GPU Architecture

What is a GPU?

A Graphics Processing Unit (GPU) is specialised hardware for graphics rendering:

• Optimised for floating-point operations on large arrays (vertices, pixels)
• Massively parallel: hundreds to thousands of SIMD processors
• High memory bandwidth (>1000 GB/s)
• Efficiently handles millions of triangles and pixels simultaneously

What Does a GPU Do?

• Low-level tasks: Clipping, rasterisation, hidden surface removal
• High-level tasks: Procedural shading, texturing, animation, simulation
• Modern GPUs: Ray tracing acceleration, physics, machine learning

GPU APIs

An Application Programming Interface (API) provides access to GPU functionality:

API	Platform	Notes
OpenGL	Cross-platform	Open standard, general 3D applications
DirectX	Windows/Xbox	Proprietary, game-focused
Vulkan	Cross-platform	Low-overhead, high performance
Metal	Apple devices	Low-level, Swift/Objective-C
OpenGL ES	Mobile	Stripped-down OpenGL for embedded systems
WebGL	Browser	JavaScript, based on OpenGL ES
WebGPU	Browser	Access to Vulkan/Metal/DirectX 12

GPGPU: General Purpose Computing

• CUDA: Nvidia’s parallel computing architecture (C-like)
• OpenCL: Open standard for parallel computing (AMD, Intel, Nvidia)

Used for physics simulation, machine learning, scientific computing.

GPU vs CPU Architecture

CPU Design

• Few powerful cores (4-64 typical)
• Optimised for sequential code and complex logic
• Large caches per core
• Branch prediction, out-of-order execution

GPU Design

• Many simpler cores (thousands)
• Optimised for parallel throughput
• Small caches shared across many threads
• SIMT (Single Instruction, Multiple Threads) execution

Memory Hierarchy

GPU Memory Types

• Global memory: Large, high-latency, accessible by all threads
• Shared memory: Fast, user-managed cache within a block
• Registers: Fastest, private to each thread
• Texture memory: Cached, optimised for 2D spatial locality
• Constant memory: Read-only, cached, broadcast to threads

Bandwidth Considerations

Modern GPUs achieve memory bandwidths exceeding 1000 GB/s through:

• Wide memory buses (256-384 bit)
• High clock speeds (GDDR6X, HBM2)
• Multiple memory controllers operating in parallel

---

Summary

• GPUs are massively parallel processors optimised for graphics workloads
• Multiple APIs exist for GPU access (OpenGL, Vulkan, DirectX, Metal)
• GPGPU enables general-purpose computation on graphics hardware
• GPU architecture prioritises throughput over single-thread performance

---

Past Paper Questions

2024 Paper 3 Question 4: What are the key differences between a GPU and a CPU architecture?

2023 Paper 3 Question 4: Explain what GPGPU means and give two applications.

2022 Paper 3 Question 4: Why are GPUs more suitable than CPUs for graphics rendering?