Graphics Processing Unit
Subject: Computer and Technology noun /ˈɡræf.ɪks ˈprɑː.sɛs.ɪŋ ˈjuː.nɪt/ 1 definition
- computer-hardware
- computer-architecture
- parallel-computing
- graphics-rendering
-
Definition (noun)
Entry #102 · UpdatedA Graphics Processing Unit (GPU) is a specialized electronic processor engineered for high-throughput, massively parallel computation across large datasets. Unlike a Central Processing Unit (CPU) that optimizes a small number of complex cores with large caches and sophisticated branch prediction for low-latency sequential control, a GPU dedicates most of its silicon area to hundreds or thousands of streamlined arithmetic logic units operating together across concurrent threads. Originally designed to execute real-time 3D rendering pipelines—transforming vertices, shading pixels, and rasterizing frames—its parallel matrix and vector capabilities now power general-purpose GPU (GPGPU) workloads such as machine learning, scientific simulation, and video encoding, while operating as an accelerator managed by a host CPU rather than independently running an operating system.
Common Types
Discrete GPU (dGPU): Mounts on a dedicated expansion card—typically via PCI Express—or standalone board module with its own high-bandwidth video memory (VRAM, such as GDDR6 or HBM) and cooling solution, maximizing rendering and compute throughput for workstations, gaming systems, and servers.
Integrated GPU (iGPU): Shares the same silicon die or chip package and unified system RAM with the host CPU, trading peak raw throughput for lower cost, compact form factor, and high power efficiency in laptops, ultrabooks, and mobile Systems-on-Chip (SoCs).
Data-Center Compute GPU: Omits display output ports in favor of high-bandwidth memory (HBM), error-correcting code (ECC) protection, specialized tensor or matrix cores, and fast chip-to-chip interconnects like NVLink to accelerate large-scale AI training, inference, and high-performance computing (HPC) clusters.
Virtual GPU (vGPU): Partitions a physical server GPU's compute slices and framebuffer across multiple isolated virtual machines or containers through hypervisor hardware virtualization, enabling multi-tenant cloud graphics workstations and shared inference workloads.
Example
When a 3D game engine renders a 4K frame at 60 frames per second, the host CPU sends scene geometry, textures, and draw commands over the PCIe bus to a discrete GPU's VRAM. Thousands of GPU shader cores then execute vertex transformations, ray-traced lighting calculations, and pixel shading in parallel to assemble the final image on screen in under 16.7 milliseconds.