Chips, servers, and computing infrastructure that power AI.

Processing Units

• GPU - Graphics chips repurposed for AI computation
• TPU - Google's custom chips designed specifically for AI
• NPU - Neural processing units optimized for AI inference
• CPU - Standard processors that can run AI models

High-End Hardware

• NVIDIA H100 - Current top-tier GPU for AI training
• NVIDIA A100 - Previous generation enterprise AI chip
• Edge TPU - Small, efficient chips for AI on devices
• Tensor Cores - Specialized units in NVIDIA GPUs for AI math

Training Infrastructure

• GPU Clusters - Multiple GPUs working together
• Distributed Training - Spreading AI training across many machines
• Data Centers - Facilities housing AI computing infrastructure
• Cloud Computing - Renting AI hardware over the internet

Optimization Techniques

• Model Parallelism - Splitting large models across multiple chips
• Data Parallelism - Processing different data on multiple chips
• Pipeline Parallelism - Breaking model processing into stages
• Quantization - Using lower precision to speed up computation

Memory and Storage

• HBM - High bandwidth memory for fast AI computation
• VRAM - Video memory where AI models are loaded
• NVMe - Fast storage for loading large AI models
• Memory Bandwidth - Speed of data transfer between components

Performance Metrics

• TOPS - Trillion operations per second, measure of AI chip speed
• FLOPS - Floating point operations per second
• Latency - Time for AI to generate a response
• Throughput - Number of AI requests processed per second

← Back to AI Glossary