CONTENT HUB
Here is our guide to decoding all the terms and abbreviations you'll see as you navigate the ever-evolving world of AI.
No glossary terms match your search.
A technique for evaluating the importance of a feature or component by temporarily removing it from a model. The model is retrained without that feature or component, and if the retrained model performs significantly worse, then the removed feature or component is deemed important.
AI systems that can act autonomously to achieve goals by planning, reasoning, making decisions, and taking actions over multiple steps.
A hypothetical form of AI that can understand, learn, and apply knowledge across a wide range of tasks at a human-like level.
The field of computer science focused on creating systems that can perform tasks that typically require human intelligence, such as reasoning, learning, and decision-making.
An AI agent is an autonomous software system that uses artificial intelligence to achieve specific goals. To separate 'agentic' AI from a conversational chatbot it must show autonomy, planning and reasoning capabilities.
An AI factory is a scalable, repeatable system that turns raw data into valuable AI-driven insights, products, or services. It typically consists of automated pipelines for data collection, pre-processing, model training, deployment, and monitoring. Think of it as an assembly line measured by tokens for artificial intelligence - standardising how businesses build, test, and scale AI solutions.
A commonly used term for an AI model - essentially the programming that tells the computer how to learn and operate on its own.
A term used to describe output from a generative AI system that favours quantity over quality.
A set of rules that allows different software systems to communicate with each other. Popular AI APIs include HuggingFace, pandas and numPy. Most APIs are built upon a framework.
Technology that converts spoken language into text. It utilises machine learning and linguistics to automatically convert spoken language into written text in real-time, and powers virtual assistants such as Siri and Alexa, transcription software, and live captioning.
A technique used in generative models, especially language generation, which helps focus on parts of input data, improving the relevance and coherence of generated content.
A performance metric for classification models, indicating how well the model distinguishes between classes. It provides a single number between 0 and 1, where higher scores generally indicate a better model - 0.5 - the model is doing no better than random guessing; 0.5 to 0.7 - poor or weak discriminative power; , 0.7 to 0.8 - acceptable performance; 0.8 to 0.9 - excellent performance; and 0.9 to 1.0 - outstanding predictive power.
A neural network used in generative AI that is trained to encode input data as representations, then decode representations back to the original data.
A training method that allows AI and neural networks to learn from their mistakes. It works by calculating the exact impact that millions of internal weights have on a model's errors, enabling fast adjustments that improve prediction accuracy over time.
A batch refers to a group of data points or requests processed together at the same time (in one iteration), rather than handling them one by one in realtime. It is used to maximise computational efficiency, reduce costs, and speed up both model training and data analysis.
A variation of gradient descent (GD) using random subsets of data for faster training. It calculates the error over the entire dataset before taking a single step. It guarantees a smooth path to the minimum but can be extremely slow and memory-intensive for large datasets.
A language model that understands context in text by looking at words in both directions. It is used to deeply analyse, classify or understand text.
Systematic error in AI outputs caused by flawed assumptions, data imbalance, or design choices. It occurs when AI models learn prejudiced assumptions from historical training data or developer inputs, subsequently amplifying inequalities in final models.
Large datasets that AI systems use to learn patterns and make predictions.
The technique of combining features or aspects from multiple sources or datasets to create new, hybrid outputs. Typically used in generative AI, blending can lead to deeper and more complex forms of analyses.
A chatbot is a software application designed to simulate human conversation through text or voice.
A task where AI predicts a category or label. Classification types can be either binary - e.g. where email is classed as either spam or not spam; or multi-label, where the model assigns multiple categories to a single piece of data - e.g., tagging a photo with multiple labels such as "beach," "sunset," and 'ocean'.
Clustering is an unsupervised machine learning technique used to group similar, unlabelled data points together. By discovering hidden patterns or natural groupings within raw data without relying on predefined categories, AI systems can automatically organise complex information.
A type of neural network commonly used for image processing and computer vision tasks.
A field of AI that enables machines to interpret and understand visual data such as images and videos.
The process of bundling an AI model, its source code, and all its required dependencies into a single, lightweight, and portable executable package, or container. This allows the AI application to run identically and flawlessly across any environment, whether that is a local laptop, a cloud server, or an edge device. Applications used to manage containerised environments include Docker and Kubernetes.
Conversational AI refers to technologies, like virtual agents and chatbots, that simulate human-like conversations. By combining NLP and LLMs, these systems understand context, recognise intent, and generate natural responses across text or voice.
