AI Glossary: 50+ Essential Terms to Understand Artificial Intelligence

A 

AI (Artificial Intelligence): The simulation of human intelligence processes by machines, including learning, reasoning, and self-correction. 

AI Accelerator: Specialized hardware or systems designed to speed up AI tasks, with NVIDIA GPUs being a leader in this space. 

Algorithm: A set of rules or instructions given to an AI system to help it learn, solve problems, and make decisions. 

Ampere Architecture: NVIDIA's GPU architecture designed to accelerate AI, data analytics, and HPC workloads. 

B 

Bare Metal: A physical server dedicated to a single tenant, offering full control over its resources without virtualization. 

Benchmarking: The process of measuring system or model performance against a standard to evaluate efficiency. 

Big Data: Large, complex datasets requiring advanced tools for processing, often used to uncover patterns. 

Black Box: An AI system where the decision-making process is opaque, even if inputs and outputs are visible. 

C 

Cluster: A group of interconnected computers (nodes) working as a unified system for complex computations. 

Colocation: Housing privately-owned servers in a third-party data center. 

Compute: The processing power available in cloud or data center environments for running applications. 

Containerization: A lightweight virtualization method that packages applications and dependencies into a container. 

Conversational AI: Technology that powers machines to engage in human-like conversations using natural language processing (NLP). 

CUDA (Compute Unified Device Architecture): NVIDIA’s parallel computing platform enabling general-purpose processing on GPUs. 

CUDA Core: The fundamental processing unit in an NVIDIA GPU, handling parallel computing operations. 

D 

Deep Learning: A subset of machine learning that uses neural networks with multiple layers to analyze complex data. 

DGX System: NVIDIA’s AI supercomputers designed for deep learning and AI workloads. 

F 

Finetuning: The process of training a pre-trained AI model on a smaller, task-specific dataset to enhance performance. 

FP16 (Half Precision Floating Point): A 16-bit floating point format that improves memory efficiency in AI training. 

FP32 (Single Precision Floating Point): A 32-bit format used in AI training and gaming for representing real numbers. 

G 

GDDR Memory: A type of memory in GPUs that provides high bandwidth for data access, critical in gaming and AI. 

Generative AI: AI that creates new content (text, images, video) based on patterns learned from existing data. 

GPU (Graphics Processing Unit): A processor designed for parallel processing, widely used in AI and machine learning. 

H 

HPC (High-Performance Computing): The use of supercomputers for solving complex problems requiring large-scale processing. 

Hypervisor: Software that runs virtual machines by separating the physical hardware from the operating system. 

I 

InfiniBand (IB): A high-speed networking standard used in HPC clusters for low-latency communication. 

Internet of Things (IoT): A network of connected devices that exchange data, often enhanced by AI for predictive tasks. 

IOPS (Input/Output Operations Per Second): A measurement of storage device performance. 

K 

Kubernetes: An open-source platform automating the deployment and management of containerized applications. 

L 

Latency: The time delay between data transmission and receipt, important for real-time applications. 

LLM (Large Language Model): AI models trained on vast amounts of text data to generate human-like language. 

Load Balancing: Distributing workloads across resources to optimize performance and avoid overloading any single system. 

M 

Machine Learning (ML): A subset of AI where machines use algorithms and statistical models to improve performance over time. 

MIG (Multi-Instance GPU): A feature allowing an NVIDIA GPU to be partitioned into smaller, isolated instances. 

N 

Natural Language Processing (NLP): AI that allows computers to understand, interpret, and generate human language. 

NIC (Network Interface Controller): A hardware component enabling a device to connect to a network. 

Node: A single machine or device in a network or cluster that processes and exchanges data. 

NVLink: NVIDIA’s high-speed interconnect for fast communication between GPUs and CPUs. 

NVSwitch: NVIDIA’s high-bandwidth switch allowing multiple GPUs to work together within a server. 

O 

OOB (Out of Band): A management method allowing access to network devices via a separate channel, even if the main network is down. 

Optimization: The process of improving system or model performance by adjusting parameters. 

P 

Pre-training: The initial training phase where AI models learn general patterns from large datasets. 

R 

RAID (Redundant Array of Independent Disks): A data storage technology combining multiple disks for redundancy and improved performance. 

S 

SLA (Service Level Agreement): A contract specifying the level of service expected from a provider. 

SLM (Small Language Model): A smaller AI model used for tasks where large-scale models are unnecessary. 

T 

Tensor: A mathematical object used in machine learning and AI, particularly in tensor processing units (TPUs). 

Tensor Core: Specialized cores in NVIDIA GPUs designed for matrix operations crucial to AI workloads. 

TensorRT: NVIDIA’s deep learning inference library optimized for real-time AI applications. 

Text-to-Image Generator: An AI tool that converts text descriptions into images. 

Text-to-Speech: AI technology that converts written text into spoken words. 

Throughput: The amount of data processed by a system in a given amount of time, often measured in bits per second (bps). 

V 

Virtualization: Creating virtual versions of resources like servers or storage to enable multiple OSs on one machine. 

Volta Architecture: NVIDIA’s GPU architecture that introduced Tensor Cores, widely used in AI tasks.