA JetPack?

NVIDIA JetPack is the software development kit for the Jetson platform. It provides the operating system, Jetson Linux (also called Linux for Tegra, based on Ubuntu), plus the AI, computer vision and multimedia libraries, including CUDA, cuDNN and TensorRT, that let a Jetson run GPU-accelerated AI. Each JetPack version supports particular Jetson modules.

7 MIN

EDGE AI · PRODUCT LIFECYCLE

How Long Is NVIDIA Jetson Supported?

NVIDIA gives Jetson modules long, published production lifecycles, which is part of why they suit industrial designs. Current Orin modules run well into the 2030s, and Industrial variants carry a ten-year production life. The exception is the older LPDDR4 generation, Xavier and TX2, which NVIDIA has brought to early end-of-life, with last-time-buy orders closing on 1 July 2026.

6 MIN

EDGE AI · NVIDIA JETSON

What Replaced the NVIDIA Jetson Xavier?

The NVIDIA Jetson Orin family replaced the Xavier generation. The Xavier NX maps to the Orin NX and the AGX Xavier to the AGX Orin, with a large jump in AI performance per watt and the same JetPack software stack. With Xavier now end-of-life and its last-time-buy closed, Orin is the migration path.

6 MIN

EDGE AI · NVIDIA JETSON

Is the Jetson Orin Nano Super Worth It?

Yes, in almost every case, because the Super is a free software update rather than new hardware. JetPack lifts an existing Jetson Orin Nano from 40 to 67 TOPS by unlocking higher clocks and power, so deployed and new designs gain roughly 70 percent more AI performance at no hardware cost, provided the cooling handles 25 W.

6 MIN

EDGE AI · NVIDIA JETSON

Jetson Orin Nano vs Orin NX: What Is the Difference?

The Jetson Orin Nano delivers up to 67 TOPS in 7 to 25 W with 4 to 8 GB of memory, suited to entry-level, low-power vision. The Orin NX steps up to 157 TOPS in 10 to 40 W with 8 to 16 GB, adds deep learning accelerators and a hardware video encoder, and handles multi-camera, heavier inference.

6 MIN

EDGE AI · NVIDIA JETSON

Does Jetson Thor Need a New Carrier Board?

Yes. The NVIDIA Jetson Thor module uses a different connector, higher power (up to 130 W) and new I/O compared with AGX Orin, so it needs a carrier board designed for Thor rather than an Orin board. Moving from AGX Orin to Thor is a platform change, not a drop-in module swap.

6 MIN

EDGE AI · NVIDIA JETSON

What Is the Difference Between a Jetson Module and a Developer Kit?

An NVIDIA Jetson developer kit is a reference carrier board with a module fitted, meant for evaluation and prototyping. A production system uses the production module on a carrier board designed for your I/O, environment and lifecycle. The silicon and software are the same; the kit is not what you deploy.

6 MIN

NETWORKING · EDGE AI

What is RDMA and why does it matter for edge AI?

RDMA (remote direct memory access) lets one computer read from or write to another computer’s memory through the network card, without passing the data through either CPU or operating system. It cuts latency and frees processor time, which matters for AI clusters and multi-node edge AI. It runs over InfiniBand, or over Ethernet as RoCEv2 or iWARP.

7 MIN

NETWORKING · EDGE AI

Is 2.5GbE fast enough for edge AI?

For many edge AI systems, yes. 2.5GbE carries about 2.35 Gbit/s of real traffic over existing Cat5e cable, enough for single-camera inference, compressed multi-camera video and gateway aggregation. It usually falls short for uncompressed machine-vision cameras, multi-sensor fusion and multi-node inference clusters, where 10GbE with RDMA has become the practical baseline.

5 MIN

Products

Shop this topic

Products for edge AI and NVIDIA Jetson

All edge AI products

CategoryNVIDIA JetsonJetson Orin and Thor modules, developer kits and computers for AI at the edge.View products

CategoryEdge AI ServersRackmount and short-depth servers with NVIDIA GPUs for on-site inference.View products

CategoryGPU CardsLow-profile, single-slot and MXM GPUs for compact and rugged computers.View products

CategoryNVIDIA Jetson Carrier BoardsCarrier boards that turn a Jetson module into a computer with the I/O you need.View products