The CPU in an AI system acts as the director. While it isn't used for calculations like GPUs, the CPU handles data preparation, system memory management, and orchestrates the overall AI workflow. Common CPUs for AI Workstations include Intel Core, Intel Xeon W and AMD Ryzen and AMD Threadripper; while Intel Xeon and AMD EPYC appear in RTX PRO, MGX and HGX servers; and NVIDIA Grace and Vera feature in Superchip servers.
CUDA (Compute Unified Device Architecture) is a software platform for parallel computing, leveraging the processing power of NVIDIA GPUs for non-graphical tasks such as AI, HPC and complex simulations.
A CUDA core is the primary processor inside NVIDIA GPUs. Thousands of CUDA cores work together in parallel; generally, more cores means higher performance. Other workloads are accelerated by specialised RT or Tensor cores.
A collection of data used for training or testing AI models.
The first stage of the AI journey where AI models are born, most commonly using libraries or frameworks. The development phase precedes training and inferencing.
NVIDIA's flagship range of AI appliances, combining hardware and ready-to-use software, including the DGX Spark and DGX Station for development; and DGX B200 and DGX B300 for the most demanding workloads including agentic and physical AI.
A digital twin is a highly detailed virtual replica of a physical object, system, or process. By continuously integrating realtime sensor data and historical information, it accurately mimics its real-world counterpart. This allows organisations to monitor performance, run simulations, and predict future outcomes safely and efficiently. Digital twins are often built using the NVIDIA Omniverse platform.
A subset of machine learning using neural networks with many layers to model complex patterns.
An advanced, specialised processor found in modern datacentres and enterprise computing used in AI. A DPU acts to offload infrastructure-heavy tasks like networking, data encryption, firewall security, and storage management. Learn more by watching our explainer video.
The process of analysing datasets to summarise their main characteristics, to uncover hidden patterns, spot anomalies, and check assumptions.
NVIDIA brings its technologies to market through a connected ecosystem of world-class partners that design, build, and deliver solutions on NVIDIA's accelerated computing and AI platforms. NVIDIA Partner Network (NPN) Program partners are trusted experts dedicated to accelerating innovation, empowering transformation, and unlocking faster time to value, applying deep domain knowledge and validated NVIDIA competencies to every engagement. Elite partners represent the deepest level of partnership with NVIDIA and demonstrate the highest level of commitment to the partnership.
A numerical representation of data (such as words or images) that captures meaning or relationships.
In generative AI, an encoder is a part of a model, often a neural network, which compresses data into a more compact representation, typically used in autoencoder architectures.
One complete pass through the entire training dataset during model training - sometimes referred to as an iteration.
Ethernet is a wired networking protocol that connects devices on a local network (LAN) using physical cables. It is used between AI servers and storage devices at speeds of up to 80,000Gb/s for high bandwidth low latency throughput. An alternative network protocol is InfiniBand.
A data integration process used to gather raw data from multiple sources. It is cleaned by removing duplicates, then formatted and structured for consistency and quality. The clean, processed data is written and stored into a target destination, making it ready for business intelligence, analytics, and AI projects.
Principles guiding ethical AI development and governance. It addresses the risks of automated decision-making - like embedded bias and "black box" complexity - to ensure algorithms are equitable, responsible, and understandable to humans.
An individual measurable property or characteristic used as input to an AI model.
A technique where models are trained across multiple devices without sharing raw data. It is designed to protect sensitive information, PII, and maintain anonymity.
A measure of a processor's performance, indicating how many floating-point calculations it can perform per second. Common examples are TeraFLOPS or TFLOPS (trillions); and PetaFLOPS or PFLOPS (quadrillions).
Adjusting a pre-trained model on a smaller, specific dataset to improve performance for a specific task.
The percentage of actual negative cases an AI model incorrectly labels as positive. It is used with TPR to calculate an ROC plot and generate an AUC score.
A massive, highly versatile artificial intelligence system trained on vast datasets. It serves as a base layer from which developers and businesses can customise and fine-tune specialised applications.
A framework (sometimes called a library) is a collection of pre-written code and tools that helps developers easily build, train, and deploy neural networks without writing complex mathematical formulas from scratch. Examples include PyTorch, TensorFlow, Keras and JAX.
A model using two networks (generator and discriminator) to create realistic synthetic data. The generator creates synthetic data, while the discriminator evaluates its authenticity, driving both to improve until the generated data is indistinguishable from real data.
An optimisation algorithm used to minimise errors by adjusting AI model parameters. There are two primary variants - Batch Gradient Descent (BGD), Stochastic Gradient Descent (SGD).
A range of consumer GPUs from NVIDIA, that offer cost-effective AI performance. These are found in entry-level to mid-range AI Workstations from 3XS Systems.
