GPU Computing

Definition

GPU computing is the use of a graphics processing unit's highly parallel architecture to accelerate general-purpose computation, not just rendering graphics, by running the same operation across thousands of data elements simultaneously.

Overview

A GPU contains thousands of relatively simple cores designed to perform the same operation on many pieces of data at once, in contrast to a CPU's smaller number of powerful cores optimized for sequential, general-purpose logic. This architecture happens to be well suited to workloads beyond graphics rendering that involve large amounts of repetitive, parallelizable math, most notably the matrix and vector operations that underpin modern machine learning and deep learning. Developers access this parallel hardware through general-purpose GPU computing frameworks, the most widely used being NVIDIA's CUDA, which lets programmers write code that runs directly on GPU cores rather than only through graphics APIs. This shift, sometimes called GPGPU (general-purpose computing on GPUs), was a key enabler of the deep learning boom, since training large neural networks requires massive numbers of parallel matrix multiplications that would be impractically slow on CPUs alone. Beyond AI, GPU computing accelerates scientific simulation, computer vision, cryptography, and financial modeling — any domain where the same calculation needs to be repeated across large datasets — and courses like PyTorch Deep Learning cover how to apply GPU-accelerated training in practice.

Key Concepts

• Thousands of parallel cores optimized for repetitive, data-parallel operations
• Accelerates matrix and vector math central to machine learning
• Accessed through general-purpose frameworks such as CUDA and OpenCL
• Complements rather than replaces CPUs, which remain better for sequential logic
• Central to modern deep learning training and inference performance
• Also used in scientific simulation, cryptography, and financial modeling

Use Cases

Frequently Asked Questions

Related in Computer Science