Graphics processing unit

Graphics processing unit
source
Name	Graphics processing unit
Caption	A modern NVIDIA GeForce Founders Edition card.
Inventor	Intel, IBM, NEC
First production	1980s
Related components	Central processing unit, Video random access memory, PCI Express

Graphics processing unit. A graphics processing unit is a specialized electronic circuit designed to rapidly manipulate and alter memory to accelerate the creation of images in a frame buffer intended for output to a display device. Initially developed to accelerate the rendering of 3D computer graphics for applications like video games and computer-aided design, their highly parallel structure has made them effective for a range of complex computational tasks beyond graphics. Modern units are essential components in systems ranging from personal computers and game consoles to supercomputers and data centers.

History

The evolution of dedicated graphics hardware began in the late 1970s and early 1980s with companies like Intel, IBM, and NEC producing early video display controllers. A significant milestone was the introduction of the Nintendo Entertainment System, which popularized dedicated graphics chips in home consoles. The 1990s saw the rise of the 3D accelerator market, driven by companies such as 3dfx Interactive with its Voodoo Graphics card and NVIDIA with the RIVA 128. This era culminated in the development of programmable shader architectures, pioneered by Microsoft's DirectX 8.0 API and hardware from ATI Technologies and NVIDIA, which transformed them into general-purpose parallel processors. The launch of NVIDIA's GeForce 256, marketed with the term, is often cited as a defining moment.

Architecture

Modern architecture is characterized by a massively parallel array of stream processors or CUDA cores organized into larger groupings such as NVIDIA's Streaming Multiprocessor or AMD's Compute Unit. These cores execute threads in a Single instruction, multiple data fashion. Key architectural components include dedicated video random access memory, high-bandwidth memory interfaces like GDDR6 or HBM2, and fixed-function hardware for tasks such as rasterization, ray tracing (via dedicated RT cores), and AI acceleration (via Tensor cores or Matrix cores). The overall design is managed by a proprietary microarchitecture, such as NVIDIA's Ada Lovelace or AMD's RDNA 3.

Performance and metrics

Performance is typically measured in operations per second, such as FLOPS, and real-world rendering speed in frames per second for specific applications like 3DMark or games. Key metrics include memory bandwidth, determined by the memory bus width and memory clock rate, and computational throughput, often quoted in texel or pixel fill rates. Benchmarking suites from organizations like UL (safety organization) and specialized reviews from media outlets like AnandTech and Tom's Hardware provide comparative analysis. Performance is also increasingly gauged by capabilities in artificial intelligence workloads, using benchmarks from the MLPerf consortium.

Applications

Beyond rendering for video games and professional visualization software like Autodesk Maya, their parallel compute capability is leveraged in general-purpose computing on graphics processing units. This includes scientific computing for simulations in fields like computational fluid dynamics, machine learning and deep learning training facilitated by frameworks such as TensorFlow and PyTorch, cryptocurrency mining (notably for Bitcoin and Ethereum), and video encoding/decoding via APIs like NVIDIA NVENC. They are also critical in emerging fields like metaverse platforms and autonomous vehicle perception systems.

Manufacturers and market

The discrete market is dominated by NVIDIA and AMD, which design and sell chips, while companies like ASUS, MSI, and Gigabyte Technology manufacture branded add-in boards. Intel has also entered the discrete market with its Intel Arc series. The integrated graphics segment is led by Intel and AMD within their central processing unit offerings. The market is highly competitive, with significant sales driven by gaming enthusiasts, data center deployments for cloud computing providers like Amazon Web Services, and demand from research institutions for high-performance computing clusters.