AI models that can create new content such as text, images, audio, or code.
A type of large language model designed for generating human-like text.
A processor optimised for parallel computations, widely used in AI training. NVIDIA produces the most popular GPUs, with recent architectures include Rubin (2027), Blackwell (2025), Ada Lovelace (2022) and Ampere (2020).
Commonly referred to as Video RAM (VRAM), it is dedicated memory onboard a GPU card to store AI model datasets, frameworks and containers. It is distinct from system memory that serves the CPU, or unified memory that is shared between a CPU and GPU in a Superchip server.
A technology that allows a single physical GPU to be divided into multiple virtual instances. This enables several users or virtual machines (VMs) to share the same hardware simultaneously, making high-performance rendering and AI processing accessible and cost-effective.
The accuracy of the training set's classification for supervised learning. It is the benchmark against which the performance of generative models is measured.
When an AI model generates incorrect or fabricated information that appears plausible.
A type of ultra-fast 3D-stacked server memory used in HPC and AI accelerators. By stacking memory chips vertically, HBM delivers massive bandwidth, incredible power efficiency, and a compact footprint that traditional GDDR memory cannot match. HBM memory is usually found alongside SXM GPUs in powerful AI servers such as NVIDIA HGX and DGX. There are several generations differentiated by an incrementing number e.g., HBM3, HBM3e, HBM4, with each offering greater performance than the last.
HGX servers are an NVIDIA range of AI hardware systems designed around the same SXM GPUs used in their DGX server range, but allowing for greater customisability of the chassis and other components. Like DGX, they are designed for the most demanding workloads including agentic and physical AI.
HPC refers to the use of supercomputers or computer clusters to solve highly complex computational tasks at speed - it is also used as an umbrella term that can include ML, DL, and AI.
A configuration setting chosen before training that controls how a model learns (e.g., learning rate).
The third and final stage of the AI journey where AI models are deployed and apply what they have learned to new data, generating predictions or outputs based on their training. The inferencing phase follows both development and training phases but may include re-training as new data becomes available to improve models.
InfiniBand is a wired networking protocol that connects devices on a LAN using cables. It is used between AI servers and storage devices at speeds of up to 800Gb/s for high bandwidth low latency throughput. An alternative network protocol is Ethernet.
One complete pass through the entire training dataset during model training - sometimes referred to as an epoch.
A lightweight data format used for data exchange. It is primarily used to send and receive structured data between a web application and a server.
In machine learning, a kernel is a function used to transform data into a higher-dimensional space, enabling better classification when data is not linearly separable in the original space.
An unsupervised machine learning algorithm used to group unlabelled data into a specific number of clusters based on their similarities. It works by finding the mathematical centre (centroid) of data points and continually refining the group.
The abstract, high-dimensional space in which the AI model represents data. It is essential in models including GANs and VAEs.
A library (sometimes called a framework) is a collection of pre-written code and tools that helps developers easily build, train, and deploy neural networks without writing complex mathematical formulas from scratch. Examples include PyTorch, TensorFlow, Keras and JAX.
Statistical technique used for modelling the relationship between a dependent variable and independent variables. While not generative, it is a foundational concept in machine learning.
A type of AI trained on massive text datasets to understand and generate human-like language and responses.
A function that measures how far an AI model's predictions are from the actual results.
A parameter that determines how much a model adjusts its weights during training.
A type of recurrent neural network designed to remember long-term dependencies.
A measure of prediction accuracy based on average absolute differences. It is a popular metric used in machine learning to evaluate how close predictions or forecasts are to the actual outcomes.
A neural network architecture used for modelling probability distributions over continuous variables. MDNs are employed in generative AI to manage uncertainty and generate probabilistic outputs.
MGX servers are an NVIDIA range of AI hardware systems designed around either a Superchip design combining a CPU, GPU, and unified memory; the RTX PRO range of PCIe GPUs. These servers are designed for dense deployments supporting platforms such as Omniverse or AI factories.
This is an NVIDIA-specific GPU virtualisation technology that enables multiple users to share access to a single GPU. It fully isolates at the hardware level, allowing memory, cache, cores, and containers to be partitioned into as many as seven independent instances.
A variation of gradient descent (GD) using random subsets of data for faster training. It splits the training data into small batches - 32 or 64 samples - to update the parameters. It strikes a balance between the stability of BGD and the efficiency of SGD.
A subset of AI where systems learn patterns from data and improve performance over time without being explicitly programmed.
A trained system that makes predictions or decisions based on input data.
