Graphics processing unit

Lore & Background

The modern GPU emerged during the 1990s, adding the ability to perform operations like drawing lines and text without CPU help, and later adding 3D functionality. Earlier graphics controllers known as video display controllers had no internal calculation capabilities, while blitters performed only basic memory movement operations. Modern GPUs include hundreds or thousands of calculation units, enabling them to handle embarrassingly parallel problems. Graphics functions are generally independent, lending themselves to implementation on separate calculation engines. The ability of GPUs to rapidly perform vast numbers of calculations has led to their adoption in artificial intelligence, neural network training, and cryptocurrency mining. In personal computers, GPUs exist in two main forms: dedicated graphics, which use on-board RAM specially selected for serial workloads such as GDDR SDRAM, and integrated graphics, which use a portion of system RAM and are less costly but tend to be less capable. Technologies such as Scalable Link Interface (SLI), NVLink, and CrossFire allow multiple GPUs to draw images simultaneously, though these are increasingly uncommon in games. Multiple GPUs are still used on supercomputers, workstations, and for AI training.

Reader's Guide

The GPU's significance lies in its transformation from a specialized graphics accelerator to a general-purpose parallel processor. The article notes that GPUs are increasingly used for AI processing and model training due to linear algebra acceleration, which is also used extensively in graphics processing. Their parallel structure makes them useful for non-graphic calculations involving embarrassingly parallel problems. In certain applications requiring massive vector operations, a GPGPU can yield several orders of magnitude higher performance than a conventional CPU. The article states that GPU acceleration plays a key role in modern large-scale modeling, with TOP500 data showing that the majority of top-performing supercomputers rely on GPU integrated systems. Since 2005, there has been interest in using GPU performance for evolutionary computation and genetic programming. The legacy of the GPU includes its adoption in diverse fields such as cryptocurrency mining, neural network training, and distributed computing projects like Folding@home, where in certain circumstances the GPU calculates forty times faster than CPUs traditionally used. The article also notes that modern GPUs include dedicated hardware blocks for ray tracing, video encoding, and AI acceleration.

Did You Know?

• In 2009, Intel held 49.4% of the GPU market share, followed by Nvidia at 27.8% and AMD/ATI at 20.6%.
• Modern GPUs typically deliver performance measured in teraflops (TFLOPS).
• Integrated GPUs use system memory with bandwidth up to 128 gigabytes per second, while discrete GPUs may exceed 1000 gigabytes per second.
• External GPUs connect via PCI Express, with Thunderbolt 3 offering up to 40 gigabits per second.