The Number of Floating-Point Operations Performed by a Graphics Processing Unit

Chapter 4 discusses graphics processing units (GPUs), specialized chips optimized for tasks like rendering 3D graphics and accelerating machine learning. Unlike CPUs, GPUs handle simpler operations such as matrix–vector multiplications, enabling thousands of cores to work in parallel with high memory bandwidth. Originally designed for gaming, GPUs evolved into general-purpose computing units (GPGPUs), becoming essential for AI, scientific simulations, and data centers. Nvidia introduced the first true GPU, GeForce 256, in 1999, and has since led the market alongside AMD and Intel. GPU performance, measured in FLOPS, has grown exponentially, doubling roughly every 1.9 years—faster than Moore’s law in recent years. Innovations like Tensor Cores, real-time ray tracing, and CUDA software have expanded GPU capabilities. The latest Nvidia RTX™ 5090 achieves 104.8 teraFLOPS at 575 watts, retailing for $1,999. GPUs are now critical for high-performance computing, though growth may slow as Moore’s law approaches its limits.