A model architecture that uses multiple specialised sub-models to improve efficiency and accuracy. Instead of using the entire network to process every piece of data, a routing mechanism selectively activates only the most relevant "experts," enabling massive model sizes with much higher computational efficiency.
A metric measuring the average squared difference between predicted and actual values. It is heavily used in regression tasks to evaluate model accuracy and guide algorithms during training.
AI that can process and combine multiple types of input or output - including text, images, audio, and video.
A computational model inspired by the human brain, composed of layers of interconnected nodes - termed neurons.
The process of generating human-like text from structured data.
A branch of AI focused on enabling computers to understand, interpret, and generate human language.
NVLink is a high-bandwidth interconnect that links compatible NVIDIA GPUs together, allowing them to share memory and data directly for improved multi-GPU performance. Sometimes abbreviated to NVL in an NVIDIA product description.
NVSwitch is a dedicated hardware chip created by NVIDIA that functions as a high-speed, multi-port switch for connecting GPUs. It works alongside NVLink to allow all GPUs in a system to communicate with one another simultaneously at maximum speed without bottlenecks.
Technology that extracts text from images or scanned documents.
Omniverse is a cloud computing platform from NVIDIA that enhances 3D visualisation and development workflows, providing a collaborative space accessible by multiple users. Omniverse follows real-world physics, ensuring simulations are as accurate as possible, guaranteeing their integrity when transferred to the real world.
A technique used in generative AI used to identify data points that are different from the training data. OOD detection helps improve the reliability of generative models.
When an AI model learns training data too well, including noise, and performs poorly on new data.
A method of performing multiple computations simultaneously, used in generative AI to speed up training and inference processes by leveraging GPUs.
AI systems that can act are designed to be embedded in physical devices, such as robots or autonomous vehicles, intended for human interaction.
Data that can be used to identify an individual.
The input or instruction given to an AI - especially generative AI - to guide its output.
Internal variables of a model that are learned during training.
A type of computing that uses quantum bits (qubits) to perform calculations at speeds much faster than classical computers. It is an emerging field with potential applications in AI and machine learning.
A request for information from a database or dataset, often used in the context of AI models retrieving data.
An AI approach that combines external information with LLMs. Instead of relying solely on static training data, RAG allows AI to search, retrieve, and synthesise up-to-date or company-specific information, drastically reducing hallucinations.
RDMA is a technology that allows data to be transferred directly between the memory of two computers, via an InfiniBand NIC, without involving the operating system or CPU on either end. It is employed within NVIDIA Networking Smart NICs, Super NICs and DPUs to offload tasks from the CPU and speed up networking for demanding HPC and AI workloads.
An AI model that provides personalised recommendations to users, commonly used in applications such as movie recommendations and product suggestions.
A supervised learning technique used to predict continuous numerical values. While classification categories data, regression finds patterns to estimate specific, measurable quantities.
A deep neural network architecture known for its skip connections, allowing it to train very deep models. ResNet is used in generative AI for various computer vision tasks.
A learning method for training AI where an agent learns dynamically through trial and error by rewarding positive behaviours and penalising negative ones, mimicking how humans naturally learn from experience. It is used to solve complex, sequential decision-making problems where the best path isn't known in advance.
A learning method for training AI where an agent learns by being given human evaluations to guide outputs.
A performance metric derived from MSE (Mean Squared Error), giving error in original units, used in model evaluation and reporting.
A neural network designed for sequential data such as text or time series.
A graphical plot that evaluates the performance of a binary classification model across all possible decision thresholds. It visualises the trade-off between sensitivity and specificity without depending on a single fixed cutoff point, and generates an AUC score.
RoCE (pronounced rocky) is a technology that allows data to be transferred directly between the memory of two computers, via an Ethernet NIC, without involving the operating system or CPU on either end. It is employed within NVIDIA Networking Smart NICs, Super NICs and DPUs to offload tasks from the CPU and speed up networking for demanding HPC and AI workloads.
RT (Ray Tracing) cores are a type of processor inside NVIDIA GPUs. They accelerate realtime ray tracing, simulating realistic light, reflections and shadows. More RT cores generally result in higher framerates when ray tracing is enabled.
A range of professional-grade GPUs from NVIDIA, that offer advanced AI performance. These are found in mid-range to high-end AI Workstations for AI development; and RTX PRO servers for Omniverse simulation from 3XS Systems.
A key component in transformers, a type of neural network architecture used in generative AI, which enables models to weigh the importance of distinct parts of input sequences when generating content.
A search technique that understands the meaning and context of words, phrases, or sentences rather than just matching exact keywords. It is commonly used in NLP tasks.
A variation of gradient descent (GD) using random subsets of data for faster training. It calculates the gradient and updates parameters using only one randomly selected training example per step. Although it speeds up training, unwanted noise can be introduced into results.
Refining a pre-trained model with labeled data for specific tasks.
Machine learning using labeled data for training.
A type of AI trained on specific text datasets to understand and generate targeted human-like language and responses.
A Smart NIC is a type of network card installed in an AI server. They are designed for high-bandwidth, low-latency throughput and employ RDMA or RoCE to remove overheads usually introduced by the CPU, system memory, and operating system. Learn more by watching our explainer video.
A Super NIC is a type of network card installed in an AI server. They are built around a high-performance network ASIC, making it a more streamlined and less computationally intensive solution compared to a DPU, so GPU-to-GPU communication is prioritised whilst minimising CPU overhead and latency. Learn more by watching our explainer video.
Training using data that generates its own labels. Instead of relying on manual human labelling, the algorithm automatically generates its own 'pseudo-labels' from the structure of the data itself.
Artificially generated data used instead of real-world data, designed to mimic the statistical properties, patterns, and structure of real-world data. It avoids PII and overcomes data scarcity.
SXM is a type of GPU format found in NVIDIA HGX and DGX servers, that acts as alternative to PCIe connections. Unlike standard PCIe GPUs, SXM modules connect directly to the server motherboard, allowing for much higher power delivery, ultra-fast memory access, and massive multi-GPU communication via technologies such as NVIDIA NVLinK. There are several generations differentiated by an incurring number e.g. SXM3, HBM4, HBM5, with each offering greater performance than the last.
Tensor cores are specialised processing units inside NVIDIA GPUs designed to accelerate matrix multiplication - the fundamental mathematical operation driving modern AI, ML, DL and HPC.
The Turing test is a behavioural benchmark for AI proposed by mathematician Alan Turing in 1950. It evaluates a machine's ability to exhibit intelligent, human-like conversation. If a human evaluator cannot reliably distinguish the machine from a human through interactions, the machine passes.
A fundamental unit of data - a word, part of a word, or symbol - processed by language models. A simple rule of thumb in English is that 1 token equals about 4 characters, or roughly 3/4 of a word. For example, 100 tokens will generally make up about 75 words.
The process of splitting text into individual units, or tokens, which can be words, subwords, or characters. Tokenisation is a crucial step in generative AI for text processing.
Tokens per second is the standard metric used to measure the speed at which an AI model processes or generates text.
A measure of how many operations (not just floating-point) a processor can perform per second, typically in trillions.
The percentage of actual positive cases the model successfully identifies. It is used with FPR to calculate an ROC plot and generate an AUC score.
A specialised processor designed for AI and machine learning tasks.
The second stage of the AI journey where AI models are scaled and fine-tuned. The training phase follows development and precedes inferencing.
A neural network architecture widely used in a deep learning architecture that processes sequential data (like words in a sentence) all at once, rather than one by one. By using an "attention mechanism," it understands the context and relationships between all parts of the data simultaneously.
Technology that converts written text into spoken words.
Learning from unlabelled data to find patterns or structures.
A standardised format for web addresses, commonly used in AI applications for data retrieval and processing.
A type of generative AI model that learns efficient data representations. VAEs represent data as a probability distribution, allowing the AI to generate entirely unique but realistic variations of training data.
A multimodal AI that combines an LLM with a vision encoder, enabling it to "see" and interpret visual data (like images and video) alongside text.
A numerical representation of data in the form of a list or array. Vectors are widely used in AI for data processing and transformations.
A technique used in NLP to represent words as vectors, capturing semantic relationships between them. Common models include Word2Vec and GloVe.
A comprehensive model of the real world that AI systems use to make predictions and decisions based on environmental data.
A JavaScript API for rendering 2D and 3D graphics in web browsers, commonly used in AI visualisation and interactive applications.
A set of methods and techniques that allow humans to understand, trust, and comprehend the decision-making processes of machine learning models.
A markup language used for structuring data in a hierarchical format. It is commonly used in AI applications for data exchange and processing.
A fundamental logic gate in computer science that outputs true only when the number of true inputs is odd. Used in neural networks and AI circuits.
A real-time object detection system that identifies and classifies objects in images or video.
A system integrator designing specialist AI hardware including AI Workstations for model development, RTX PRO servers for general purpose AI training, simulation and visualisation, and Superchip, MGX and HGX servers for demanding AI workloads such as LLMs, generative AI and agentic AI.